release more models

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 Stability AI
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSE-MODEL

@@ -0,0 +1,84 @@
+Copyright (c) 2022 Stability AI and contributors
+
+CreativeML Open RAIL++-M License
+dated November 24, 2022
+
+Section I: PREAMBLE
+
+Multimodal generative models are being widely adopted and used, and have the potential to transform the way artists, among other individuals, conceive and benefit from AI or ML technologies as a tool for content creation.
+
+Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations.
+
+In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the Model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for art and content generation.
+
+Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this License aims to strike a balance between both in order to enable responsible open-science in the field of AI.
+
+This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model.
+
+NOW THEREFORE, You and Licensor agree as follows:
+
+1. Definitions
+
+- "License" means the terms and conditions for use, reproduction, and Distribution as defined in this document.
+- "Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License.
+- "Output" means the results of operating a Model as embodied in informational content resulting therefrom.
+- "Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material.
+- "Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model.
+- "Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any.
+- "Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access.
+- "Licensor" means the copyright owner or entity authorized by the copyright owner that is granting the License, including the persons or entities that may have rights in the Model and/or distributing the Model.
+- "You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, image generator.
+- "Third Parties" means individuals or legal entities that are not under common control with Licensor or You.
+- "Contribution" means any work of authorship, including the original version of the Model and any modifications or additions to that Model or Derivatives of the Model thereof, that is intentionally submitted to Licensor for inclusion in the Model by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Model, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
+- "Contributor" means Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Model.
+
+Section II: INTELLECTUAL PROPERTY RIGHTS
+
+Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III.
+
+2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model.
+3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Model to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material or a Contribution incorporated within the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or Work shall terminate as of the date such litigation is asserted or filed.
+
+Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION
+
+4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions:
+Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material.
+You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License;
+You must cause any modified files to carry prominent notices stating that You changed the files;
+You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model.
+You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. - for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License.
+5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5).
+6. The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License.
+
+Section IV: OTHER PROVISIONS
+
+7. Updates and Runtime Restrictions. To the maximum extent permitted by law, Licensor reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License.
+8. Trademarks and related. Nothing in this License permits You to make use of Licensors’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by the Licensors.
+9. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Model and the Complementary Material (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License.
+10. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
+11. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
+12. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein.
+
+END OF TERMS AND CONDITIONS
+
+
+
+
+Attachment A
+
+Use Restrictions
+
+You agree not to use the Model or Derivatives of the Model:
+
+- In any way that violates any applicable national, federal, state, local or international law or regulation;
+- For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
+- To generate or disseminate verifiably false information and/or content with the purpose of harming others;
+- To generate or disseminate personal identifiable information that can be used to harm an individual;
+- To defame, disparage or otherwise harass others;
+- For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation;
+- For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics;
+- To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
+- For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories;
+- To provide medical advice and medical results interpretation;
+- To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use).
+

README.md

@@ -0,0 +1,239 @@
+# Stable Diffusion 2.0
+![t2i](assets/stable-samples/txt2img/768/merged-0006.png)
+![t2i](assets/stable-samples/txt2img/768/merged-0002.png)
+![t2i](assets/stable-samples/txt2img/768/merged-0005.png)
+
+This repository contains [Stable Diffusion](https://github.com/CompVis/stable-diffusion) models trained from scratch and will be continuously updated with
+new checkpoints. The following list provides an overview of all currently available models. More coming soon.
+## News
+**November 2022**
+- New stable diffusion model (_Stable Diffusion 2.0-v_) at 768x768 resolution. Same number of parameters in the U-Net as 1.5, but uses [OpenCLIP-ViT/H](https://github.com/mlfoundations/open_clip) as the text encoder and is trained from scratch. _SD 2.0-v_ is a so-called [v-prediction](https://arxiv.org/abs/2202.00512) model. 
+- The above model is finetuned from _SD 2.0-base_, which was trained as a standard noise-prediction model on 512x512 images and is also made available.
+- Added a [x4 upscaling latent text-guided diffusion model](#image-upscaling-with-stable-diffusion).
+- New [depth-guided stable diffusion model](#depth-conditional-stable-diffusion), finetuned from _SD 2.0-base_. The model is conditioned on monocular depth estimates inferred via [MiDaS](https://github.com/isl-org/MiDaS) and can be used for structure-preserving img2img and shape-conditional synthesis.
+
+  ![d2i](assets/stable-samples/depth2img/depth2img01.png)
+- A [text-guided inpainting model](#image-inpainting-with-stable-diffusion), finetuned from SD _2.0-base_.
+
+We follow the [original repository](https://github.com/CompVis/stable-diffusion) and provide basic inference scripts to sample from the models.
+
+________________
+*The original Stable Diffusion model was created in a collaboration with [CompVis](https://arxiv.org/abs/2202.00512) and [RunwayML](https://runwayml.com/) and builds upon the work:*
+
+[**High-Resolution Image Synthesis with Latent Diffusion Models**](https://ommer-lab.com/research/latent-diffusion-models/)<br/>
+[Robin Rombach](https://github.com/rromb)\*,
+[Andreas Blattmann](https://github.com/ablattmann)\*,
+[Dominik Lorenz](https://github.com/qp-qp)\,
+[Patrick Esser](https://github.com/pesser),
+[Björn Ommer](https://hci.iwr.uni-heidelberg.de/Staff/bommer)<br/>
+_[CVPR '22 Oral](https://openaccess.thecvf.com/content/CVPR2022/html/Rombach_High-Resolution_Image_Synthesis_With_Latent_Diffusion_Models_CVPR_2022_paper.html) |
+[GitHub](https://github.com/CompVis/latent-diffusion) | [arXiv](https://arxiv.org/abs/2112.10752) | [Project page](https://ommer-lab.com/research/latent-diffusion-models/)_
+
+and [many others](#shout-outs).
+
+Stable Diffusion is a latent text-to-image diffusion model.
+________________________________
+  
+## Requirements
+
+You can update an existing [latent diffusion](https://github.com/CompVis/latent-diffusion) environment by running
+
+```
+conda install pytorch==1.12.1 torchvision==0.13.1 -c pytorch
+pip install transformers==4.19.2 diffusers invisible-watermark
+pip install -e .
+``` 
+#### xformers efficient attention
+For more efficiency and speed on GPUs, 
+we highly recommended installing the [xformers](https://github.com/facebookresearch/xformers)
+library.
+
+Tested on A100 with CUDA 11.4.
+Installation needs a somewhat recent version of nvcc and gcc/g++, obtain those, e.g., via 
+```commandline
+export CUDA_HOME=/usr/local/cuda-11.4
+conda install -c nvidia/label/cuda-11.4.0 cuda-nvcc
+conda install -c conda-forge gcc
+conda install -c conda-forge gxx_linux-64=9.5.0
+```
+
+Then, run the following (compiling takes up to 30 min).
+
+```commandline
+cd ..
+git clone https://github.com/facebookresearch/xformers.git
+cd xformers
+git submodule update --init --recursive
+pip install -r requirements.txt
+pip install -e .
+cd ../stable-diffusion
+```
+Upon successful installation, the code will automatically default to [memory efficient attention](https://github.com/facebookresearch/xformers)
+for the self- and cross-attention layers in the U-Net and autoencoder.
+
+## General Disclaimer
+Stable Diffusion models are general text-to-image diffusion models and therefore mirror biases and (mis-)conceptions that are present
+in their training data. Although efforts were made to reduce the inclusion of explicit pornographic material, **we do not recommend using the provided weights for services or products without additional safety mechanisms and considerations.
+The weights are research artifacts and should be treated as such.**
+Details on the training procedure and data, as well as the intended use of the model can be found in the corresponding [model card](modelcard.md).
+The weights are available via [the StabilityAI organization at Hugging Face](https://huggingface.co/StabilityAI) under the [CreativeML Open RAIL++-M License](LICENSE-MODEL). 
+
+
+
+## Stable Diffusion v2.0
+
+Stable Diffusion v2.0 refers to a specific configuration of the model
+architecture that uses a downsampling-factor 8 autoencoder with an 865M UNet
+and OpenCLIP ViT-H/14 text encoder for the diffusion model. The _SD 2.0-v_ model produces 768x768 px outputs. 
+
+Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,
+5.0, 6.0, 7.0, 8.0) and 50 DDIM sampling steps show the relative improvements of the checkpoints:
+
+![sd evaluation results](assets/model-variants.jpg)
+
+
+
+### Text-to-Image
+![txt2img-stable2](assets/stable-samples/txt2img/merged-0003.png)
+![txt2img-stable2](assets/stable-samples/txt2img/merged-0001.png)
+
+Stable Diffusion 2.0 is a latent diffusion model conditioned on the penultimate text embeddings of a CLIP ViT-H/14 text encoder.
+We provide a [reference script for sampling](#reference-sampling-script).
+#### Reference Sampling Script
+
+This script incorporates an [invisible watermarking](https://github.com/ShieldMnt/invisible-watermark) of the outputs, to help viewers [identify the images as machine-generated](scripts/tests/test_watermark.py).
+We provide the configs for the _SD2.0-v_ (768px) and _SD2.0-base_ (512px) model.
+
+First, download the weights for [_SD2.0-v_](https://huggingface.co/stabilityai/stable-diffusion-2) and [_SD2.0-base_](https://huggingface.co/stabilityai/stable-diffusion-2-base). 
+
+To sample from the _SD2.0-v_ model, run the following:
+
+```
+python scripts/txt2img.py --prompt "a professional photograph of an astronaut riding a horse" --ckpt <path/to/768model.ckpt/> --config configs/stable-diffusion/v2-inference-v.yaml --H 768 --W 768  
+```
+or try out the Web Demo: [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/stabilityai/stable-diffusion).
+
+To sample from the base model, use
+```
+python scripts/txt2img.py --prompt "a professional photograph of an astronaut riding a horse" --ckpt <path/to/model.ckpt/> --config <path/to/config.yaml/>  
+```
+
+By default, this uses the [DDIM sampler](https://arxiv.org/abs/2010.02502), and renders images of size 768x768 (which it was trained on) in 50 steps. 
+Empirically, the v-models can be sampled with higher guidance scales.
+
+Note: The inference config for all model versions is designed to be used with EMA-only checkpoints. 
+For this reason `use_ema=False` is set in the configuration, otherwise the code will try to switch from
+non-EMA to EMA weights. 
+
+### Image Modification with Stable Diffusion
+
+![depth2img-stable2](assets/stable-samples/depth2img/merged-0000.png)
+#### Depth-Conditional Stable Diffusion
+
+To augment the well-established [img2img](https://github.com/CompVis/stable-diffusion#image-modification-with-stable-diffusion) functionality of Stable Diffusion, we provide a _shape-preserving_ stable diffusion model.
+
+
+Note that the original method for image modification introduces significant semantic changes w.r.t. the initial image.
+If that is not desired, download our [depth-conditional stable diffusion](https://huggingface.co/stabilityai/stable-diffusion-2-depth) model and the `dpt_hybrid` MiDaS [model weights](https://github.com/intel-isl/DPT/releases/download/1_0/dpt_hybrid-midas-501f0c75.pt), place the latter in a folder `midas_models` and sample via 
+```
+python scripts/streamlit/depth2img.py streamlit run scripts/demo/depth2img.py configs/stable-diffusion/v2-midas-inference.yaml <path-to-ckpt>
+```
+
+This method can be used on the samples of the base model itself.
+For example, take [this sample](assets/stable-samples/depth2img/old_man.png) generated by an anonymous discord user.
+Using the [streamlit](https://streamlit.io/) script `depth2img.py`, the MiDaS model first infers a monocular depth estimate given this input, 
+and the diffusion model is then conditioned on the (relative) depth output.
+
+<p align="center">
+<b> depth2image </b><br/>
+<img src=assets/stable-samples/depth2img/d2i.gif/>
+</p>
+
+This model is particularly useful for a photorealistic style; see the [examples](assets/stable-samples/depth2img).
+For a maximum strength of 1.0, the model removes all pixel-based information and only relies on the text prompt and the inferred monocular depth estimate.
+
+![depth2img-stable3](assets/stable-samples/depth2img/merged-0005.png)
+
+#### Classic Img2Img
+
+For running the "classic" img2img, use
+```
+python scripts/img2img.py --prompt "A fantasy landscape, trending on artstation" --init-img <path-to-img.jpg> --strength 0.8 --ckpt <path/to/model.ckpt>
+```
+and adapt the checkpoint and config paths accordingly.
+
+### Image Upscaling with Stable Diffusion
+![upscaling-x4](assets/stable-samples/upscaling/merged-dog.png)
+After [downloading the weights](https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler), run
+```
+python scripts/gradio/inpainting.py scripts/gradio/superresolution.py configs/stable-diffusion/x4-upscaling.yaml <path-to-checkpoint>
+```
+
+or
+
+```
+streamlit run scripts/streamlit/superresolution.py -- configs/stable-diffusion/upscaling_x4.yaml <path-to-checkpoint>
+```
+
+for a Gradio or Streamlit demo of the text-guided x4 superresolution model.  
+This model can be used both on real inputs and on synthesized examples. For the latter, we recommend setting a higher 
+`noise_level`, e.g. `noise_level=100`.
+
+### Image Inpainting with Stable Diffusion
+
+![inpainting-stable2](assets/stable-inpainting/merged-leopards.png)
+
+[Download the SD 2.0-inpainting checkpoint](https://huggingface.co/stabilityai/stable-diffusion-2-inpainting) and run
+
+```
+python scripts/gradio/inpainting.py configs/stable-diffusion/v2-inpainting-inference.yaml <path-to-checkpoint>
+```
+
+or
+
+```
+streamlit run scripts/streamlit/inpainting.py -- configs/stable-diffusion/v2-inpainting-inference.yaml <path-to-checkpoint>
+```
+
+for a Gradio or Streamlit demo of the inpainting model. 
+This scripts adds invisible watermarking to the demo in the [RunwayML](https://github.com/runwayml/stable-diffusion/blob/main/scripts/inpaint_st.py) repository, but both should work interchangeably with the checkpoints/configs.  
+
+
+
+## Shout-Outs
+- Thanks to [Hugging Face](https://huggingface.co/) and in particular [Apolinário](https://github.com/apolinario)  for support with our model releases!
+- Stable Diffusion would not be possible without [LAION](https://laion.ai/) and their efforts to create open, large-scale datasets.
+- The [DeepFloyd team](https://twitter.com/deepfloydai) at Stability AI, for creating the subset of [LAION-5B](https://laion.ai/blog/laion-5b/) dataset used to train the model.
+- Stable Diffusion 2.0 uses [OpenCLIP](https://laion.ai/blog/large-openclip/), trained by [Romain Beaumont](https://github.com/rom1504).  
+- Our codebase for the diffusion models builds heavily on [OpenAI's ADM codebase](https://github.com/openai/guided-diffusion)
+and [https://github.com/lucidrains/denoising-diffusion-pytorch](https://github.com/lucidrains/denoising-diffusion-pytorch). 
+Thanks for open-sourcing!
+- [CompVis](https://github.com/CompVis/stable-diffusion) initial stable diffusion release
+- [Patrick](https://github.com/pesser)'s [implementation](https://github.com/runwayml/stable-diffusion/blob/main/scripts/inpaint_st.py) of the streamlit demo for inpainting.
+- `img2img` is an application of [SDEdit](https://arxiv.org/abs/2108.01073) by [Chenlin Meng](https://cs.stanford.edu/~chenlin/) from the [Stanford AI Lab](https://cs.stanford.edu/~ermon/website/). 
+- [Kat's implementation]((https://github.com/CompVis/latent-diffusion/pull/51)) of the [PLMS](https://arxiv.org/abs/2202.09778) sampler, and [more](https://github.com/crowsonkb/k-diffusion).
+- [DPMSolver](https://arxiv.org/abs/2206.00927) [integration](https://github.com/CompVis/stable-diffusion/pull/440) by [Cheng Lu](https://github.com/LuChengTHU).
+- Facebook's [xformers](https://github.com/facebookresearch/xformers) for efficient attention computation.
+- [MiDaS](https://github.com/isl-org/MiDaS) for monocular depth estimation.
+
+
+## License
+
+The code in this repository is released under the MIT License.
+
+The weights are available via [the StabilityAI organization at Hugging Face](https://huggingface.co/StabilityAI), and released under the [CreativeML Open RAIL++-M License](LICENSE-MODEL) License.
+
+## BibTeX
+
+```
+@misc{rombach2021highresolution,
+      title={High-Resolution Image Synthesis with Latent Diffusion Models}, 
+      author={Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and Björn Ommer},
+      year={2021},
+      eprint={2112.10752},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV}
+}
+```
+
+

configs/stable-diffusion/v2-inference-v.yaml

@@ -0,0 +1,68 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    parameterization: "v"
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False # we set this to false because this is an inference only config
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        use_fp16: True
+        image_size: 32 # unused
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_head_channels: 64 # need to fix for flash-attn
+        use_spatial_transformer: True
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          #attn_type: "vanilla-xformers"
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+      params:
+        freeze: True
+        layer: "penultimate"

configs/stable-diffusion/v2-inference.yaml

@@ -0,0 +1,67 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False # we set this to false because this is an inference only config
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        use_fp16: True
+        image_size: 32 # unused
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_head_channels: 64 # need to fix for flash-attn
+        use_spatial_transformer: True
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          #attn_type: "vanilla-xformers"
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+      params:
+        freeze: True
+        layer: "penultimate"

configs/stable-diffusion/v2-inpainting-inference.yaml

@@ -0,0 +1,158 @@
+model:
+  base_learning_rate: 5.0e-05
+  target: ldm.models.diffusion.ddpm.LatentInpaintDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: hybrid
+    scale_factor: 0.18215
+    monitor: val/loss_simple_ema
+    finetune_keys: null
+    use_ema: False
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        image_size: 32 # unused
+        in_channels: 9
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_head_channels: 64 # need to fix for flash-attn
+        use_spatial_transformer: True
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          #attn_type: "vanilla-xformers"
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+            - 1
+            - 2
+            - 4
+            - 4
+          num_res_blocks: 2
+          attn_resolutions: [ ]
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+      params:
+        freeze: True
+        layer: "penultimate"
+
+
+data:
+  target: ldm.data.laion.WebDataModuleFromConfig
+  params:
+    tar_base: null  # for concat as in LAION-A
+    p_unsafe_threshold: 0.1
+    filter_word_list: "data/filters.yaml"
+    max_pwatermark: 0.45
+    batch_size: 8
+    num_workers: 6
+    multinode: True
+    min_size: 512
+    train:
+      shards:
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-0/{00000..18699}.tar -"
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-1/{00000..18699}.tar -"
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-2/{00000..18699}.tar -"
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-3/{00000..18699}.tar -"
+        - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-4/{00000..18699}.tar -"  #{00000-94333}.tar"
+      shuffle: 10000
+      image_key: jpg
+      image_transforms:
+      - target: torchvision.transforms.Resize
+        params:
+          size: 512
+          interpolation: 3
+      - target: torchvision.transforms.RandomCrop
+        params:
+          size: 512
+      postprocess:
+        target: ldm.data.laion.AddMask
+        params:
+          mode: "512train-large"
+          p_drop: 0.25
+    # NOTE use enough shards to avoid empty validation loops in workers
+    validation:
+      shards:
+        - "pipe:aws s3 cp s3://deep-floyd-s3/datasets/laion_cleaned-part5/{93001..94333}.tar - "
+      shuffle: 0
+      image_key: jpg
+      image_transforms:
+      - target: torchvision.transforms.Resize
+        params:
+          size: 512
+          interpolation: 3
+      - target: torchvision.transforms.CenterCrop
+        params:
+          size: 512
+      postprocess:
+        target: ldm.data.laion.AddMask
+        params:
+          mode: "512train-large"
+          p_drop: 0.25
+
+lightning:
+  find_unused_parameters: True
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 10000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        enable_autocast: False
+        disabled: False
+        batch_frequency: 1000
+        max_images: 4
+        increase_log_steps: False
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          inpaint: False
+          plot_progressive_rows: False
+          plot_diffusion_rows: False
+          N: 4
+          unconditional_guidance_scale: 5.0
+          unconditional_guidance_label: [""]
+          ddim_steps: 50  # todo check these out for depth2img,
+          ddim_eta: 0.0   # todo check these out for depth2img,
+
+  trainer:
+    benchmark: True
+    val_check_interval: 5000000
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1

configs/stable-diffusion/v2-midas-inference.yaml

@@ -0,0 +1,74 @@
+model:
+  base_learning_rate: 5.0e-07
+  target: ldm.models.diffusion.ddpm.LatentDepth2ImageDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: hybrid
+    scale_factor: 0.18215
+    monitor: val/loss_simple_ema
+    finetune_keys: null
+    use_ema: False
+
+    depth_stage_config:
+      target: ldm.modules.midas.api.MiDaSInference
+      params:
+        model_type: "dpt_hybrid"
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        image_size: 32 # unused
+        in_channels: 5
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_head_channels: 64 # need to fix for flash-attn
+        use_spatial_transformer: True
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          #attn_type: "vanilla-xformers"
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+            - 1
+            - 2
+            - 4
+            - 4
+          num_res_blocks: 2
+          attn_resolutions: [ ]
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+      params:
+        freeze: True
+        layer: "penultimate"
+
+

configs/stable-diffusion/x4-upscaling.yaml

@@ -0,0 +1,76 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentUpscaleDiffusion
+  params:
+    parameterization: "v"
+    low_scale_key: "lr"
+    linear_start: 0.0001
+    linear_end: 0.02
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 128
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: "hybrid-adm"
+    monitor: val/loss_simple_ema
+    scale_factor: 0.08333
+    use_ema: False
+
+    low_scale_config:
+      target: ldm.modules.diffusionmodules.upscaling.ImageConcatWithNoiseAugmentation
+      params:
+        noise_schedule_config: # image space
+          linear_start: 0.0001
+          linear_end: 0.02
+        max_noise_level: 350
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        num_classes: 1000  # timesteps for noise conditioning (here constant, just need one)
+        image_size: 128
+        in_channels: 7
+        out_channels: 4
+        model_channels: 256
+        attention_resolutions: [ 2,4,8]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 2, 4]
+        disable_self_attentions: [True, True, True, False]
+        disable_middle_self_attn: False
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        legacy: False
+        use_linear_in_transformer: True
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        ddconfig:
+          # attn_type: "vanilla-xformers" this model needs efficient attention to be feasible on HR data, also the decoder seems to break in half precision (UNet is fine though)
+          double_z: True
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [ 1,2,4 ]  # num_down = len(ch_mult)-1
+          num_res_blocks: 2
+          attn_resolutions: [ ]
+          dropout: 0.0
+
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+      params:
+        freeze: True
+        layer: "penultimate"
+

environment.yaml

@@ -0,0 +1,29 @@
+name: ldm
+channels:
+  - pytorch
+  - defaults
+dependencies:
+  - python=3.8.5
+  - pip=20.3
+  - cudatoolkit=11.3
+  - pytorch=1.12.1
+  - torchvision=0.13.1
+  - numpy=1.23.1
+  - pip:
+    - albumentations==1.3.0
+    - opencv-python==4.6.0.66
+    - imageio==2.9.0
+    - imageio-ffmpeg==0.4.2
+    - pytorch-lightning==1.4.2
+    - omegaconf==2.1.1
+    - test-tube>=0.7.5
+    - streamlit==1.12.1
+    - einops==0.3.0
+    - transformers==4.19.2
+    - webdataset==0.2.5
+    - kornia==0.6
+    - open_clip_torch==2.0.2
+    - invisible-watermark>=0.1.5
+    - streamlit-drawable-canvas==0.8.0
+    - torchmetrics==0.6.0
+    - -e .

ldm/data/util.py

@@ -0,0 +1,24 @@
+import torch
+
+from ldm.modules.midas.api import load_midas_transform
+
+
+class AddMiDaS(object):
+    def __init__(self, model_type):
+        super().__init__()
+        self.transform = load_midas_transform(model_type)
+
+    def pt2np(self, x):
+        x = ((x + 1.0) * .5).detach().cpu().numpy()
+        return x
+
+    def np2pt(self, x):
+        x = torch.from_numpy(x) * 2 - 1.
+        return x
+
+    def __call__(self, sample):
+        # sample['jpg'] is tensor hwc in [-1, 1] at this point
+        x = self.pt2np(sample['jpg'])
+        x = self.transform({"image": x})["image"]
+        sample['midas_in'] = x
+        return sample

ldm/models/autoencoder.py

@@ -0,0 +1,219 @@
+import torch
+import pytorch_lightning as pl
+import torch.nn.functional as F
+from contextlib import contextmanager
+
+from ldm.modules.diffusionmodules.model import Encoder, Decoder
+from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
+from ldm.util import instantiate_from_config
+from ldm.modules.ema import LitEma
+
+
+class AutoencoderKL(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 ema_decay=None,
+                 learn_logvar=False
+                 ):
+        super().__init__()
+        self.learn_logvar = learn_logvar
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        assert ddconfig["double_z"]
+        self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+
+        self.use_ema = ema_decay is not None
+        if self.use_ema:
+            self.ema_decay = ema_decay
+            assert 0. < ema_decay < 1.
+            self.model_ema = LitEma(self, decay=ema_decay)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path}")
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.parameters())
+            self.model_ema.copy_to(self)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self)
+
+    def encode(self, x):
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+        return dec
+
+    def forward(self, input, sample_posterior=True):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        return dec, posterior
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+
+        if optimizer_idx == 0:
+            # train encoder+decoder+logvar
+            aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # train the discriminator
+            discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                                last_layer=self.get_last_layer(), split="train")
+
+            self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        log_dict = self._validation_step(batch, batch_idx)
+        with self.ema_scope():
+            log_dict_ema = self._validation_step(batch, batch_idx, postfix="_ema")
+        return log_dict
+
+    def _validation_step(self, batch, batch_idx, postfix=""):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+        aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
+                                        last_layer=self.get_last_layer(), split="val"+postfix)
+
+        discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
+                                            last_layer=self.get_last_layer(), split="val"+postfix)
+
+        self.log(f"val{postfix}/rec_loss", log_dict_ae[f"val{postfix}/rec_loss"])
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        ae_params_list = list(self.encoder.parameters()) + list(self.decoder.parameters()) + list(
+            self.quant_conv.parameters()) + list(self.post_quant_conv.parameters())
+        if self.learn_logvar:
+            print(f"{self.__class__.__name__}: Learning logvar")
+            ae_params_list.append(self.loss.logvar)
+        opt_ae = torch.optim.Adam(ae_params_list,
+                                  lr=lr, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr, betas=(0.5, 0.9))
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    @torch.no_grad()
+    def log_images(self, batch, only_inputs=False, log_ema=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if not only_inputs:
+            xrec, posterior = self(x)
+            if x.shape[1] > 3:
+                # colorize with random projection
+                assert xrec.shape[1] > 3
+                x = self.to_rgb(x)
+                xrec = self.to_rgb(xrec)
+            log["samples"] = self.decode(torch.randn_like(posterior.sample()))
+            log["reconstructions"] = xrec
+            if log_ema or self.use_ema:
+                with self.ema_scope():
+                    xrec_ema, posterior_ema = self(x)
+                    if x.shape[1] > 3:
+                        # colorize with random projection
+                        assert xrec_ema.shape[1] > 3
+                        xrec_ema = self.to_rgb(xrec_ema)
+                    log["samples_ema"] = self.decode(torch.randn_like(posterior_ema.sample()))
+                    log["reconstructions_ema"] = xrec_ema
+        log["inputs"] = x
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class IdentityFirstStage(torch.nn.Module):
+    def __init__(self, *args, vq_interface=False, **kwargs):
+        self.vq_interface = vq_interface
+        super().__init__()
+
+    def encode(self, x, *args, **kwargs):
+        return x
+
+    def decode(self, x, *args, **kwargs):
+        return x
+
+    def quantize(self, x, *args, **kwargs):
+        if self.vq_interface:
+            return x, None, [None, None, None]
+        return x
+
+    def forward(self, x, *args, **kwargs):
+        return x
+

ldm/models/diffusion/ddim.py

@@ -0,0 +1,336 @@
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+
+from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               dynamic_threshold=None,
+               ucg_schedule=None,
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                ctmp = conditioning[list(conditioning.keys())[0]]
+                while isinstance(ctmp, list): ctmp = ctmp[0]
+                cbs = ctmp.shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+
+            elif isinstance(conditioning, list):
+                for ctmp in conditioning:
+                    if ctmp.shape[0] != batch_size:
+                        print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        print(f'Data shape for DDIM sampling is {size}, eta {eta}')
+
+        samples, intermediates = self.ddim_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    dynamic_threshold=dynamic_threshold,
+                                                    ucg_schedule=ucg_schedule
+                                                    )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None,
+                      ucg_schedule=None):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1. - mask) * img
+
+            if ucg_schedule is not None:
+                assert len(ucg_schedule) == len(time_range)
+                unconditional_guidance_scale = ucg_schedule[i]
+
+            outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      dynamic_threshold=dynamic_threshold)
+            img, pred_x0 = outs
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,
+                      dynamic_threshold=None):
+        b, *_, device = *x.shape, x.device
+
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+            model_output = self.model.apply_model(x, t, c)
+        else:
+            x_in = torch.cat([x] * 2)
+            t_in = torch.cat([t] * 2)
+            if isinstance(c, dict):
+                assert isinstance(unconditional_conditioning, dict)
+                c_in = dict()
+                for k in c:
+                    if isinstance(c[k], list):
+                        c_in[k] = [torch.cat([
+                            unconditional_conditioning[k][i],
+                            c[k][i]]) for i in range(len(c[k]))]
+                    else:
+                        c_in[k] = torch.cat([
+                                unconditional_conditioning[k],
+                                c[k]])
+            elif isinstance(c, list):
+                c_in = list()
+                assert isinstance(unconditional_conditioning, list)
+                for i in range(len(c)):
+                    c_in.append(torch.cat([unconditional_conditioning[i], c[i]]))
+            else:
+                c_in = torch.cat([unconditional_conditioning, c])
+            model_uncond, model_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+            model_output = model_uncond + unconditional_guidance_scale * (model_t - model_uncond)
+
+        if self.model.parameterization == "v":
+            e_t = self.model.predict_eps_from_z_and_v(x, t, model_output)
+        else:
+            e_t = model_output
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps", 'not implemented'
+            e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+        # select parameters corresponding to the currently considered timestep
+        a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+        a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+        sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
+
+        # current prediction for x_0
+        if self.model.parameterization != "v":
+            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        else:
+            pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
+
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+
+        if dynamic_threshold is not None:
+            raise NotImplementedError()
+
+        # direction pointing to x_t
+        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+        return x_prev, pred_x0
+
+    @torch.no_grad()
+    def encode(self, x0, c, t_enc, use_original_steps=False, return_intermediates=None,
+               unconditional_guidance_scale=1.0, unconditional_conditioning=None, callback=None):
+        num_reference_steps = self.ddpm_num_timesteps if use_original_steps else self.ddim_timesteps.shape[0]
+
+        assert t_enc <= num_reference_steps
+        num_steps = t_enc
+
+        if use_original_steps:
+            alphas_next = self.alphas_cumprod[:num_steps]
+            alphas = self.alphas_cumprod_prev[:num_steps]
+        else:
+            alphas_next = self.ddim_alphas[:num_steps]
+            alphas = torch.tensor(self.ddim_alphas_prev[:num_steps])
+
+        x_next = x0
+        intermediates = []
+        inter_steps = []
+        for i in tqdm(range(num_steps), desc='Encoding Image'):
+            t = torch.full((x0.shape[0],), i, device=self.model.device, dtype=torch.long)
+            if unconditional_guidance_scale == 1.:
+                noise_pred = self.model.apply_model(x_next, t, c)
+            else:
+                assert unconditional_conditioning is not None
+                e_t_uncond, noise_pred = torch.chunk(
+                    self.model.apply_model(torch.cat((x_next, x_next)), torch.cat((t, t)),
+                                           torch.cat((unconditional_conditioning, c))), 2)
+                noise_pred = e_t_uncond + unconditional_guidance_scale * (noise_pred - e_t_uncond)
+
+            xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next
+            weighted_noise_pred = alphas_next[i].sqrt() * (
+                    (1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt()) * noise_pred
+            x_next = xt_weighted + weighted_noise_pred
+            if return_intermediates and i % (
+                    num_steps // return_intermediates) == 0 and i < num_steps - 1:
+                intermediates.append(x_next)
+                inter_steps.append(i)
+            elif return_intermediates and i >= num_steps - 2:
+                intermediates.append(x_next)
+                inter_steps.append(i)
+            if callback: callback(i)
+
+        out = {'x_encoded': x_next, 'intermediate_steps': inter_steps}
+        if return_intermediates:
+            out.update({'intermediates': intermediates})
+        return x_next, out
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+        return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
+                extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
+
+    @torch.no_grad()
+    def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
+               use_original_steps=False, callback=None):
+
+        timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
+        x_dec = x_latent
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
+            x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
+                                          unconditional_guidance_scale=unconditional_guidance_scale,
+                                          unconditional_conditioning=unconditional_conditioning)
+            if callback: callback(i)
+        return x_dec

ldm/models/diffusion/dpm_solver/__init__.py

@@ -0,0 +1 @@
+from .sampler import DPMSolverSampler

ldm/models/diffusion/dpm_solver/sampler.py

@@ -0,0 +1,87 @@
+"""SAMPLING ONLY."""
+import torch
+
+from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver
+
+
+MODEL_TYPES = {
+    "eps": "noise",
+    "v": "v"
+}
+
+
+class DPMSolverSampler(object):
+    def __init__(self, model, **kwargs):
+        super().__init__()
+        self.model = model
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device)
+        self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod))
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+
+        print(f'Data shape for DPM-Solver sampling is {size}, sampling steps {S}')
+
+        device = self.model.betas.device
+        if x_T is None:
+            img = torch.randn(size, device=device)
+        else:
+            img = x_T
+
+        ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod)
+
+        model_fn = model_wrapper(
+            lambda x, t, c: self.model.apply_model(x, t, c),
+            ns,
+            model_type=MODEL_TYPES[self.model.parameterization],
+            guidance_type="classifier-free",
+            condition=conditioning,
+            unconditional_condition=unconditional_conditioning,
+            guidance_scale=unconditional_guidance_scale,
+        )
+
+        dpm_solver = DPM_Solver(model_fn, ns, predict_x0=True, thresholding=False)
+        x = dpm_solver.sample(img, steps=S, skip_type="time_uniform", method="multistep", order=2, lower_order_final=True)
+
+        return x.to(device), None

ldm/models/diffusion/plms.py

@@ -0,0 +1,244 @@
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+from functools import partial
+
+from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
+from ldm.models.diffusion.sampling_util import norm_thresholding
+
+
+class PLMSSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        if ddim_eta != 0:
+            raise ValueError('ddim_eta must be 0 for PLMS')
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               dynamic_threshold=None,
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        print(f'Data shape for PLMS sampling is {size}')
+
+        samples, intermediates = self.plms_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    dynamic_threshold=dynamic_threshold,
+                                                    )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def plms_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,
+                      dynamic_threshold=None):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        print(f"Running PLMS Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
+        old_eps = []
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+            ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1. - mask) * img
+
+            outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      old_eps=old_eps, t_next=ts_next,
+                                      dynamic_threshold=dynamic_threshold)
+            img, pred_x0, e_t = outs
+            old_eps.append(e_t)
+            if len(old_eps) >= 4:
+                old_eps.pop(0)
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None,
+                      dynamic_threshold=None):
+        b, *_, device = *x.shape, x.device
+
+        def get_model_output(x, t):
+            if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+                e_t = self.model.apply_model(x, t, c)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t] * 2)
+                c_in = torch.cat([unconditional_conditioning, c])
+                e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+                e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+
+            if score_corrector is not None:
+                assert self.model.parameterization == "eps"
+                e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+            return e_t
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+
+        def get_x_prev_and_pred_x0(e_t, index):
+            # select parameters corresponding to the currently considered timestep
+            a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+            a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+            sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+            sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
+
+            # current prediction for x_0
+            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+            if quantize_denoised:
+                pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+            if dynamic_threshold is not None:
+                pred_x0 = norm_thresholding(pred_x0, dynamic_threshold)
+            # direction pointing to x_t
+            dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+            noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+            if noise_dropout > 0.:
+                noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+            x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+            return x_prev, pred_x0
+
+        e_t = get_model_output(x, t)
+        if len(old_eps) == 0:
+            # Pseudo Improved Euler (2nd order)
+            x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
+            e_t_next = get_model_output(x_prev, t_next)
+            e_t_prime = (e_t + e_t_next) / 2
+        elif len(old_eps) == 1:
+            # 2nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (3 * e_t - old_eps[-1]) / 2
+        elif len(old_eps) == 2:
+            # 3nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
+        elif len(old_eps) >= 3:
+            # 4nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
+
+        x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
+
+        return x_prev, pred_x0, e_t

ldm/models/diffusion/sampling_util.py

@@ -0,0 +1,22 @@
+import torch
+import numpy as np
+
+
+def append_dims(x, target_dims):
+    """Appends dimensions to the end of a tensor until it has target_dims dimensions.
+    From https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/utils.py"""
+    dims_to_append = target_dims - x.ndim
+    if dims_to_append < 0:
+        raise ValueError(f'input has {x.ndim} dims but target_dims is {target_dims}, which is less')
+    return x[(...,) + (None,) * dims_to_append]
+
+
+def norm_thresholding(x0, value):
+    s = append_dims(x0.pow(2).flatten(1).mean(1).sqrt().clamp(min=value), x0.ndim)
+    return x0 * (value / s)
+
+
+def spatial_norm_thresholding(x0, value):
+    # b c h w
+    s = x0.pow(2).mean(1, keepdim=True).sqrt().clamp(min=value)
+    return x0 * (value / s)

ldm/modules/attention.py

@@ -0,0 +1,331 @@
+from inspect import isfunction
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+from einops import rearrange, repeat
+from typing import Optional, Any
+
+from ldm.modules.diffusionmodules.util import checkpoint
+
+
+try:
+    import xformers
+    import xformers.ops
+    XFORMERS_IS_AVAILBLE = True
+except:
+    XFORMERS_IS_AVAILBLE = False
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = rearrange(q, 'b c h w -> b (h w) c')
+        k = rearrange(k, 'b c h w -> b c (h w)')
+        w_ = torch.einsum('bij,bjk->bik', q, k)
+
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, 'b c h w -> b c (h w)')
+        w_ = rearrange(w_, 'b i j -> b j i')
+        h_ = torch.einsum('bij,bjk->bik', v, w_)
+        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, context=None, mask=None):
+        h = self.heads
+
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+        del q, k
+
+        if exists(mask):
+            mask = rearrange(mask, 'b ... -> b (...)')
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+
+        # attention, what we cannot get enough of
+        sim = sim.softmax(dim=-1)
+
+        out = einsum('b i j, b j d -> b i d', sim, v)
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+        return self.to_out(out)
+
+
+class MemoryEfficientCrossAttention(nn.Module):
+    # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        print(f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using "
+              f"{heads} heads.")
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
+        self.attention_op: Optional[Any] = None
+
+    def forward(self, x, context=None, mask=None):
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        b, _, _ = q.shape
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, t.shape[1], self.heads, self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * self.heads, t.shape[1], self.dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+
+        # actually compute the attention, what we cannot get enough of
+        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)
+
+        if exists(mask):
+            raise NotImplementedError
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, self.heads, out.shape[1], self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, out.shape[1], self.heads * self.dim_head)
+        )
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    ATTENTION_MODES = {
+        "softmax": CrossAttention,  # vanilla attention
+        "softmax-xformers": MemoryEfficientCrossAttention
+    }
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
+                 disable_self_attn=False):
+        super().__init__()
+        attn_mode = "softmax-xformers" if XFORMERS_IS_AVAILBLE else "softmax"
+        assert attn_mode in self.ATTENTION_MODES
+        attn_cls = self.ATTENTION_MODES[attn_mode]
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+                              context_dim=context_dim if self.disable_self_attn else None)  # is a self-attention if not self.disable_self_attn
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim,
+                              heads=n_heads, dim_head=d_head, dropout=dropout)  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context=None):
+        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    NEW: use_linear for more efficiency instead of the 1x1 convs
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None,
+                 disable_self_attn=False, use_linear=False,
+                 use_checkpoint=True):
+        super().__init__()
+        if exists(context_dim) and not isinstance(context_dim, list):
+            context_dim = [context_dim]
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+        if not use_linear:
+            self.proj_in = nn.Conv2d(in_channels,
+                                     inner_dim,
+                                     kernel_size=1,
+                                     stride=1,
+                                     padding=0)
+        else:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d],
+                                   disable_self_attn=disable_self_attn, checkpoint=use_checkpoint)
+                for d in range(depth)]
+        )
+        if not use_linear:
+            self.proj_out = zero_module(nn.Conv2d(inner_dim,
+                                                  in_channels,
+                                                  kernel_size=1,
+                                                  stride=1,
+                                                  padding=0))
+        else:
+            self.proj_out = zero_module(nn.Linear(in_channels, inner_dim))
+        self.use_linear = use_linear
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        if not isinstance(context, list):
+            context = [context]
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            x = block(x, context=context[i])
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = self.proj_out(x)
+        return x + x_in
+

ldm/modules/diffusionmodules/model.py

@@ -0,0 +1,852 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+from typing import Optional, Any
+
+from ldm.modules.attention import MemoryEfficientCrossAttention
+
+try:
+    import xformers
+    import xformers.ops
+    XFORMERS_IS_AVAILBLE = True
+except:
+    XFORMERS_IS_AVAILBLE = False
+    print("No module 'xformers'. Proceeding without it.")
+
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0,1,0,0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return x*torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = q.reshape(b,c,h*w)
+        q = q.permute(0,2,1)   # b,hw,c
+        k = k.reshape(b,c,h*w) # b,c,hw
+        w_ = torch.bmm(q,k)     # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b,c,h*w)
+        w_ = w_.permute(0,2,1)   # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v,w_)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b,c,h,w)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+class MemoryEfficientAttnBlock(nn.Module):
+    """
+        Uses xformers efficient implementation,
+        see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+        Note: this is a single-head self-attention operation
+    """
+    #
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+        self.attention_op: Optional[Any] = None
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        B, C, H, W = q.shape
+        q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v))
+
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(B, t.shape[1], 1, C)
+            .permute(0, 2, 1, 3)
+            .reshape(B * 1, t.shape[1], C)
+            .contiguous(),
+            (q, k, v),
+        )
+        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)
+
+        out = (
+            out.unsqueeze(0)
+            .reshape(B, 1, out.shape[1], C)
+            .permute(0, 2, 1, 3)
+            .reshape(B, out.shape[1], C)
+        )
+        out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C)
+        out = self.proj_out(out)
+        return x+out
+
+
+class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention):
+    def forward(self, x, context=None, mask=None):
+        b, c, h, w = x.shape
+        x = rearrange(x, 'b c h w -> b (h w) c')
+        out = super().forward(x, context=context, mask=mask)
+        out = rearrange(out, 'b (h w) c -> b c h w', h=h, w=w, c=c)
+        return x + out
+
+
+def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
+    assert attn_type in ["vanilla", "vanilla-xformers", "memory-efficient-cross-attn", "linear", "none"], f'attn_type {attn_type} unknown'
+    if XFORMERS_IS_AVAILBLE and attn_type == "vanilla":
+        attn_type = "vanilla-xformers"
+    print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        assert attn_kwargs is None
+        return AttnBlock(in_channels)
+    elif attn_type == "vanilla-xformers":
+        print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...")
+        return MemoryEfficientAttnBlock(in_channels)
+    elif type == "memory-efficient-cross-attn":
+        attn_kwargs["query_dim"] = in_channels
+        return MemoryEfficientCrossAttentionWrapper(**attn_kwargs)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        raise NotImplementedError()
+
+
+class Model(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch*4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList([
+                torch.nn.Linear(self.ch,
+                                self.temb_ch),
+                torch.nn.Linear(self.temb_ch,
+                                self.temb_ch),
+            ])
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            skip_in = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch*in_ch_mult[i_level]
+                block.append(ResnetBlock(in_channels=block_in+skip_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, t=None, context=None):
+        #assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 **ignore_kwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 attn_type="vanilla", **ignorekwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+        print("Working with z of shape {} = {} dimensions.".format(
+            self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class SimpleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, *args, **kwargs):
+        super().__init__()
+        self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
+                                     ResnetBlock(in_channels=in_channels,
+                                                 out_channels=2 * in_channels,
+                                                 temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=2 * in_channels,
+                                                out_channels=4 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=4 * in_channels,
+                                                out_channels=2 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     nn.Conv2d(2*in_channels, in_channels, 1),
+                                     Upsample(in_channels, with_conv=True)])
+        # end
+        self.norm_out = Normalize(in_channels)
+        self.conv_out = torch.nn.Conv2d(in_channels,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        for i, layer in enumerate(self.model):
+            if i in [1,2,3]:
+                x = layer(x, None)
+            else:
+                x = layer(x)
+
+        h = self.norm_out(x)
+        h = nonlinearity(h)
+        x = self.conv_out(h)
+        return x
+
+
+class UpsampleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
+                 ch_mult=(2,2), dropout=0.0):
+        super().__init__()
+        # upsampling
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        block_in = in_channels
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.res_blocks = nn.ModuleList()
+        self.upsample_blocks = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            res_block = []
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                res_block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+            self.res_blocks.append(nn.ModuleList(res_block))
+            if i_level != self.num_resolutions - 1:
+                self.upsample_blocks.append(Upsample(block_in, True))
+                curr_res = curr_res * 2
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # upsampling
+        h = x
+        for k, i_level in enumerate(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.res_blocks[i_level][i_block](h, None)
+            if i_level != self.num_resolutions - 1:
+                h = self.upsample_blocks[k](h)
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class LatentRescaler(nn.Module):
+    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
+        super().__init__()
+        # residual block, interpolate, residual block
+        self.factor = factor
+        self.conv_in = nn.Conv2d(in_channels,
+                                 mid_channels,
+                                 kernel_size=3,
+                                 stride=1,
+                                 padding=1)
+        self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+        self.attn = AttnBlock(mid_channels)
+        self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+
+        self.conv_out = nn.Conv2d(mid_channels,
+                                  out_channels,
+                                  kernel_size=1,
+                                  )
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        for block in self.res_block1:
+            x = block(x, None)
+        x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
+        x = self.attn(x)
+        for block in self.res_block2:
+            x = block(x, None)
+        x = self.conv_out(x)
+        return x
+
+
+class MergedRescaleEncoder(nn.Module):
+    def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True,
+                 ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        intermediate_chn = ch * ch_mult[-1]
+        self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
+                               z_channels=intermediate_chn, double_z=False, resolution=resolution,
+                               attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
+                               out_ch=None)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
+                                       mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.rescaler(x)
+        return x
+
+
+class MergedRescaleDecoder(nn.Module):
+    def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
+                 dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        tmp_chn = z_channels*ch_mult[-1]
+        self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
+                               resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
+                               ch_mult=ch_mult, resolution=resolution, ch=ch)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
+                                       out_channels=tmp_chn, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Upsampler(nn.Module):
+    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
+        super().__init__()
+        assert out_size >= in_size
+        num_blocks = int(np.log2(out_size//in_size))+1
+        factor_up = 1.+ (out_size % in_size)
+        print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
+        self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
+                                       out_channels=in_channels)
+        self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
+                               attn_resolutions=[], in_channels=None, ch=in_channels,
+                               ch_mult=[ch_mult for _ in range(num_blocks)])
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Resize(nn.Module):
+    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
+        super().__init__()
+        self.with_conv = learned
+        self.mode = mode
+        if self.with_conv:
+            print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
+            raise NotImplementedError()
+            assert in_channels is not None
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=4,
+                                        stride=2,
+                                        padding=1)
+
+    def forward(self, x, scale_factor=1.0):
+        if scale_factor==1.0:
+            return x
+        else:
+            x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
+        return x

ldm/modules/diffusionmodules/openaimodel.py

@@ -0,0 +1,786 @@
+from abc import abstractmethod
+import math
+
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ldm.modules.diffusionmodules.util import (
+    checkpoint,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    zero_module,
+    normalization,
+    timestep_embedding,
+)
+from ldm.modules.attention import SpatialTransformer
+from ldm.util import exists
+
+
+# dummy replace
+def convert_module_to_f16(x):
+    pass
+
+def convert_module_to_f32(x):
+    pass
+
+
+## go
+class AttentionPool2d(nn.Module):
+    """
+    Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
+    """
+
+    def __init__(
+        self,
+        spacial_dim: int,
+        embed_dim: int,
+        num_heads_channels: int,
+        output_dim: int = None,
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5)
+        self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
+        self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
+        self.num_heads = embed_dim // num_heads_channels
+        self.attention = QKVAttention(self.num_heads)
+
+    def forward(self, x):
+        b, c, *_spatial = x.shape
+        x = x.reshape(b, c, -1)  # NC(HW)
+        x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1)  # NC(HW+1)
+        x = x + self.positional_embedding[None, :, :].to(x.dtype)  # NC(HW+1)
+        x = self.qkv_proj(x)
+        x = self.attention(x)
+        x = self.c_proj(x)
+        return x[:, :, 0]
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(
+                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
+            )
+        else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+class TransposedUpsample(nn.Module):
+    'Learned 2x upsampling without padding'
+    def __init__(self, channels, out_channels=None, ks=5):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+
+        self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2)
+
+    def forward(self,x):
+        return self.up(x)
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, 3, padding=1
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        return checkpoint(
+            self._forward, (x, emb), self.parameters(), self.use_checkpoint
+        )
+
+
+    def _forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = th.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    An attention block that allows spatial positions to attend to each other.
+    Originally ported from here, but adapted to the N-d case.
+    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
+    """
+
+    def __init__(
+        self,
+        channels,
+        num_heads=1,
+        num_head_channels=-1,
+        use_checkpoint=False,
+        use_new_attention_order=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        if num_head_channels == -1:
+            self.num_heads = num_heads
+        else:
+            assert (
+                channels % num_head_channels == 0
+            ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
+            self.num_heads = channels // num_head_channels
+        self.use_checkpoint = use_checkpoint
+        self.norm = normalization(channels)
+        self.qkv = conv_nd(1, channels, channels * 3, 1)
+        if use_new_attention_order:
+            # split qkv before split heads
+            self.attention = QKVAttention(self.num_heads)
+        else:
+            # split heads before split qkv
+            self.attention = QKVAttentionLegacy(self.num_heads)
+
+        self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
+
+    def forward(self, x):
+        return checkpoint(self._forward, (x,), self.parameters(), True)   # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
+        #return pt_checkpoint(self._forward, x)  # pytorch
+
+    def _forward(self, x):
+        b, c, *spatial = x.shape
+        x = x.reshape(b, c, -1)
+        qkv = self.qkv(self.norm(x))
+        h = self.attention(qkv)
+        h = self.proj_out(h)
+        return (x + h).reshape(b, c, *spatial)
+
+
+def count_flops_attn(model, _x, y):
+    """
+    A counter for the `thop` package to count the operations in an
+    attention operation.
+    Meant to be used like:
+        macs, params = thop.profile(
+            model,
+            inputs=(inputs, timestamps),
+            custom_ops={QKVAttention: QKVAttention.count_flops},
+        )
+    """
+    b, c, *spatial = y[0].shape
+    num_spatial = int(np.prod(spatial))
+    # We perform two matmuls with the same number of ops.
+    # The first computes the weight matrix, the second computes
+    # the combination of the value vectors.
+    matmul_ops = 2 * b * (num_spatial ** 2) * c
+    model.total_ops += th.DoubleTensor([matmul_ops])
+
+
+class QKVAttentionLegacy(nn.Module):
+    """
+    A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts", q * scale, k * scale
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v)
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class QKVAttention(nn.Module):
+    """
+    A module which performs QKV attention and splits in a different order.
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.chunk(3, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts",
+            (q * scale).view(bs * self.n_heads, ch, length),
+            (k * scale).view(bs * self.n_heads, ch, length),
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        disable_self_attentions=None,
+        num_attention_blocks=None,
+        disable_middle_self_attn=False,
+        use_linear_in_transformer=False,
+    ):
+        super().__init__()
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            from omegaconf.listconfig import ListConfig
+            if type(context_dim) == ListConfig:
+                context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+            assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
+            print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
+                  f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
+                  f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
+                  f"attention will still not be set.")
+
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        if self.num_classes is not None:
+            if isinstance(self.num_classes, int):
+                self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+            elif self.num_classes == "continuous":
+                print("setting up linear c_adm embedding layer")
+                self.label_emb = nn.Linear(1, time_embed_dim)
+            else:
+                raise ValueError()
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
+                        layers.append(
+                            AttentionBlock(
+                                ch,
+                                use_checkpoint=use_checkpoint,
+                                num_heads=num_heads,
+                                num_head_channels=dim_head,
+                                use_new_attention_order=use_new_attention_order,
+                            ) if not use_spatial_transformer else SpatialTransformer(
+                                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
+                                use_checkpoint=use_checkpoint
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                use_new_attention_order=use_new_attention_order,
+            ) if not use_spatial_transformer else SpatialTransformer(  # always uses a self-attn
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                            disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
+                            use_checkpoint=use_checkpoint
+                        ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(self.num_res_blocks[level] + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
+                        layers.append(
+                            AttentionBlock(
+                                ch,
+                                use_checkpoint=use_checkpoint,
+                                num_heads=num_heads_upsample,
+                                num_head_channels=dim_head,
+                                use_new_attention_order=use_new_attention_order,
+                            ) if not use_spatial_transformer else SpatialTransformer(
+                                ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
+                                use_checkpoint=use_checkpoint
+                            )
+                        )
+                if level and i == self.num_res_blocks[level]:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+            normalization(ch),
+            conv_nd(dims, model_channels, n_embed, 1),
+            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+        )
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+        self.output_blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+        self.output_blocks.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps=None, context=None, y=None,**kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+
+        if self.num_classes is not None:
+            assert y.shape[0] == x.shape[0]
+            emb = emb + self.label_emb(y)
+
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb, context)
+            hs.append(h)
+        h = self.middle_block(h, emb, context)
+        for module in self.output_blocks:
+            h = th.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)

ldm/modules/diffusionmodules/upscaling.py

@@ -0,0 +1,81 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from functools import partial
+
+from ldm.modules.diffusionmodules.util import extract_into_tensor, make_beta_schedule
+from ldm.util import default
+
+
+class AbstractLowScaleModel(nn.Module):
+    # for concatenating a downsampled image to the latent representation
+    def __init__(self, noise_schedule_config=None):
+        super(AbstractLowScaleModel, self).__init__()
+        if noise_schedule_config is not None:
+            self.register_schedule(**noise_schedule_config)
+
+    def register_schedule(self, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                   cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+    def forward(self, x):
+        return x, None
+
+    def decode(self, x):
+        return x
+
+
+class SimpleImageConcat(AbstractLowScaleModel):
+    # no noise level conditioning
+    def __init__(self):
+        super(SimpleImageConcat, self).__init__(noise_schedule_config=None)
+        self.max_noise_level = 0
+
+    def forward(self, x):
+        # fix to constant noise level
+        return x, torch.zeros(x.shape[0], device=x.device).long()
+
+
+class ImageConcatWithNoiseAugmentation(AbstractLowScaleModel):
+    def __init__(self, noise_schedule_config, max_noise_level=1000, to_cuda=False):
+        super().__init__(noise_schedule_config=noise_schedule_config)
+        self.max_noise_level = max_noise_level
+
+    def forward(self, x, noise_level=None):
+        if noise_level is None:
+            noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long()
+        else:
+            assert isinstance(noise_level, torch.Tensor)
+        z = self.q_sample(x, noise_level)
+        return z, noise_level
+
+
+

ldm/modules/diffusionmodules/util.py

@@ -0,0 +1,270 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+
+import os
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import repeat
+
+from ldm.util import instantiate_from_config
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+
+def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
+    if ddim_discr_method == 'uniform':
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+    elif ddim_discr_method == 'quad':
+        ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
+    else:
+        raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    steps_out = ddim_timesteps + 1
+    if verbose:
+        print(f'Selected timesteps for ddim sampler: {steps_out}')
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
+    if verbose:
+        print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
+        print(f'For the chosen value of eta, which is {eta}, '
+              f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+        ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(),
+                                   "dtype": torch.get_autocast_gpu_dtype(),
+                                   "cache_enabled": torch.is_autocast_cache_enabled()}
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad(), \
+                torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs):
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def normalization(channels):
+    """
+    Make a standard normalization layer.
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    return GroupNorm32(32, channels)
+
+
+# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
+class SiLU(nn.Module):
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class GroupNorm32(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.float()).type(x.dtype)
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+class HybridConditioner(nn.Module):
+
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()

ldm/modules/distributions/distributions.py

@@ -0,0 +1,92 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

ldm/modules/ema.py

@@ -0,0 +1,80 @@
+import torch
+from torch import nn
+
+
+class LitEma(nn.Module):
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError('Decay must be between 0 and 1')
+
+        self.m_name2s_name = {}
+        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
+        self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates
+        else torch.tensor(-1, dtype=torch.int))
+
+        for name, p in model.named_parameters():
+            if p.requires_grad:
+                # remove as '.'-character is not allowed in buffers
+                s_name = name.replace('.', '')
+                self.m_name2s_name.update({name: s_name})
+                self.register_buffer(s_name, p.clone().detach().data)
+
+        self.collected_params = []
+
+    def reset_num_updates(self):
+        del self.num_updates
+        self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int))
+
+    def forward(self, model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with torch.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if m_param[key].requires_grad:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
+                    shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
+                else:
+                    assert not key in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if m_param[key].requires_grad:
+                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
+            else:
+                assert not key in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            param.data.copy_(c_param.data)

ldm/modules/encoders/modules.py

@@ -0,0 +1,213 @@
+import torch
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint
+
+from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel
+
+import open_clip
+from ldm.util import default, count_params
+
+
+class AbstractEncoder(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+class IdentityEncoder(AbstractEncoder):
+
+    def encode(self, x):
+        return x
+
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+        self.n_classes = n_classes
+        self.ucg_rate = ucg_rate
+
+    def forward(self, batch, key=None, disable_dropout=False):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]
+        if self.ucg_rate > 0. and not disable_dropout:
+            mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate)
+            c = mask * c + (1-mask) * torch.ones_like(c)*(self.n_classes-1)
+            c = c.long()
+        c = self.embedding(c)
+        return c
+
+    def get_unconditional_conditioning(self, bs, device="cuda"):
+        uc_class = self.n_classes - 1  # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000)
+        uc = torch.ones((bs,), device=device) * uc_class
+        uc = {self.key: uc}
+        return uc
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class FrozenT5Embedder(AbstractEncoder):
+    """Uses the T5 transformer encoder for text"""
+    def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, freeze=True):  # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
+        super().__init__()
+        self.tokenizer = T5Tokenizer.from_pretrained(version)
+        self.transformer = T5EncoderModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length   # TODO: typical value?
+        if freeze:
+            self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        #self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+
+class FrozenCLIPEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from huggingface)"""
+    LAYERS = [
+        "last",
+        "pooled",
+        "hidden"
+    ]
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77,
+                 freeze=True, layer="last", layer_idx=None):  # clip-vit-base-patch32
+        super().__init__()
+        assert layer in self.LAYERS
+        self.tokenizer = CLIPTokenizer.from_pretrained(version)
+        self.transformer = CLIPTextModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        self.layer_idx = layer_idx
+        if layer == "hidden":
+            assert layer_idx is not None
+            assert 0 <= abs(layer_idx) <= 12
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        #self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer=="hidden")
+        if self.layer == "last":
+            z = outputs.last_hidden_state
+        elif self.layer == "pooled":
+            z = outputs.pooler_output[:, None, :]
+        else:
+            z = outputs.hidden_states[self.layer_idx]
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+
+class FrozenOpenCLIPEmbedder(AbstractEncoder):
+    """
+    Uses the OpenCLIP transformer encoder for text
+    """
+    LAYERS = [
+        #"pooled",
+        "last",
+        "penultimate"
+    ]
+    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
+                 freeze=True, layer="last"):
+        super().__init__()
+        assert layer in self.LAYERS
+        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version)
+        del model.visual
+        self.model = model
+
+        self.device = device
+        self.max_length = max_length
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        if self.layer == "last":
+            self.layer_idx = 0
+        elif self.layer == "penultimate":
+            self.layer_idx = 1
+        else:
+            raise NotImplementedError()
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        tokens = open_clip.tokenize(text)
+        z = self.encode_with_transformer(tokens.to(self.device))
+        return z
+
+    def encode_with_transformer(self, text):
+        x = self.model.token_embedding(text)  # [batch_size, n_ctx, d_model]
+        x = x + self.model.positional_embedding
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.model.ln_final(x)
+        return x
+
+    def text_transformer_forward(self, x: torch.Tensor, attn_mask = None):
+        for i, r in enumerate(self.model.transformer.resblocks):
+            if i == len(self.model.transformer.resblocks) - self.layer_idx:
+                break
+            if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting():
+                x = checkpoint(r, x, attn_mask)
+            else:
+                x = r(x, attn_mask=attn_mask)
+        return x
+
+    def encode(self, text):
+        return self(text)
+
+
+class FrozenCLIPT5Encoder(AbstractEncoder):
+    def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda",
+                 clip_max_length=77, t5_max_length=77):
+        super().__init__()
+        self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length)
+        self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length)
+        print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder)*1.e-6:.2f} M parameters, "
+              f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder)*1.e-6:.2f} M params.")
+
+    def encode(self, text):
+        return self(text)
+
+    def forward(self, text):
+        clip_z = self.clip_encoder.encode(text)
+        t5_z = self.t5_encoder.encode(text)
+        return [clip_z, t5_z]
+
+

ldm/modules/image_degradation/__init__.py

@@ -0,0 +1,2 @@
+from ldm.modules.image_degradation.bsrgan import degradation_bsrgan_variant as degradation_fn_bsr
+from ldm.modules.image_degradation.bsrgan_light import degradation_bsrgan_variant as degradation_fn_bsr_light

ldm/modules/image_degradation/bsrgan.py

@@ -0,0 +1,730 @@
+# -*- coding: utf-8 -*-
+"""
+# --------------------------------------------
+# Super-Resolution
+# --------------------------------------------
+#
+# Kai Zhang (cskaizhang@gmail.com)
+# https://github.com/cszn
+# From 2019/03--2021/08
+# --------------------------------------------
+"""
+
+import numpy as np
+import cv2
+import torch
+
+from functools import partial
+import random
+from scipy import ndimage
+import scipy
+import scipy.stats as ss
+from scipy.interpolate import interp2d
+from scipy.linalg import orth
+import albumentations
+
+import ldm.modules.image_degradation.utils_image as util
+
+
+def modcrop_np(img, sf):
+    '''
+    Args:
+        img: numpy image, WxH or WxHxC
+        sf: scale factor
+    Return:
+        cropped image
+    '''
+    w, h = img.shape[:2]
+    im = np.copy(img)
+    return im[:w - w % sf, :h - h % sf, ...]
+
+
+"""
+# --------------------------------------------
+# anisotropic Gaussian kernels
+# --------------------------------------------
+"""
+
+
+def analytic_kernel(k):
+    """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
+    k_size = k.shape[0]
+    # Calculate the big kernels size
+    big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
+    # Loop over the small kernel to fill the big one
+    for r in range(k_size):
+        for c in range(k_size):
+            big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
+    # Crop the edges of the big kernel to ignore very small values and increase run time of SR
+    crop = k_size // 2
+    cropped_big_k = big_k[crop:-crop, crop:-crop]
+    # Normalize to 1
+    return cropped_big_k / cropped_big_k.sum()
+
+
+def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
+    """ generate an anisotropic Gaussian kernel
+    Args:
+        ksize : e.g., 15, kernel size
+        theta : [0,  pi], rotation angle range
+        l1    : [0.1,50], scaling of eigenvalues
+        l2    : [0.1,l1], scaling of eigenvalues
+        If l1 = l2, will get an isotropic Gaussian kernel.
+    Returns:
+        k     : kernel
+    """
+
+    v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
+    V = np.array([[v[0], v[1]], [v[1], -v[0]]])
+    D = np.array([[l1, 0], [0, l2]])
+    Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
+    k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
+
+    return k
+
+
+def gm_blur_kernel(mean, cov, size=15):
+    center = size / 2.0 + 0.5
+    k = np.zeros([size, size])
+    for y in range(size):
+        for x in range(size):
+            cy = y - center + 1
+            cx = x - center + 1
+            k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
+
+    k = k / np.sum(k)
+    return k
+
+
+def shift_pixel(x, sf, upper_left=True):
+    """shift pixel for super-resolution with different scale factors
+    Args:
+        x: WxHxC or WxH
+        sf: scale factor
+        upper_left: shift direction
+    """
+    h, w = x.shape[:2]
+    shift = (sf - 1) * 0.5
+    xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
+    if upper_left:
+        x1 = xv + shift
+        y1 = yv + shift
+    else:
+        x1 = xv - shift
+        y1 = yv - shift
+
+    x1 = np.clip(x1, 0, w - 1)
+    y1 = np.clip(y1, 0, h - 1)
+
+    if x.ndim == 2:
+        x = interp2d(xv, yv, x)(x1, y1)
+    if x.ndim == 3:
+        for i in range(x.shape[-1]):
+            x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
+
+    return x
+
+
+def blur(x, k):
+    '''
+    x: image, NxcxHxW
+    k: kernel, Nx1xhxw
+    '''
+    n, c = x.shape[:2]
+    p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
+    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
+    k = k.repeat(1, c, 1, 1)
+    k = k.view(-1, 1, k.shape[2], k.shape[3])
+    x = x.view(1, -1, x.shape[2], x.shape[3])
+    x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
+    x = x.view(n, c, x.shape[2], x.shape[3])
+
+    return x
+
+
+def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
+    """"
+    # modified version of https://github.com/assafshocher/BlindSR_dataset_generator
+    # Kai Zhang
+    # min_var = 0.175 * sf  # variance of the gaussian kernel will be sampled between min_var and max_var
+    # max_var = 2.5 * sf
+    """
+    # Set random eigen-vals (lambdas) and angle (theta) for COV matrix
+    lambda_1 = min_var + np.random.rand() * (max_var - min_var)
+    lambda_2 = min_var + np.random.rand() * (max_var - min_var)
+    theta = np.random.rand() * np.pi  # random theta
+    noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
+
+    # Set COV matrix using Lambdas and Theta
+    LAMBDA = np.diag([lambda_1, lambda_2])
+    Q = np.array([[np.cos(theta), -np.sin(theta)],
+                  [np.sin(theta), np.cos(theta)]])
+    SIGMA = Q @ LAMBDA @ Q.T
+    INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
+
+    # Set expectation position (shifting kernel for aligned image)
+    MU = k_size // 2 - 0.5 * (scale_factor - 1)  # - 0.5 * (scale_factor - k_size % 2)
+    MU = MU[None, None, :, None]
+
+    # Create meshgrid for Gaussian
+    [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
+    Z = np.stack([X, Y], 2)[:, :, :, None]
+
+    # Calcualte Gaussian for every pixel of the kernel
+    ZZ = Z - MU
+    ZZ_t = ZZ.transpose(0, 1, 3, 2)
+    raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
+
+    # shift the kernel so it will be centered
+    # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
+
+    # Normalize the kernel and return
+    # kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
+    kernel = raw_kernel / np.sum(raw_kernel)
+    return kernel
+
+
+def fspecial_gaussian(hsize, sigma):
+    hsize = [hsize, hsize]
+    siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
+    std = sigma
+    [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
+    arg = -(x * x + y * y) / (2 * std * std)
+    h = np.exp(arg)
+    h[h < scipy.finfo(float).eps * h.max()] = 0
+    sumh = h.sum()
+    if sumh != 0:
+        h = h / sumh
+    return h
+
+
+def fspecial_laplacian(alpha):
+    alpha = max([0, min([alpha, 1])])
+    h1 = alpha / (alpha + 1)
+    h2 = (1 - alpha) / (alpha + 1)
+    h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
+    h = np.array(h)
+    return h
+
+
+def fspecial(filter_type, *args, **kwargs):
+    '''
+    python code from:
+    https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
+    '''
+    if filter_type == 'gaussian':
+        return fspecial_gaussian(*args, **kwargs)
+    if filter_type == 'laplacian':
+        return fspecial_laplacian(*args, **kwargs)
+
+
+"""
+# --------------------------------------------
+# degradation models
+# --------------------------------------------
+"""
+
+
+def bicubic_degradation(x, sf=3):
+    '''
+    Args:
+        x: HxWxC image, [0, 1]
+        sf: down-scale factor
+    Return:
+        bicubicly downsampled LR image
+    '''
+    x = util.imresize_np(x, scale=1 / sf)
+    return x
+
+
+def srmd_degradation(x, k, sf=3):
+    ''' blur + bicubic downsampling
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2018learning,
+          title={Learning a single convolutional super-resolution network for multiple degradations},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={3262--3271},
+          year={2018}
+        }
+    '''
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')  # 'nearest' | 'mirror'
+    x = bicubic_degradation(x, sf=sf)
+    return x
+
+
+def dpsr_degradation(x, k, sf=3):
+    ''' bicubic downsampling + blur
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2019deep,
+          title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={1671--1681},
+          year={2019}
+        }
+    '''
+    x = bicubic_degradation(x, sf=sf)
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    return x
+
+
+def classical_degradation(x, k, sf=3):
+    ''' blur + downsampling
+    Args:
+        x: HxWxC image, [0, 1]/[0, 255]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    '''
+    x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
+    st = 0
+    return x[st::sf, st::sf, ...]
+
+
+def add_sharpening(img, weight=0.5, radius=50, threshold=10):
+    """USM sharpening. borrowed from real-ESRGAN
+    Input image: I; Blurry image: B.
+    1. K = I + weight * (I - B)
+    2. Mask = 1 if abs(I - B) > threshold, else: 0
+    3. Blur mask:
+    4. Out = Mask * K + (1 - Mask) * I
+    Args:
+        img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
+        weight (float): Sharp weight. Default: 1.
+        radius (float): Kernel size of Gaussian blur. Default: 50.
+        threshold (int):
+    """
+    if radius % 2 == 0:
+        radius += 1
+    blur = cv2.GaussianBlur(img, (radius, radius), 0)
+    residual = img - blur
+    mask = np.abs(residual) * 255 > threshold
+    mask = mask.astype('float32')
+    soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
+
+    K = img + weight * residual
+    K = np.clip(K, 0, 1)
+    return soft_mask * K + (1 - soft_mask) * img
+
+
+def add_blur(img, sf=4):
+    wd2 = 4.0 + sf
+    wd = 2.0 + 0.2 * sf
+    if random.random() < 0.5:
+        l1 = wd2 * random.random()
+        l2 = wd2 * random.random()
+        k = anisotropic_Gaussian(ksize=2 * random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
+    else:
+        k = fspecial('gaussian', 2 * random.randint(2, 11) + 3, wd * random.random())
+    img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
+
+    return img
+
+
+def add_resize(img, sf=4):
+    rnum = np.random.rand()
+    if rnum > 0.8:  # up
+        sf1 = random.uniform(1, 2)
+    elif rnum < 0.7:  # down
+        sf1 = random.uniform(0.5 / sf, 1)
+    else:
+        sf1 = 1.0
+    img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
+    img = np.clip(img, 0.0, 1.0)
+
+    return img
+
+
+# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+#     noise_level = random.randint(noise_level1, noise_level2)
+#     rnum = np.random.rand()
+#     if rnum > 0.6:  # add color Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+#     elif rnum < 0.4:  # add grayscale Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+#     else:  # add  noise
+#         L = noise_level2 / 255.
+#         D = np.diag(np.random.rand(3))
+#         U = orth(np.random.rand(3, 3))
+#         conv = np.dot(np.dot(np.transpose(U), D), U)
+#         img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+#     img = np.clip(img, 0.0, 1.0)
+#     return img
+
+def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    rnum = np.random.rand()
+    if rnum > 0.6:  # add color Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:  # add grayscale Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:  # add  noise
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_speckle_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    img = np.clip(img, 0.0, 1.0)
+    rnum = random.random()
+    if rnum > 0.6:
+        img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:
+        img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_Poisson_noise(img):
+    img = np.clip((img * 255.0).round(), 0, 255) / 255.
+    vals = 10 ** (2 * random.random() + 2.0)  # [2, 4]
+    if random.random() < 0.5:
+        img = np.random.poisson(img * vals).astype(np.float32) / vals
+    else:
+        img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
+        img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
+        noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
+        img += noise_gray[:, :, np.newaxis]
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_JPEG_noise(img):
+    quality_factor = random.randint(30, 95)
+    img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
+    result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
+    img = cv2.imdecode(encimg, 1)
+    img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
+    return img
+
+
+def random_crop(lq, hq, sf=4, lq_patchsize=64):
+    h, w = lq.shape[:2]
+    rnd_h = random.randint(0, h - lq_patchsize)
+    rnd_w = random.randint(0, w - lq_patchsize)
+    lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
+
+    rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
+    hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
+    return lq, hq
+
+
+def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = img.shape[:2]
+    img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = img.shape[:2]
+
+    if h < lq_patchsize * sf or w < lq_patchsize * sf:
+        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+
+    hq = img.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
+                             interpolation=random.choice([1, 2, 3]))
+        else:
+            img = util.imresize_np(img, 1 / 2, True)
+        img = np.clip(img, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            img = add_blur(img, sf=sf)
+
+        elif i == 1:
+            img = add_blur(img, sf=sf)
+
+        elif i == 2:
+            a, b = img.shape[1], img.shape[0]
+            # downsample2
+            if random.random() < 0.75:
+                sf1 = random.uniform(1, 2 * sf)
+                img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
+                                 interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                img = img[0::sf, 0::sf, ...]  # nearest downsampling
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                img = add_JPEG_noise(img)
+
+        elif i == 6:
+            # add processed camera sensor noise
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    img = add_JPEG_noise(img)
+
+    # random crop
+    img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
+
+    return img, hq
+
+
+# todo no isp_model?
+def degradation_bsrgan_variant(image, sf=4, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    image = util.uint2single(image)
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = image.shape[:2]
+    image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = image.shape[:2]
+
+    hq = image.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
+                               interpolation=random.choice([1, 2, 3]))
+        else:
+            image = util.imresize_np(image, 1 / 2, True)
+        image = np.clip(image, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            image = add_blur(image, sf=sf)
+
+        elif i == 1:
+            image = add_blur(image, sf=sf)
+
+        elif i == 2:
+            a, b = image.shape[1], image.shape[0]
+            # downsample2
+            if random.random() < 0.75:
+                sf1 = random.uniform(1, 2 * sf)
+                image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
+                                   interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                image = image[0::sf, 0::sf, ...]  # nearest downsampling
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            image = add_Gaussian_noise(image, noise_level1=2, noise_level2=25)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                image = add_JPEG_noise(image)
+
+        # elif i == 6:
+        #     # add processed camera sensor noise
+        #     if random.random() < isp_prob and isp_model is not None:
+        #         with torch.no_grad():
+        #             img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    image = add_JPEG_noise(image)
+    image = util.single2uint(image)
+    example = {"image":image}
+    return example
+
+
+# TODO incase there is a pickle error one needs to replace a += x with a = a + x in add_speckle_noise etc...
+def degradation_bsrgan_plus(img, sf=4, shuffle_prob=0.5, use_sharp=True, lq_patchsize=64, isp_model=None):
+    """
+    This is an extended degradation model by combining
+    the degradation models of BSRGAN and Real-ESRGAN
+    ----------
+    img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
+    sf: scale factor
+    use_shuffle: the degradation shuffle
+    use_sharp: sharpening the img
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+
+    h1, w1 = img.shape[:2]
+    img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = img.shape[:2]
+
+    if h < lq_patchsize * sf or w < lq_patchsize * sf:
+        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+
+    if use_sharp:
+        img = add_sharpening(img)
+    hq = img.copy()
+
+    if random.random() < shuffle_prob:
+        shuffle_order = random.sample(range(13), 13)
+    else:
+        shuffle_order = list(range(13))
+        # local shuffle for noise, JPEG is always the last one
+        shuffle_order[2:6] = random.sample(shuffle_order[2:6], len(range(2, 6)))
+        shuffle_order[9:13] = random.sample(shuffle_order[9:13], len(range(9, 13)))
+
+    poisson_prob, speckle_prob, isp_prob = 0.1, 0.1, 0.1
+
+    for i in shuffle_order:
+        if i == 0:
+            img = add_blur(img, sf=sf)
+        elif i == 1:
+            img = add_resize(img, sf=sf)
+        elif i == 2:
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
+        elif i == 3:
+            if random.random() < poisson_prob:
+                img = add_Poisson_noise(img)
+        elif i == 4:
+            if random.random() < speckle_prob:
+                img = add_speckle_noise(img)
+        elif i == 5:
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+        elif i == 6:
+            img = add_JPEG_noise(img)
+        elif i == 7:
+            img = add_blur(img, sf=sf)
+        elif i == 8:
+            img = add_resize(img, sf=sf)
+        elif i == 9:
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25)
+        elif i == 10:
+            if random.random() < poisson_prob:
+                img = add_Poisson_noise(img)
+        elif i == 11:
+            if random.random() < speckle_prob:
+                img = add_speckle_noise(img)
+        elif i == 12:
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+        else:
+            print('check the shuffle!')
+
+    # resize to desired size
+    img = cv2.resize(img, (int(1 / sf * hq.shape[1]), int(1 / sf * hq.shape[0])),
+                     interpolation=random.choice([1, 2, 3]))
+
+    # add final JPEG compression noise
+    img = add_JPEG_noise(img)
+
+    # random crop
+    img, hq = random_crop(img, hq, sf, lq_patchsize)
+
+    return img, hq
+
+
+if __name__ == '__main__':
+	print("hey")
+	img = util.imread_uint('utils/test.png', 3)
+	print(img)
+	img = util.uint2single(img)
+	print(img)
+	img = img[:448, :448]
+	h = img.shape[0] // 4
+	print("resizing to", h)
+	sf = 4
+	deg_fn = partial(degradation_bsrgan_variant, sf=sf)
+	for i in range(20):
+		print(i)
+		img_lq = deg_fn(img)
+		print(img_lq)
+		img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img)["image"]
+		print(img_lq.shape)
+		print("bicubic", img_lq_bicubic.shape)
+		print(img_hq.shape)
+		lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+		                        interpolation=0)
+		lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+		                        interpolation=0)
+		img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
+		util.imsave(img_concat, str(i) + '.png')
+
+

ldm/modules/image_degradation/bsrgan_light.py

@@ -0,0 +1,651 @@
+# -*- coding: utf-8 -*-
+import numpy as np
+import cv2
+import torch
+
+from functools import partial
+import random
+from scipy import ndimage
+import scipy
+import scipy.stats as ss
+from scipy.interpolate import interp2d
+from scipy.linalg import orth
+import albumentations
+
+import ldm.modules.image_degradation.utils_image as util
+
+"""
+# --------------------------------------------
+# Super-Resolution
+# --------------------------------------------
+#
+# Kai Zhang (cskaizhang@gmail.com)
+# https://github.com/cszn
+# From 2019/03--2021/08
+# --------------------------------------------
+"""
+
+def modcrop_np(img, sf):
+    '''
+    Args:
+        img: numpy image, WxH or WxHxC
+        sf: scale factor
+    Return:
+        cropped image
+    '''
+    w, h = img.shape[:2]
+    im = np.copy(img)
+    return im[:w - w % sf, :h - h % sf, ...]
+
+
+"""
+# --------------------------------------------
+# anisotropic Gaussian kernels
+# --------------------------------------------
+"""
+
+
+def analytic_kernel(k):
+    """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)"""
+    k_size = k.shape[0]
+    # Calculate the big kernels size
+    big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2))
+    # Loop over the small kernel to fill the big one
+    for r in range(k_size):
+        for c in range(k_size):
+            big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k
+    # Crop the edges of the big kernel to ignore very small values and increase run time of SR
+    crop = k_size // 2
+    cropped_big_k = big_k[crop:-crop, crop:-crop]
+    # Normalize to 1
+    return cropped_big_k / cropped_big_k.sum()
+
+
+def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6):
+    """ generate an anisotropic Gaussian kernel
+    Args:
+        ksize : e.g., 15, kernel size
+        theta : [0,  pi], rotation angle range
+        l1    : [0.1,50], scaling of eigenvalues
+        l2    : [0.1,l1], scaling of eigenvalues
+        If l1 = l2, will get an isotropic Gaussian kernel.
+    Returns:
+        k     : kernel
+    """
+
+    v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.]))
+    V = np.array([[v[0], v[1]], [v[1], -v[0]]])
+    D = np.array([[l1, 0], [0, l2]])
+    Sigma = np.dot(np.dot(V, D), np.linalg.inv(V))
+    k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize)
+
+    return k
+
+
+def gm_blur_kernel(mean, cov, size=15):
+    center = size / 2.0 + 0.5
+    k = np.zeros([size, size])
+    for y in range(size):
+        for x in range(size):
+            cy = y - center + 1
+            cx = x - center + 1
+            k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov)
+
+    k = k / np.sum(k)
+    return k
+
+
+def shift_pixel(x, sf, upper_left=True):
+    """shift pixel for super-resolution with different scale factors
+    Args:
+        x: WxHxC or WxH
+        sf: scale factor
+        upper_left: shift direction
+    """
+    h, w = x.shape[:2]
+    shift = (sf - 1) * 0.5
+    xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0)
+    if upper_left:
+        x1 = xv + shift
+        y1 = yv + shift
+    else:
+        x1 = xv - shift
+        y1 = yv - shift
+
+    x1 = np.clip(x1, 0, w - 1)
+    y1 = np.clip(y1, 0, h - 1)
+
+    if x.ndim == 2:
+        x = interp2d(xv, yv, x)(x1, y1)
+    if x.ndim == 3:
+        for i in range(x.shape[-1]):
+            x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1)
+
+    return x
+
+
+def blur(x, k):
+    '''
+    x: image, NxcxHxW
+    k: kernel, Nx1xhxw
+    '''
+    n, c = x.shape[:2]
+    p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2
+    x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate')
+    k = k.repeat(1, c, 1, 1)
+    k = k.view(-1, 1, k.shape[2], k.shape[3])
+    x = x.view(1, -1, x.shape[2], x.shape[3])
+    x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c)
+    x = x.view(n, c, x.shape[2], x.shape[3])
+
+    return x
+
+
+def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0):
+    """"
+    # modified version of https://github.com/assafshocher/BlindSR_dataset_generator
+    # Kai Zhang
+    # min_var = 0.175 * sf  # variance of the gaussian kernel will be sampled between min_var and max_var
+    # max_var = 2.5 * sf
+    """
+    # Set random eigen-vals (lambdas) and angle (theta) for COV matrix
+    lambda_1 = min_var + np.random.rand() * (max_var - min_var)
+    lambda_2 = min_var + np.random.rand() * (max_var - min_var)
+    theta = np.random.rand() * np.pi  # random theta
+    noise = -noise_level + np.random.rand(*k_size) * noise_level * 2
+
+    # Set COV matrix using Lambdas and Theta
+    LAMBDA = np.diag([lambda_1, lambda_2])
+    Q = np.array([[np.cos(theta), -np.sin(theta)],
+                  [np.sin(theta), np.cos(theta)]])
+    SIGMA = Q @ LAMBDA @ Q.T
+    INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :]
+
+    # Set expectation position (shifting kernel for aligned image)
+    MU = k_size // 2 - 0.5 * (scale_factor - 1)  # - 0.5 * (scale_factor - k_size % 2)
+    MU = MU[None, None, :, None]
+
+    # Create meshgrid for Gaussian
+    [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1]))
+    Z = np.stack([X, Y], 2)[:, :, :, None]
+
+    # Calcualte Gaussian for every pixel of the kernel
+    ZZ = Z - MU
+    ZZ_t = ZZ.transpose(0, 1, 3, 2)
+    raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise)
+
+    # shift the kernel so it will be centered
+    # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor)
+
+    # Normalize the kernel and return
+    # kernel = raw_kernel_centered / np.sum(raw_kernel_centered)
+    kernel = raw_kernel / np.sum(raw_kernel)
+    return kernel
+
+
+def fspecial_gaussian(hsize, sigma):
+    hsize = [hsize, hsize]
+    siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0]
+    std = sigma
+    [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1))
+    arg = -(x * x + y * y) / (2 * std * std)
+    h = np.exp(arg)
+    h[h < scipy.finfo(float).eps * h.max()] = 0
+    sumh = h.sum()
+    if sumh != 0:
+        h = h / sumh
+    return h
+
+
+def fspecial_laplacian(alpha):
+    alpha = max([0, min([alpha, 1])])
+    h1 = alpha / (alpha + 1)
+    h2 = (1 - alpha) / (alpha + 1)
+    h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]]
+    h = np.array(h)
+    return h
+
+
+def fspecial(filter_type, *args, **kwargs):
+    '''
+    python code from:
+    https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py
+    '''
+    if filter_type == 'gaussian':
+        return fspecial_gaussian(*args, **kwargs)
+    if filter_type == 'laplacian':
+        return fspecial_laplacian(*args, **kwargs)
+
+
+"""
+# --------------------------------------------
+# degradation models
+# --------------------------------------------
+"""
+
+
+def bicubic_degradation(x, sf=3):
+    '''
+    Args:
+        x: HxWxC image, [0, 1]
+        sf: down-scale factor
+    Return:
+        bicubicly downsampled LR image
+    '''
+    x = util.imresize_np(x, scale=1 / sf)
+    return x
+
+
+def srmd_degradation(x, k, sf=3):
+    ''' blur + bicubic downsampling
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2018learning,
+          title={Learning a single convolutional super-resolution network for multiple degradations},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={3262--3271},
+          year={2018}
+        }
+    '''
+    x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap')  # 'nearest' | 'mirror'
+    x = bicubic_degradation(x, sf=sf)
+    return x
+
+
+def dpsr_degradation(x, k, sf=3):
+    ''' bicubic downsampling + blur
+    Args:
+        x: HxWxC image, [0, 1]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    Reference:
+        @inproceedings{zhang2019deep,
+          title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels},
+          author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
+          booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
+          pages={1671--1681},
+          year={2019}
+        }
+    '''
+    x = bicubic_degradation(x, sf=sf)
+    x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    return x
+
+
+def classical_degradation(x, k, sf=3):
+    ''' blur + downsampling
+    Args:
+        x: HxWxC image, [0, 1]/[0, 255]
+        k: hxw, double
+        sf: down-scale factor
+    Return:
+        downsampled LR image
+    '''
+    x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap')
+    # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2))
+    st = 0
+    return x[st::sf, st::sf, ...]
+
+
+def add_sharpening(img, weight=0.5, radius=50, threshold=10):
+    """USM sharpening. borrowed from real-ESRGAN
+    Input image: I; Blurry image: B.
+    1. K = I + weight * (I - B)
+    2. Mask = 1 if abs(I - B) > threshold, else: 0
+    3. Blur mask:
+    4. Out = Mask * K + (1 - Mask) * I
+    Args:
+        img (Numpy array): Input image, HWC, BGR; float32, [0, 1].
+        weight (float): Sharp weight. Default: 1.
+        radius (float): Kernel size of Gaussian blur. Default: 50.
+        threshold (int):
+    """
+    if radius % 2 == 0:
+        radius += 1
+    blur = cv2.GaussianBlur(img, (radius, radius), 0)
+    residual = img - blur
+    mask = np.abs(residual) * 255 > threshold
+    mask = mask.astype('float32')
+    soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0)
+
+    K = img + weight * residual
+    K = np.clip(K, 0, 1)
+    return soft_mask * K + (1 - soft_mask) * img
+
+
+def add_blur(img, sf=4):
+    wd2 = 4.0 + sf
+    wd = 2.0 + 0.2 * sf
+
+    wd2 = wd2/4
+    wd = wd/4
+
+    if random.random() < 0.5:
+        l1 = wd2 * random.random()
+        l2 = wd2 * random.random()
+        k = anisotropic_Gaussian(ksize=random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2)
+    else:
+        k = fspecial('gaussian', random.randint(2, 4) + 3, wd * random.random())
+    img = ndimage.convolve(img, np.expand_dims(k, axis=2), mode='mirror')
+
+    return img
+
+
+def add_resize(img, sf=4):
+    rnum = np.random.rand()
+    if rnum > 0.8:  # up
+        sf1 = random.uniform(1, 2)
+    elif rnum < 0.7:  # down
+        sf1 = random.uniform(0.5 / sf, 1)
+    else:
+        sf1 = 1.0
+    img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3]))
+    img = np.clip(img, 0.0, 1.0)
+
+    return img
+
+
+# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+#     noise_level = random.randint(noise_level1, noise_level2)
+#     rnum = np.random.rand()
+#     if rnum > 0.6:  # add color Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+#     elif rnum < 0.4:  # add grayscale Gaussian noise
+#         img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+#     else:  # add  noise
+#         L = noise_level2 / 255.
+#         D = np.diag(np.random.rand(3))
+#         U = orth(np.random.rand(3, 3))
+#         conv = np.dot(np.dot(np.transpose(U), D), U)
+#         img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+#     img = np.clip(img, 0.0, 1.0)
+#     return img
+
+def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    rnum = np.random.rand()
+    if rnum > 0.6:  # add color Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:  # add grayscale Gaussian noise
+        img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:  # add  noise
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_speckle_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    img = np.clip(img, 0.0, 1.0)
+    rnum = random.random()
+    if rnum > 0.6:
+        img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4:
+        img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:
+        L = noise_level2 / 255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3, 3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_Poisson_noise(img):
+    img = np.clip((img * 255.0).round(), 0, 255) / 255.
+    vals = 10 ** (2 * random.random() + 2.0)  # [2, 4]
+    if random.random() < 0.5:
+        img = np.random.poisson(img * vals).astype(np.float32) / vals
+    else:
+        img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114])
+        img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255.
+        noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray
+        img += noise_gray[:, :, np.newaxis]
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+
+def add_JPEG_noise(img):
+    quality_factor = random.randint(80, 95)
+    img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR)
+    result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
+    img = cv2.imdecode(encimg, 1)
+    img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB)
+    return img
+
+
+def random_crop(lq, hq, sf=4, lq_patchsize=64):
+    h, w = lq.shape[:2]
+    rnd_h = random.randint(0, h - lq_patchsize)
+    rnd_w = random.randint(0, w - lq_patchsize)
+    lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :]
+
+    rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf)
+    hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :]
+    return lq, hq
+
+
+def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf)
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = img.shape[:2]
+    img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = img.shape[:2]
+
+    if h < lq_patchsize * sf or w < lq_patchsize * sf:
+        raise ValueError(f'img size ({h1}X{w1}) is too small!')
+
+    hq = img.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])),
+                             interpolation=random.choice([1, 2, 3]))
+        else:
+            img = util.imresize_np(img, 1 / 2, True)
+        img = np.clip(img, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            img = add_blur(img, sf=sf)
+
+        elif i == 1:
+            img = add_blur(img, sf=sf)
+
+        elif i == 2:
+            a, b = img.shape[1], img.shape[0]
+            # downsample2
+            if random.random() < 0.75:
+                sf1 = random.uniform(1, 2 * sf)
+                img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])),
+                                 interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                img = ndimage.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                img = img[0::sf, 0::sf, ...]  # nearest downsampling
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            img = np.clip(img, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            img = add_Gaussian_noise(img, noise_level1=2, noise_level2=8)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                img = add_JPEG_noise(img)
+
+        elif i == 6:
+            # add processed camera sensor noise
+            if random.random() < isp_prob and isp_model is not None:
+                with torch.no_grad():
+                    img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    img = add_JPEG_noise(img)
+
+    # random crop
+    img, hq = random_crop(img, hq, sf_ori, lq_patchsize)
+
+    return img, hq
+
+
+# todo no isp_model?
+def degradation_bsrgan_variant(image, sf=4, isp_model=None, up=False):
+    """
+    This is the degradation model of BSRGAN from the paper
+    "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution"
+    ----------
+    sf: scale factor
+    isp_model: camera ISP model
+    Returns
+    -------
+    img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1]
+    hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1]
+    """
+    image = util.uint2single(image)
+    isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25
+    sf_ori = sf
+
+    h1, w1 = image.shape[:2]
+    image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...]  # mod crop
+    h, w = image.shape[:2]
+
+    hq = image.copy()
+
+    if sf == 4 and random.random() < scale2_prob:  # downsample1
+        if np.random.rand() < 0.5:
+            image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])),
+                               interpolation=random.choice([1, 2, 3]))
+        else:
+            image = util.imresize_np(image, 1 / 2, True)
+        image = np.clip(image, 0.0, 1.0)
+        sf = 2
+
+    shuffle_order = random.sample(range(7), 7)
+    idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3)
+    if idx1 > idx2:  # keep downsample3 last
+        shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1]
+
+    for i in shuffle_order:
+
+        if i == 0:
+            image = add_blur(image, sf=sf)
+
+        # elif i == 1:
+        #     image = add_blur(image, sf=sf)
+
+        if i == 0:
+            pass
+
+        elif i == 2:
+            a, b = image.shape[1], image.shape[0]
+            # downsample2
+            if random.random() < 0.8:
+                sf1 = random.uniform(1, 2 * sf)
+                image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])),
+                                   interpolation=random.choice([1, 2, 3]))
+            else:
+                k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf))
+                k_shifted = shift_pixel(k, sf)
+                k_shifted = k_shifted / k_shifted.sum()  # blur with shifted kernel
+                image = ndimage.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror')
+                image = image[0::sf, 0::sf, ...]  # nearest downsampling
+
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 3:
+            # downsample3
+            image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3]))
+            image = np.clip(image, 0.0, 1.0)
+
+        elif i == 4:
+            # add Gaussian noise
+            image = add_Gaussian_noise(image, noise_level1=1, noise_level2=2)
+
+        elif i == 5:
+            # add JPEG noise
+            if random.random() < jpeg_prob:
+                image = add_JPEG_noise(image)
+        #
+        # elif i == 6:
+        #     # add processed camera sensor noise
+        #     if random.random() < isp_prob and isp_model is not None:
+        #         with torch.no_grad():
+        #             img, hq = isp_model.forward(img.copy(), hq)
+
+    # add final JPEG compression noise
+    image = add_JPEG_noise(image)
+    image = util.single2uint(image)
+    if up:
+        image = cv2.resize(image, (w1, h1), interpolation=cv2.INTER_CUBIC)  # todo: random, as above? want to condition on it then
+    example = {"image": image}
+    return example
+
+
+
+
+if __name__ == '__main__':
+    print("hey")
+    img = util.imread_uint('utils/test.png', 3)
+    img = img[:448, :448]
+    h = img.shape[0] // 4
+    print("resizing to", h)
+    sf = 4
+    deg_fn = partial(degradation_bsrgan_variant, sf=sf)
+    for i in range(20):
+        print(i)
+        img_hq = img
+        img_lq = deg_fn(img)["image"]
+        img_hq, img_lq = util.uint2single(img_hq), util.uint2single(img_lq)
+        print(img_lq)
+        img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img_hq)["image"]
+        print(img_lq.shape)
+        print("bicubic", img_lq_bicubic.shape)
+        print(img_hq.shape)
+        lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+                                interpolation=0)
+        lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic),
+                                        (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])),
+                                        interpolation=0)
+        img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1)
+        util.imsave(img_concat, str(i) + '.png')

ldm/modules/image_degradation/utils_image.py

@@ -0,0 +1,916 @@
+import os
+import math
+import random
+import numpy as np
+import torch
+import cv2
+from torchvision.utils import make_grid
+from datetime import datetime
+#import matplotlib.pyplot as plt   # TODO: check with Dominik, also bsrgan.py vs bsrgan_light.py
+
+
+os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
+
+
+'''
+# --------------------------------------------
+# Kai Zhang (github: https://github.com/cszn)
+# 03/Mar/2019
+# --------------------------------------------
+# https://github.com/twhui/SRGAN-pyTorch
+# https://github.com/xinntao/BasicSR
+# --------------------------------------------
+'''
+
+
+IMG_EXTENSIONS = ['.jpg', '.JPG', '.jpeg', '.JPEG', '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', '.tif']
+
+
+def is_image_file(filename):
+    return any(filename.endswith(extension) for extension in IMG_EXTENSIONS)
+
+
+def get_timestamp():
+    return datetime.now().strftime('%y%m%d-%H%M%S')
+
+
+def imshow(x, title=None, cbar=False, figsize=None):
+    plt.figure(figsize=figsize)
+    plt.imshow(np.squeeze(x), interpolation='nearest', cmap='gray')
+    if title:
+        plt.title(title)
+    if cbar:
+        plt.colorbar()
+    plt.show()
+
+
+def surf(Z, cmap='rainbow', figsize=None):
+    plt.figure(figsize=figsize)
+    ax3 = plt.axes(projection='3d')
+
+    w, h = Z.shape[:2]
+    xx = np.arange(0,w,1)
+    yy = np.arange(0,h,1)
+    X, Y = np.meshgrid(xx, yy)
+    ax3.plot_surface(X,Y,Z,cmap=cmap)
+    #ax3.contour(X,Y,Z, zdim='z',offset=-2，cmap=cmap)
+    plt.show()
+
+
+'''
+# --------------------------------------------
+# get image pathes
+# --------------------------------------------
+'''
+
+
+def get_image_paths(dataroot):
+    paths = None  # return None if dataroot is None
+    if dataroot is not None:
+        paths = sorted(_get_paths_from_images(dataroot))
+    return paths
+
+
+def _get_paths_from_images(path):
+    assert os.path.isdir(path), '{:s} is not a valid directory'.format(path)
+    images = []
+    for dirpath, _, fnames in sorted(os.walk(path)):
+        for fname in sorted(fnames):
+            if is_image_file(fname):
+                img_path = os.path.join(dirpath, fname)
+                images.append(img_path)
+    assert images, '{:s} has no valid image file'.format(path)
+    return images
+
+
+'''
+# --------------------------------------------
+# split large images into small images 
+# --------------------------------------------
+'''
+
+
+def patches_from_image(img, p_size=512, p_overlap=64, p_max=800):
+    w, h = img.shape[:2]
+    patches = []
+    if w > p_max and h > p_max:
+        w1 = list(np.arange(0, w-p_size, p_size-p_overlap, dtype=np.int))
+        h1 = list(np.arange(0, h-p_size, p_size-p_overlap, dtype=np.int))
+        w1.append(w-p_size)
+        h1.append(h-p_size)
+#        print(w1)
+#        print(h1)
+        for i in w1:
+            for j in h1:
+                patches.append(img[i:i+p_size, j:j+p_size,:])
+    else:
+        patches.append(img)
+
+    return patches
+
+
+def imssave(imgs, img_path):
+    """
+    imgs: list, N images of size WxHxC
+    """
+    img_name, ext = os.path.splitext(os.path.basename(img_path))
+
+    for i, img in enumerate(imgs):
+        if img.ndim == 3:
+            img = img[:, :, [2, 1, 0]]
+        new_path = os.path.join(os.path.dirname(img_path), img_name+str('_s{:04d}'.format(i))+'.png')
+        cv2.imwrite(new_path, img)
+
+
+def split_imageset(original_dataroot, taget_dataroot, n_channels=3, p_size=800, p_overlap=96, p_max=1000):
+    """
+    split the large images from original_dataroot into small overlapped images with size (p_size)x(p_size),
+    and save them into taget_dataroot; only the images with larger size than (p_max)x(p_max)
+    will be splitted.
+    Args:
+        original_dataroot:
+        taget_dataroot:
+        p_size: size of small images
+        p_overlap: patch size in training is a good choice
+        p_max: images with smaller size than (p_max)x(p_max) keep unchanged.
+    """
+    paths = get_image_paths(original_dataroot)
+    for img_path in paths:
+        # img_name, ext = os.path.splitext(os.path.basename(img_path))
+        img = imread_uint(img_path, n_channels=n_channels)
+        patches = patches_from_image(img, p_size, p_overlap, p_max)
+        imssave(patches, os.path.join(taget_dataroot,os.path.basename(img_path)))
+        #if original_dataroot == taget_dataroot:
+        #del img_path
+
+'''
+# --------------------------------------------
+# makedir
+# --------------------------------------------
+'''
+
+
+def mkdir(path):
+    if not os.path.exists(path):
+        os.makedirs(path)
+
+
+def mkdirs(paths):
+    if isinstance(paths, str):
+        mkdir(paths)
+    else:
+        for path in paths:
+            mkdir(path)
+
+
+def mkdir_and_rename(path):
+    if os.path.exists(path):
+        new_name = path + '_archived_' + get_timestamp()
+        print('Path already exists. Rename it to [{:s}]'.format(new_name))
+        os.rename(path, new_name)
+    os.makedirs(path)
+
+
+'''
+# --------------------------------------------
+# read image from path
+# opencv is fast, but read BGR numpy image
+# --------------------------------------------
+'''
+
+
+# --------------------------------------------
+# get uint8 image of size HxWxn_channles (RGB)
+# --------------------------------------------
+def imread_uint(path, n_channels=3):
+    #  input: path
+    # output: HxWx3(RGB or GGG), or HxWx1 (G)
+    if n_channels == 1:
+        img = cv2.imread(path, 0)  # cv2.IMREAD_GRAYSCALE
+        img = np.expand_dims(img, axis=2)  # HxWx1
+    elif n_channels == 3:
+        img = cv2.imread(path, cv2.IMREAD_UNCHANGED)  # BGR or G
+        if img.ndim == 2:
+            img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)  # GGG
+        else:
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)  # RGB
+    return img
+
+
+# --------------------------------------------
+# matlab's imwrite
+# --------------------------------------------
+def imsave(img, img_path):
+    img = np.squeeze(img)
+    if img.ndim == 3:
+        img = img[:, :, [2, 1, 0]]
+    cv2.imwrite(img_path, img)
+
+def imwrite(img, img_path):
+    img = np.squeeze(img)
+    if img.ndim == 3:
+        img = img[:, :, [2, 1, 0]]
+    cv2.imwrite(img_path, img)
+
+
+
+# --------------------------------------------
+# get single image of size HxWxn_channles (BGR)
+# --------------------------------------------
+def read_img(path):
+    # read image by cv2
+    # return: Numpy float32, HWC, BGR, [0,1]
+    img = cv2.imread(path, cv2.IMREAD_UNCHANGED)  # cv2.IMREAD_GRAYSCALE
+    img = img.astype(np.float32) / 255.
+    if img.ndim == 2:
+        img = np.expand_dims(img, axis=2)
+    # some images have 4 channels
+    if img.shape[2] > 3:
+        img = img[:, :, :3]
+    return img
+
+
+'''
+# --------------------------------------------
+# image format conversion
+# --------------------------------------------
+# numpy(single) <--->  numpy(unit)
+# numpy(single) <--->  tensor
+# numpy(unit)   <--->  tensor
+# --------------------------------------------
+'''
+
+
+# --------------------------------------------
+# numpy(single) [0, 1] <--->  numpy(unit)
+# --------------------------------------------
+
+
+def uint2single(img):
+
+    return np.float32(img/255.)
+
+
+def single2uint(img):
+
+    return np.uint8((img.clip(0, 1)*255.).round())
+
+
+def uint162single(img):
+
+    return np.float32(img/65535.)
+
+
+def single2uint16(img):
+
+    return np.uint16((img.clip(0, 1)*65535.).round())
+
+
+# --------------------------------------------
+# numpy(unit) (HxWxC or HxW) <--->  tensor
+# --------------------------------------------
+
+
+# convert uint to 4-dimensional torch tensor
+def uint2tensor4(img):
+    if img.ndim == 2:
+        img = np.expand_dims(img, axis=2)
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.).unsqueeze(0)
+
+
+# convert uint to 3-dimensional torch tensor
+def uint2tensor3(img):
+    if img.ndim == 2:
+        img = np.expand_dims(img, axis=2)
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.)
+
+
+# convert 2/3/4-dimensional torch tensor to uint
+def tensor2uint(img):
+    img = img.data.squeeze().float().clamp_(0, 1).cpu().numpy()
+    if img.ndim == 3:
+        img = np.transpose(img, (1, 2, 0))
+    return np.uint8((img*255.0).round())
+
+
+# --------------------------------------------
+# numpy(single) (HxWxC) <--->  tensor
+# --------------------------------------------
+
+
+# convert single (HxWxC) to 3-dimensional torch tensor
+def single2tensor3(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float()
+
+
+# convert single (HxWxC) to 4-dimensional torch tensor
+def single2tensor4(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().unsqueeze(0)
+
+
+# convert torch tensor to single
+def tensor2single(img):
+    img = img.data.squeeze().float().cpu().numpy()
+    if img.ndim == 3:
+        img = np.transpose(img, (1, 2, 0))
+
+    return img
+
+# convert torch tensor to single
+def tensor2single3(img):
+    img = img.data.squeeze().float().cpu().numpy()
+    if img.ndim == 3:
+        img = np.transpose(img, (1, 2, 0))
+    elif img.ndim == 2:
+        img = np.expand_dims(img, axis=2)
+    return img
+
+
+def single2tensor5(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float().unsqueeze(0)
+
+
+def single32tensor5(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).float().unsqueeze(0).unsqueeze(0)
+
+
+def single42tensor4(img):
+    return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float()
+
+
+# from skimage.io import imread, imsave
+def tensor2img(tensor, out_type=np.uint8, min_max=(0, 1)):
+    '''
+    Converts a torch Tensor into an image Numpy array of BGR channel order
+    Input: 4D(B,(3/1),H,W), 3D(C,H,W), or 2D(H,W), any range, RGB channel order
+    Output: 3D(H,W,C) or 2D(H,W), [0,255], np.uint8 (default)
+    '''
+    tensor = tensor.squeeze().float().cpu().clamp_(*min_max)  # squeeze first, then clamp
+    tensor = (tensor - min_max[0]) / (min_max[1] - min_max[0])  # to range [0,1]
+    n_dim = tensor.dim()
+    if n_dim == 4:
+        n_img = len(tensor)
+        img_np = make_grid(tensor, nrow=int(math.sqrt(n_img)), normalize=False).numpy()
+        img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0))  # HWC, BGR
+    elif n_dim == 3:
+        img_np = tensor.numpy()
+        img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0))  # HWC, BGR
+    elif n_dim == 2:
+        img_np = tensor.numpy()
+    else:
+        raise TypeError(
+            'Only support 4D, 3D and 2D tensor. But received with dimension: {:d}'.format(n_dim))
+    if out_type == np.uint8:
+        img_np = (img_np * 255.0).round()
+        # Important. Unlike matlab, numpy.unit8() WILL NOT round by default.
+    return img_np.astype(out_type)
+
+
+'''
+# --------------------------------------------
+# Augmentation, flipe and/or rotate
+# --------------------------------------------
+# The following two are enough.
+# (1) augmet_img: numpy image of WxHxC or WxH
+# (2) augment_img_tensor4: tensor image 1xCxWxH
+# --------------------------------------------
+'''
+
+
+def augment_img(img, mode=0):
+    '''Kai Zhang (github: https://github.com/cszn)
+    '''
+    if mode == 0:
+        return img
+    elif mode == 1:
+        return np.flipud(np.rot90(img))
+    elif mode == 2:
+        return np.flipud(img)
+    elif mode == 3:
+        return np.rot90(img, k=3)
+    elif mode == 4:
+        return np.flipud(np.rot90(img, k=2))
+    elif mode == 5:
+        return np.rot90(img)
+    elif mode == 6:
+        return np.rot90(img, k=2)
+    elif mode == 7:
+        return np.flipud(np.rot90(img, k=3))
+
+
+def augment_img_tensor4(img, mode=0):
+    '''Kai Zhang (github: https://github.com/cszn)
+    '''
+    if mode == 0:
+        return img
+    elif mode == 1:
+        return img.rot90(1, [2, 3]).flip([2])
+    elif mode == 2:
+        return img.flip([2])
+    elif mode == 3:
+        return img.rot90(3, [2, 3])
+    elif mode == 4:
+        return img.rot90(2, [2, 3]).flip([2])
+    elif mode == 5:
+        return img.rot90(1, [2, 3])
+    elif mode == 6:
+        return img.rot90(2, [2, 3])
+    elif mode == 7:
+        return img.rot90(3, [2, 3]).flip([2])
+
+
+def augment_img_tensor(img, mode=0):
+    '''Kai Zhang (github: https://github.com/cszn)
+    '''
+    img_size = img.size()
+    img_np = img.data.cpu().numpy()
+    if len(img_size) == 3:
+        img_np = np.transpose(img_np, (1, 2, 0))
+    elif len(img_size) == 4:
+        img_np = np.transpose(img_np, (2, 3, 1, 0))
+    img_np = augment_img(img_np, mode=mode)
+    img_tensor = torch.from_numpy(np.ascontiguousarray(img_np))
+    if len(img_size) == 3:
+        img_tensor = img_tensor.permute(2, 0, 1)
+    elif len(img_size) == 4:
+        img_tensor = img_tensor.permute(3, 2, 0, 1)
+
+    return img_tensor.type_as(img)
+
+
+def augment_img_np3(img, mode=0):
+    if mode == 0:
+        return img
+    elif mode == 1:
+        return img.transpose(1, 0, 2)
+    elif mode == 2:
+        return img[::-1, :, :]
+    elif mode == 3:
+        img = img[::-1, :, :]
+        img = img.transpose(1, 0, 2)
+        return img
+    elif mode == 4:
+        return img[:, ::-1, :]
+    elif mode == 5:
+        img = img[:, ::-1, :]
+        img = img.transpose(1, 0, 2)
+        return img
+    elif mode == 6:
+        img = img[:, ::-1, :]
+        img = img[::-1, :, :]
+        return img
+    elif mode == 7:
+        img = img[:, ::-1, :]
+        img = img[::-1, :, :]
+        img = img.transpose(1, 0, 2)
+        return img
+
+
+def augment_imgs(img_list, hflip=True, rot=True):
+    # horizontal flip OR rotate
+    hflip = hflip and random.random() < 0.5
+    vflip = rot and random.random() < 0.5
+    rot90 = rot and random.random() < 0.5
+
+    def _augment(img):
+        if hflip:
+            img = img[:, ::-1, :]
+        if vflip:
+            img = img[::-1, :, :]
+        if rot90:
+            img = img.transpose(1, 0, 2)
+        return img
+
+    return [_augment(img) for img in img_list]
+
+
+'''
+# --------------------------------------------
+# modcrop and shave
+# --------------------------------------------
+'''
+
+
+def modcrop(img_in, scale):
+    # img_in: Numpy, HWC or HW
+    img = np.copy(img_in)
+    if img.ndim == 2:
+        H, W = img.shape
+        H_r, W_r = H % scale, W % scale
+        img = img[:H - H_r, :W - W_r]
+    elif img.ndim == 3:
+        H, W, C = img.shape
+        H_r, W_r = H % scale, W % scale
+        img = img[:H - H_r, :W - W_r, :]
+    else:
+        raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim))
+    return img
+
+
+def shave(img_in, border=0):
+    # img_in: Numpy, HWC or HW
+    img = np.copy(img_in)
+    h, w = img.shape[:2]
+    img = img[border:h-border, border:w-border]
+    return img
+
+
+'''
+# --------------------------------------------
+# image processing process on numpy image
+# channel_convert(in_c, tar_type, img_list):
+# rgb2ycbcr(img, only_y=True):
+# bgr2ycbcr(img, only_y=True):
+# ycbcr2rgb(img):
+# --------------------------------------------
+'''
+
+
+def rgb2ycbcr(img, only_y=True):
+    '''same as matlab rgb2ycbcr
+    only_y: only return Y channel
+    Input:
+        uint8, [0, 255]
+        float, [0, 1]
+    '''
+    in_img_type = img.dtype
+    img.astype(np.float32)
+    if in_img_type != np.uint8:
+        img *= 255.
+    # convert
+    if only_y:
+        rlt = np.dot(img, [65.481, 128.553, 24.966]) / 255.0 + 16.0
+    else:
+        rlt = np.matmul(img, [[65.481, -37.797, 112.0], [128.553, -74.203, -93.786],
+                              [24.966, 112.0, -18.214]]) / 255.0 + [16, 128, 128]
+    if in_img_type == np.uint8:
+        rlt = rlt.round()
+    else:
+        rlt /= 255.
+    return rlt.astype(in_img_type)
+
+
+def ycbcr2rgb(img):
+    '''same as matlab ycbcr2rgb
+    Input:
+        uint8, [0, 255]
+        float, [0, 1]
+    '''
+    in_img_type = img.dtype
+    img.astype(np.float32)
+    if in_img_type != np.uint8:
+        img *= 255.
+    # convert
+    rlt = np.matmul(img, [[0.00456621, 0.00456621, 0.00456621], [0, -0.00153632, 0.00791071],
+                          [0.00625893, -0.00318811, 0]]) * 255.0 + [-222.921, 135.576, -276.836]
+    if in_img_type == np.uint8:
+        rlt = rlt.round()
+    else:
+        rlt /= 255.
+    return rlt.astype(in_img_type)
+
+
+def bgr2ycbcr(img, only_y=True):
+    '''bgr version of rgb2ycbcr
+    only_y: only return Y channel
+    Input:
+        uint8, [0, 255]
+        float, [0, 1]
+    '''
+    in_img_type = img.dtype
+    img.astype(np.float32)
+    if in_img_type != np.uint8:
+        img *= 255.
+    # convert
+    if only_y:
+        rlt = np.dot(img, [24.966, 128.553, 65.481]) / 255.0 + 16.0
+    else:
+        rlt = np.matmul(img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786],
+                              [65.481, -37.797, 112.0]]) / 255.0 + [16, 128, 128]
+    if in_img_type == np.uint8:
+        rlt = rlt.round()
+    else:
+        rlt /= 255.
+    return rlt.astype(in_img_type)
+
+
+def channel_convert(in_c, tar_type, img_list):
+    # conversion among BGR, gray and y
+    if in_c == 3 and tar_type == 'gray':  # BGR to gray
+        gray_list = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img in img_list]
+        return [np.expand_dims(img, axis=2) for img in gray_list]
+    elif in_c == 3 and tar_type == 'y':  # BGR to y
+        y_list = [bgr2ycbcr(img, only_y=True) for img in img_list]
+        return [np.expand_dims(img, axis=2) for img in y_list]
+    elif in_c == 1 and tar_type == 'RGB':  # gray/y to BGR
+        return [cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) for img in img_list]
+    else:
+        return img_list
+
+
+'''
+# --------------------------------------------
+# metric, PSNR and SSIM
+# --------------------------------------------
+'''
+
+
+# --------------------------------------------
+# PSNR
+# --------------------------------------------
+def calculate_psnr(img1, img2, border=0):
+    # img1 and img2 have range [0, 255]
+    #img1 = img1.squeeze()
+    #img2 = img2.squeeze()
+    if not img1.shape == img2.shape:
+        raise ValueError('Input images must have the same dimensions.')
+    h, w = img1.shape[:2]
+    img1 = img1[border:h-border, border:w-border]
+    img2 = img2[border:h-border, border:w-border]
+
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    mse = np.mean((img1 - img2)**2)
+    if mse == 0:
+        return float('inf')
+    return 20 * math.log10(255.0 / math.sqrt(mse))
+
+
+# --------------------------------------------
+# SSIM
+# --------------------------------------------
+def calculate_ssim(img1, img2, border=0):
+    '''calculate SSIM
+    the same outputs as MATLAB's
+    img1, img2: [0, 255]
+    '''
+    #img1 = img1.squeeze()
+    #img2 = img2.squeeze()
+    if not img1.shape == img2.shape:
+        raise ValueError('Input images must have the same dimensions.')
+    h, w = img1.shape[:2]
+    img1 = img1[border:h-border, border:w-border]
+    img2 = img2[border:h-border, border:w-border]
+
+    if img1.ndim == 2:
+        return ssim(img1, img2)
+    elif img1.ndim == 3:
+        if img1.shape[2] == 3:
+            ssims = []
+            for i in range(3):
+                ssims.append(ssim(img1[:,:,i], img2[:,:,i]))
+            return np.array(ssims).mean()
+        elif img1.shape[2] == 1:
+            return ssim(np.squeeze(img1), np.squeeze(img2))
+    else:
+        raise ValueError('Wrong input image dimensions.')
+
+
+def ssim(img1, img2):
+    C1 = (0.01 * 255)**2
+    C2 = (0.03 * 255)**2
+
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    kernel = cv2.getGaussianKernel(11, 1.5)
+    window = np.outer(kernel, kernel.transpose())
+
+    mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5]  # valid
+    mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]
+    mu1_sq = mu1**2
+    mu2_sq = mu2**2
+    mu1_mu2 = mu1 * mu2
+    sigma1_sq = cv2.filter2D(img1**2, -1, window)[5:-5, 5:-5] - mu1_sq
+    sigma2_sq = cv2.filter2D(img2**2, -1, window)[5:-5, 5:-5] - mu2_sq
+    sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2
+
+    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) *
+                                                            (sigma1_sq + sigma2_sq + C2))
+    return ssim_map.mean()
+
+
+'''
+# --------------------------------------------
+# matlab's bicubic imresize (numpy and torch) [0, 1]
+# --------------------------------------------
+'''
+
+
+# matlab 'imresize' function, now only support 'bicubic'
+def cubic(x):
+    absx = torch.abs(x)
+    absx2 = absx**2
+    absx3 = absx**3
+    return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \
+        (-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx))
+
+
+def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing):
+    if (scale < 1) and (antialiasing):
+        # Use a modified kernel to simultaneously interpolate and antialias- larger kernel width
+        kernel_width = kernel_width / scale
+
+    # Output-space coordinates
+    x = torch.linspace(1, out_length, out_length)
+
+    # Input-space coordinates. Calculate the inverse mapping such that 0.5
+    # in output space maps to 0.5 in input space, and 0.5+scale in output
+    # space maps to 1.5 in input space.
+    u = x / scale + 0.5 * (1 - 1 / scale)
+
+    # What is the left-most pixel that can be involved in the computation?
+    left = torch.floor(u - kernel_width / 2)
+
+    # What is the maximum number of pixels that can be involved in the
+    # computation?  Note: it's OK to use an extra pixel here; if the
+    # corresponding weights are all zero, it will be eliminated at the end
+    # of this function.
+    P = math.ceil(kernel_width) + 2
+
+    # The indices of the input pixels involved in computing the k-th output
+    # pixel are in row k of the indices matrix.
+    indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view(
+        1, P).expand(out_length, P)
+
+    # The weights used to compute the k-th output pixel are in row k of the
+    # weights matrix.
+    distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices
+    # apply cubic kernel
+    if (scale < 1) and (antialiasing):
+        weights = scale * cubic(distance_to_center * scale)
+    else:
+        weights = cubic(distance_to_center)
+    # Normalize the weights matrix so that each row sums to 1.
+    weights_sum = torch.sum(weights, 1).view(out_length, 1)
+    weights = weights / weights_sum.expand(out_length, P)
+
+    # If a column in weights is all zero, get rid of it. only consider the first and last column.
+    weights_zero_tmp = torch.sum((weights == 0), 0)
+    if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6):
+        indices = indices.narrow(1, 1, P - 2)
+        weights = weights.narrow(1, 1, P - 2)
+    if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6):
+        indices = indices.narrow(1, 0, P - 2)
+        weights = weights.narrow(1, 0, P - 2)
+    weights = weights.contiguous()
+    indices = indices.contiguous()
+    sym_len_s = -indices.min() + 1
+    sym_len_e = indices.max() - in_length
+    indices = indices + sym_len_s - 1
+    return weights, indices, int(sym_len_s), int(sym_len_e)
+
+
+# --------------------------------------------
+# imresize for tensor image [0, 1]
+# --------------------------------------------
+def imresize(img, scale, antialiasing=True):
+    # Now the scale should be the same for H and W
+    # input: img: pytorch tensor, CHW or HW [0,1]
+    # output: CHW or HW [0,1] w/o round
+    need_squeeze = True if img.dim() == 2 else False
+    if need_squeeze:
+        img.unsqueeze_(0)
+    in_C, in_H, in_W = img.size()
+    out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale)
+    kernel_width = 4
+    kernel = 'cubic'
+
+    # Return the desired dimension order for performing the resize.  The
+    # strategy is to perform the resize first along the dimension with the
+    # smallest scale factor.
+    # Now we do not support this.
+
+    # get weights and indices
+    weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices(
+        in_H, out_H, scale, kernel, kernel_width, antialiasing)
+    weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices(
+        in_W, out_W, scale, kernel, kernel_width, antialiasing)
+    # process H dimension
+    # symmetric copying
+    img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W)
+    img_aug.narrow(1, sym_len_Hs, in_H).copy_(img)
+
+    sym_patch = img[:, :sym_len_Hs, :]
+    inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(1, inv_idx)
+    img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv)
+
+    sym_patch = img[:, -sym_len_He:, :]
+    inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(1, inv_idx)
+    img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv)
+
+    out_1 = torch.FloatTensor(in_C, out_H, in_W)
+    kernel_width = weights_H.size(1)
+    for i in range(out_H):
+        idx = int(indices_H[i][0])
+        for j in range(out_C):
+            out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i])
+
+    # process W dimension
+    # symmetric copying
+    out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We)
+    out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1)
+
+    sym_patch = out_1[:, :, :sym_len_Ws]
+    inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(2, inv_idx)
+    out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv)
+
+    sym_patch = out_1[:, :, -sym_len_We:]
+    inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(2, inv_idx)
+    out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv)
+
+    out_2 = torch.FloatTensor(in_C, out_H, out_W)
+    kernel_width = weights_W.size(1)
+    for i in range(out_W):
+        idx = int(indices_W[i][0])
+        for j in range(out_C):
+            out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i])
+    if need_squeeze:
+        out_2.squeeze_()
+    return out_2
+
+
+# --------------------------------------------
+# imresize for numpy image [0, 1]
+# --------------------------------------------
+def imresize_np(img, scale, antialiasing=True):
+    # Now the scale should be the same for H and W
+    # input: img: Numpy, HWC or HW [0,1]
+    # output: HWC or HW [0,1] w/o round
+    img = torch.from_numpy(img)
+    need_squeeze = True if img.dim() == 2 else False
+    if need_squeeze:
+        img.unsqueeze_(2)
+
+    in_H, in_W, in_C = img.size()
+    out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale)
+    kernel_width = 4
+    kernel = 'cubic'
+
+    # Return the desired dimension order for performing the resize.  The
+    # strategy is to perform the resize first along the dimension with the
+    # smallest scale factor.
+    # Now we do not support this.
+
+    # get weights and indices
+    weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices(
+        in_H, out_H, scale, kernel, kernel_width, antialiasing)
+    weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices(
+        in_W, out_W, scale, kernel, kernel_width, antialiasing)
+    # process H dimension
+    # symmetric copying
+    img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C)
+    img_aug.narrow(0, sym_len_Hs, in_H).copy_(img)
+
+    sym_patch = img[:sym_len_Hs, :, :]
+    inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(0, inv_idx)
+    img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv)
+
+    sym_patch = img[-sym_len_He:, :, :]
+    inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(0, inv_idx)
+    img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv)
+
+    out_1 = torch.FloatTensor(out_H, in_W, in_C)
+    kernel_width = weights_H.size(1)
+    for i in range(out_H):
+        idx = int(indices_H[i][0])
+        for j in range(out_C):
+            out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i])
+
+    # process W dimension
+    # symmetric copying
+    out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C)
+    out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1)
+
+    sym_patch = out_1[:, :sym_len_Ws, :]
+    inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(1, inv_idx)
+    out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv)
+
+    sym_patch = out_1[:, -sym_len_We:, :]
+    inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long()
+    sym_patch_inv = sym_patch.index_select(1, inv_idx)
+    out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv)
+
+    out_2 = torch.FloatTensor(out_H, out_W, in_C)
+    kernel_width = weights_W.size(1)
+    for i in range(out_W):
+        idx = int(indices_W[i][0])
+        for j in range(out_C):
+            out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i])
+    if need_squeeze:
+        out_2.squeeze_()
+
+    return out_2.numpy()
+
+
+if __name__ == '__main__':
+    print('---')
+#    img = imread_uint('test.bmp', 3)
+#    img = uint2single(img)
+#    img_bicubic = imresize_np(img, 1/4)

ldm/modules/midas/api.py

@@ -0,0 +1,170 @@
+# based on https://github.com/isl-org/MiDaS
+
+import cv2
+import torch
+import torch.nn as nn
+from torchvision.transforms import Compose
+
+from ldm.modules.midas.midas.dpt_depth import DPTDepthModel
+from ldm.modules.midas.midas.midas_net import MidasNet
+from ldm.modules.midas.midas.midas_net_custom import MidasNet_small
+from ldm.modules.midas.midas.transforms import Resize, NormalizeImage, PrepareForNet
+
+
+ISL_PATHS = {
+    "dpt_large": "midas_models/dpt_large-midas-2f21e586.pt",
+    "dpt_hybrid": "midas_models/dpt_hybrid-midas-501f0c75.pt",
+    "midas_v21": "",
+    "midas_v21_small": "",
+}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+def load_midas_transform(model_type):
+    # https://github.com/isl-org/MiDaS/blob/master/run.py
+    # load transform only
+    if model_type == "dpt_large":  # DPT-Large
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "dpt_hybrid":  # DPT-Hybrid
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "midas_v21":
+        net_w, net_h = 384, 384
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+    elif model_type == "midas_v21_small":
+        net_w, net_h = 256, 256
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+    else:
+        assert False, f"model_type '{model_type}' not implemented, use: --model_type large"
+
+    transform = Compose(
+        [
+            Resize(
+                net_w,
+                net_h,
+                resize_target=None,
+                keep_aspect_ratio=True,
+                ensure_multiple_of=32,
+                resize_method=resize_mode,
+                image_interpolation_method=cv2.INTER_CUBIC,
+            ),
+            normalization,
+            PrepareForNet(),
+        ]
+    )
+
+    return transform
+
+
+def load_model(model_type):
+    # https://github.com/isl-org/MiDaS/blob/master/run.py
+    # load network
+    model_path = ISL_PATHS[model_type]
+    if model_type == "dpt_large":  # DPT-Large
+        model = DPTDepthModel(
+            path=model_path,
+            backbone="vitl16_384",
+            non_negative=True,
+        )
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "dpt_hybrid":  # DPT-Hybrid
+        model = DPTDepthModel(
+            path=model_path,
+            backbone="vitb_rn50_384",
+            non_negative=True,
+        )
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "midas_v21":
+        model = MidasNet(model_path, non_negative=True)
+        net_w, net_h = 384, 384
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(
+            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+        )
+
+    elif model_type == "midas_v21_small":
+        model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True,
+                               non_negative=True, blocks={'expand': True})
+        net_w, net_h = 256, 256
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(
+            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+        )
+
+    else:
+        print(f"model_type '{model_type}' not implemented, use: --model_type large")
+        assert False
+
+    transform = Compose(
+        [
+            Resize(
+                net_w,
+                net_h,
+                resize_target=None,
+                keep_aspect_ratio=True,
+                ensure_multiple_of=32,
+                resize_method=resize_mode,
+                image_interpolation_method=cv2.INTER_CUBIC,
+            ),
+            normalization,
+            PrepareForNet(),
+        ]
+    )
+
+    return model.eval(), transform
+
+
+class MiDaSInference(nn.Module):
+    MODEL_TYPES_TORCH_HUB = [
+        "DPT_Large",
+        "DPT_Hybrid",
+        "MiDaS_small"
+    ]
+    MODEL_TYPES_ISL = [
+        "dpt_large",
+        "dpt_hybrid",
+        "midas_v21",
+        "midas_v21_small",
+    ]
+
+    def __init__(self, model_type):
+        super().__init__()
+        assert (model_type in self.MODEL_TYPES_ISL)
+        model, _ = load_model(model_type)
+        self.model = model
+        self.model.train = disabled_train
+
+    def forward(self, x):
+        # x in 0..1 as produced by calling self.transform on a 0..1 float64 numpy array
+        # NOTE: we expect that the correct transform has been called during dataloading.
+        with torch.no_grad():
+            prediction = self.model(x)
+            prediction = torch.nn.functional.interpolate(
+                prediction.unsqueeze(1),
+                size=x.shape[2:],
+                mode="bicubic",
+                align_corners=False,
+            )
+        assert prediction.shape == (x.shape[0], 1, x.shape[2], x.shape[3])
+        return prediction
+

ldm/modules/midas/midas/base_model.py

@@ -0,0 +1,16 @@
+import torch
+
+
+class BaseModel(torch.nn.Module):
+    def load(self, path):
+        """Load model from file.
+
+        Args:
+            path (str): file path
+        """
+        parameters = torch.load(path, map_location=torch.device('cpu'))
+
+        if "optimizer" in parameters:
+            parameters = parameters["model"]
+
+        self.load_state_dict(parameters)

ldm/modules/midas/midas/blocks.py

@@ -0,0 +1,342 @@
+import torch
+import torch.nn as nn
+
+from .vit import (
+    _make_pretrained_vitb_rn50_384,
+    _make_pretrained_vitl16_384,
+    _make_pretrained_vitb16_384,
+    forward_vit,
+)
+
+def _make_encoder(backbone, features, use_pretrained, groups=1, expand=False, exportable=True, hooks=None, use_vit_only=False, use_readout="ignore",):
+    if backbone == "vitl16_384":
+        pretrained = _make_pretrained_vitl16_384(
+            use_pretrained, hooks=hooks, use_readout=use_readout
+        )
+        scratch = _make_scratch(
+            [256, 512, 1024, 1024], features, groups=groups, expand=expand
+        )  # ViT-L/16 - 85.0% Top1 (backbone)
+    elif backbone == "vitb_rn50_384":
+        pretrained = _make_pretrained_vitb_rn50_384(
+            use_pretrained,
+            hooks=hooks,
+            use_vit_only=use_vit_only,
+            use_readout=use_readout,
+        )
+        scratch = _make_scratch(
+            [256, 512, 768, 768], features, groups=groups, expand=expand
+        )  # ViT-H/16 - 85.0% Top1 (backbone)
+    elif backbone == "vitb16_384":
+        pretrained = _make_pretrained_vitb16_384(
+            use_pretrained, hooks=hooks, use_readout=use_readout
+        )
+        scratch = _make_scratch(
+            [96, 192, 384, 768], features, groups=groups, expand=expand
+        )  # ViT-B/16 - 84.6% Top1 (backbone)
+    elif backbone == "resnext101_wsl":
+        pretrained = _make_pretrained_resnext101_wsl(use_pretrained)
+        scratch = _make_scratch([256, 512, 1024, 2048], features, groups=groups, expand=expand)     # efficientnet_lite3  
+    elif backbone == "efficientnet_lite3":
+        pretrained = _make_pretrained_efficientnet_lite3(use_pretrained, exportable=exportable)
+        scratch = _make_scratch([32, 48, 136, 384], features, groups=groups, expand=expand)  # efficientnet_lite3     
+    else:
+        print(f"Backbone '{backbone}' not implemented")
+        assert False
+        
+    return pretrained, scratch
+
+
+def _make_scratch(in_shape, out_shape, groups=1, expand=False):
+    scratch = nn.Module()
+
+    out_shape1 = out_shape
+    out_shape2 = out_shape
+    out_shape3 = out_shape
+    out_shape4 = out_shape
+    if expand==True:
+        out_shape1 = out_shape
+        out_shape2 = out_shape*2
+        out_shape3 = out_shape*4
+        out_shape4 = out_shape*8
+
+    scratch.layer1_rn = nn.Conv2d(
+        in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer2_rn = nn.Conv2d(
+        in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer3_rn = nn.Conv2d(
+        in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer4_rn = nn.Conv2d(
+        in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+
+    return scratch
+
+
+def _make_pretrained_efficientnet_lite3(use_pretrained, exportable=False):
+    efficientnet = torch.hub.load(
+        "rwightman/gen-efficientnet-pytorch",
+        "tf_efficientnet_lite3",
+        pretrained=use_pretrained,
+        exportable=exportable
+    )
+    return _make_efficientnet_backbone(efficientnet)
+
+
+def _make_efficientnet_backbone(effnet):
+    pretrained = nn.Module()
+
+    pretrained.layer1 = nn.Sequential(
+        effnet.conv_stem, effnet.bn1, effnet.act1, *effnet.blocks[0:2]
+    )
+    pretrained.layer2 = nn.Sequential(*effnet.blocks[2:3])
+    pretrained.layer3 = nn.Sequential(*effnet.blocks[3:5])
+    pretrained.layer4 = nn.Sequential(*effnet.blocks[5:9])
+
+    return pretrained
+    
+
+def _make_resnet_backbone(resnet):
+    pretrained = nn.Module()
+    pretrained.layer1 = nn.Sequential(
+        resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool, resnet.layer1
+    )
+
+    pretrained.layer2 = resnet.layer2
+    pretrained.layer3 = resnet.layer3
+    pretrained.layer4 = resnet.layer4
+
+    return pretrained
+
+
+def _make_pretrained_resnext101_wsl(use_pretrained):
+    resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl")
+    return _make_resnet_backbone(resnet)
+
+
+
+class Interpolate(nn.Module):
+    """Interpolation module.
+    """
+
+    def __init__(self, scale_factor, mode, align_corners=False):
+        """Init.
+
+        Args:
+            scale_factor (float): scaling
+            mode (str): interpolation mode
+        """
+        super(Interpolate, self).__init__()
+
+        self.interp = nn.functional.interpolate
+        self.scale_factor = scale_factor
+        self.mode = mode
+        self.align_corners = align_corners
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: interpolated data
+        """
+
+        x = self.interp(
+            x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners
+        )
+
+        return x
+
+
+class ResidualConvUnit(nn.Module):
+    """Residual convolution module.
+    """
+
+    def __init__(self, features):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.conv1 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True
+        )
+
+        self.conv2 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True
+        )
+
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+        out = self.relu(x)
+        out = self.conv1(out)
+        out = self.relu(out)
+        out = self.conv2(out)
+
+        return out + x
+
+
+class FeatureFusionBlock(nn.Module):
+    """Feature fusion block.
+    """
+
+    def __init__(self, features):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock, self).__init__()
+
+        self.resConfUnit1 = ResidualConvUnit(features)
+        self.resConfUnit2 = ResidualConvUnit(features)
+
+    def forward(self, *xs):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            output += self.resConfUnit1(xs[1])
+
+        output = self.resConfUnit2(output)
+
+        output = nn.functional.interpolate(
+            output, scale_factor=2, mode="bilinear", align_corners=True
+        )
+
+        return output
+
+
+
+
+class ResidualConvUnit_custom(nn.Module):
+    """Residual convolution module.
+    """
+
+    def __init__(self, features, activation, bn):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.bn = bn
+
+        self.groups=1
+
+        self.conv1 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+        
+        self.conv2 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+
+        if self.bn==True:
+            self.bn1 = nn.BatchNorm2d(features)
+            self.bn2 = nn.BatchNorm2d(features)
+
+        self.activation = activation
+
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+        
+        out = self.activation(x)
+        out = self.conv1(out)
+        if self.bn==True:
+            out = self.bn1(out)
+       
+        out = self.activation(out)
+        out = self.conv2(out)
+        if self.bn==True:
+            out = self.bn2(out)
+
+        if self.groups > 1:
+            out = self.conv_merge(out)
+
+        return self.skip_add.add(out, x)
+
+        # return out + x
+
+
+class FeatureFusionBlock_custom(nn.Module):
+    """Feature fusion block.
+    """
+
+    def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock_custom, self).__init__()
+
+        self.deconv = deconv
+        self.align_corners = align_corners
+
+        self.groups=1
+
+        self.expand = expand
+        out_features = features
+        if self.expand==True:
+            out_features = features//2
+        
+        self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1)
+
+        self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn)
+        self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn)
+        
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, *xs):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            res = self.resConfUnit1(xs[1])
+            output = self.skip_add.add(output, res)
+            # output += res
+
+        output = self.resConfUnit2(output)
+
+        output = nn.functional.interpolate(
+            output, scale_factor=2, mode="bilinear", align_corners=self.align_corners
+        )
+
+        output = self.out_conv(output)
+
+        return output
+

ldm/modules/midas/midas/dpt_depth.py

@@ -0,0 +1,109 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .base_model import BaseModel
+from .blocks import (
+    FeatureFusionBlock,
+    FeatureFusionBlock_custom,
+    Interpolate,
+    _make_encoder,
+    forward_vit,
+)
+
+
+def _make_fusion_block(features, use_bn):
+    return FeatureFusionBlock_custom(
+        features,
+        nn.ReLU(False),
+        deconv=False,
+        bn=use_bn,
+        expand=False,
+        align_corners=True,
+    )
+
+
+class DPT(BaseModel):
+    def __init__(
+        self,
+        head,
+        features=256,
+        backbone="vitb_rn50_384",
+        readout="project",
+        channels_last=False,
+        use_bn=False,
+    ):
+
+        super(DPT, self).__init__()
+
+        self.channels_last = channels_last
+
+        hooks = {
+            "vitb_rn50_384": [0, 1, 8, 11],
+            "vitb16_384": [2, 5, 8, 11],
+            "vitl16_384": [5, 11, 17, 23],
+        }
+
+        # Instantiate backbone and reassemble blocks
+        self.pretrained, self.scratch = _make_encoder(
+            backbone,
+            features,
+            False, # Set to true of you want to train from scratch, uses ImageNet weights
+            groups=1,
+            expand=False,
+            exportable=False,
+            hooks=hooks[backbone],
+            use_readout=readout,
+        )
+
+        self.scratch.refinenet1 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet2 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet3 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet4 = _make_fusion_block(features, use_bn)
+
+        self.scratch.output_conv = head
+
+
+    def forward(self, x):
+        if self.channels_last == True:
+            x.contiguous(memory_format=torch.channels_last)
+
+        layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x)
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+
+        out = self.scratch.output_conv(path_1)
+
+        return out
+
+
+class DPTDepthModel(DPT):
+    def __init__(self, path=None, non_negative=True, **kwargs):
+        features = kwargs["features"] if "features" in kwargs else 256
+
+        head = nn.Sequential(
+            nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1),
+            Interpolate(scale_factor=2, mode="bilinear", align_corners=True),
+            nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1),
+            nn.ReLU(True),
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+            nn.Identity(),
+        )
+
+        super().__init__(head, **kwargs)
+
+        if path is not None:
+           self.load(path)
+
+    def forward(self, x):
+        return super().forward(x).squeeze(dim=1)
+

ldm/modules/midas/midas/midas_net.py

@@ -0,0 +1,76 @@
+"""MidashNet: Network for monocular depth estimation trained by mixing several datasets.
+This file contains code that is adapted from
+https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
+"""
+import torch
+import torch.nn as nn
+
+from .base_model import BaseModel
+from .blocks import FeatureFusionBlock, Interpolate, _make_encoder
+
+
+class MidasNet(BaseModel):
+    """Network for monocular depth estimation.
+    """
+
+    def __init__(self, path=None, features=256, non_negative=True):
+        """Init.
+
+        Args:
+            path (str, optional): Path to saved model. Defaults to None.
+            features (int, optional): Number of features. Defaults to 256.
+            backbone (str, optional): Backbone network for encoder. Defaults to resnet50
+        """
+        print("Loading weights: ", path)
+
+        super(MidasNet, self).__init__()
+
+        use_pretrained = False if path is None else True
+
+        self.pretrained, self.scratch = _make_encoder(backbone="resnext101_wsl", features=features, use_pretrained=use_pretrained)
+
+        self.scratch.refinenet4 = FeatureFusionBlock(features)
+        self.scratch.refinenet3 = FeatureFusionBlock(features)
+        self.scratch.refinenet2 = FeatureFusionBlock(features)
+        self.scratch.refinenet1 = FeatureFusionBlock(features)
+
+        self.scratch.output_conv = nn.Sequential(
+            nn.Conv2d(features, 128, kernel_size=3, stride=1, padding=1),
+            Interpolate(scale_factor=2, mode="bilinear"),
+            nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1),
+            nn.ReLU(True),
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+        )
+
+        if path:
+            self.load(path)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input data (image)
+
+        Returns:
+            tensor: depth
+        """
+
+        layer_1 = self.pretrained.layer1(x)
+        layer_2 = self.pretrained.layer2(layer_1)
+        layer_3 = self.pretrained.layer3(layer_2)
+        layer_4 = self.pretrained.layer4(layer_3)
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+
+        out = self.scratch.output_conv(path_1)
+
+        return torch.squeeze(out, dim=1)

ldm/modules/midas/midas/midas_net_custom.py

@@ -0,0 +1,128 @@
+"""MidashNet: Network for monocular depth estimation trained by mixing several datasets.
+This file contains code that is adapted from
+https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
+"""
+import torch
+import torch.nn as nn
+
+from .base_model import BaseModel
+from .blocks import FeatureFusionBlock, FeatureFusionBlock_custom, Interpolate, _make_encoder
+
+
+class MidasNet_small(BaseModel):
+    """Network for monocular depth estimation.
+    """
+
+    def __init__(self, path=None, features=64, backbone="efficientnet_lite3", non_negative=True, exportable=True, channels_last=False, align_corners=True,
+        blocks={'expand': True}):
+        """Init.
+
+        Args:
+            path (str, optional): Path to saved model. Defaults to None.
+            features (int, optional): Number of features. Defaults to 256.
+            backbone (str, optional): Backbone network for encoder. Defaults to resnet50
+        """
+        print("Loading weights: ", path)
+
+        super(MidasNet_small, self).__init__()
+
+        use_pretrained = False if path else True
+                
+        self.channels_last = channels_last
+        self.blocks = blocks
+        self.backbone = backbone
+
+        self.groups = 1
+
+        features1=features
+        features2=features
+        features3=features
+        features4=features
+        self.expand = False
+        if "expand" in self.blocks and self.blocks['expand'] == True:
+            self.expand = True
+            features1=features
+            features2=features*2
+            features3=features*4
+            features4=features*8
+
+        self.pretrained, self.scratch = _make_encoder(self.backbone, features, use_pretrained, groups=self.groups, expand=self.expand, exportable=exportable)
+  
+        self.scratch.activation = nn.ReLU(False)    
+
+        self.scratch.refinenet4 = FeatureFusionBlock_custom(features4, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet3 = FeatureFusionBlock_custom(features3, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet2 = FeatureFusionBlock_custom(features2, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet1 = FeatureFusionBlock_custom(features1, self.scratch.activation, deconv=False, bn=False, align_corners=align_corners)
+
+        
+        self.scratch.output_conv = nn.Sequential(
+            nn.Conv2d(features, features//2, kernel_size=3, stride=1, padding=1, groups=self.groups),
+            Interpolate(scale_factor=2, mode="bilinear"),
+            nn.Conv2d(features//2, 32, kernel_size=3, stride=1, padding=1),
+            self.scratch.activation,
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+            nn.Identity(),
+        )
+        
+        if path:
+            self.load(path)
+
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input data (image)
+
+        Returns:
+            tensor: depth
+        """
+        if self.channels_last==True:
+            print("self.channels_last = ", self.channels_last)
+            x.contiguous(memory_format=torch.channels_last)
+
+
+        layer_1 = self.pretrained.layer1(x)
+        layer_2 = self.pretrained.layer2(layer_1)
+        layer_3 = self.pretrained.layer3(layer_2)
+        layer_4 = self.pretrained.layer4(layer_3)
+        
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+        
+        out = self.scratch.output_conv(path_1)
+
+        return torch.squeeze(out, dim=1)
+
+
+
+def fuse_model(m):
+    prev_previous_type = nn.Identity()
+    prev_previous_name = ''
+    previous_type = nn.Identity()
+    previous_name = ''
+    for name, module in m.named_modules():
+        if prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d and type(module) == nn.ReLU:
+            # print("FUSED ", prev_previous_name, previous_name, name)
+            torch.quantization.fuse_modules(m, [prev_previous_name, previous_name, name], inplace=True)
+        elif prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d:
+            # print("FUSED ", prev_previous_name, previous_name)
+            torch.quantization.fuse_modules(m, [prev_previous_name, previous_name], inplace=True)
+        # elif previous_type == nn.Conv2d and type(module) == nn.ReLU:
+        #    print("FUSED ", previous_name, name)
+        #    torch.quantization.fuse_modules(m, [previous_name, name], inplace=True)
+
+        prev_previous_type = previous_type
+        prev_previous_name = previous_name
+        previous_type = type(module)
+        previous_name = name

ldm/modules/midas/midas/transforms.py

@@ -0,0 +1,234 @@
+import numpy as np
+import cv2
+import math
+
+
+def apply_min_size(sample, size, image_interpolation_method=cv2.INTER_AREA):
+    """Rezise the sample to ensure the given size. Keeps aspect ratio.
+
+    Args:
+        sample (dict): sample
+        size (tuple): image size
+
+    Returns:
+        tuple: new size
+    """
+    shape = list(sample["disparity"].shape)
+
+    if shape[0] >= size[0] and shape[1] >= size[1]:
+        return sample
+
+    scale = [0, 0]
+    scale[0] = size[0] / shape[0]
+    scale[1] = size[1] / shape[1]
+
+    scale = max(scale)
+
+    shape[0] = math.ceil(scale * shape[0])
+    shape[1] = math.ceil(scale * shape[1])
+
+    # resize
+    sample["image"] = cv2.resize(
+        sample["image"], tuple(shape[::-1]), interpolation=image_interpolation_method
+    )
+
+    sample["disparity"] = cv2.resize(
+        sample["disparity"], tuple(shape[::-1]), interpolation=cv2.INTER_NEAREST
+    )
+    sample["mask"] = cv2.resize(
+        sample["mask"].astype(np.float32),
+        tuple(shape[::-1]),
+        interpolation=cv2.INTER_NEAREST,
+    )
+    sample["mask"] = sample["mask"].astype(bool)
+
+    return tuple(shape)
+
+
+class Resize(object):
+    """Resize sample to given size (width, height).
+    """
+
+    def __init__(
+        self,
+        width,
+        height,
+        resize_target=True,
+        keep_aspect_ratio=False,
+        ensure_multiple_of=1,
+        resize_method="lower_bound",
+        image_interpolation_method=cv2.INTER_AREA,
+    ):
+        """Init.
+
+        Args:
+            width (int): desired output width
+            height (int): desired output height
+            resize_target (bool, optional):
+                True: Resize the full sample (image, mask, target).
+                False: Resize image only.
+                Defaults to True.
+            keep_aspect_ratio (bool, optional):
+                True: Keep the aspect ratio of the input sample.
+                Output sample might not have the given width and height, and
+                resize behaviour depends on the parameter 'resize_method'.
+                Defaults to False.
+            ensure_multiple_of (int, optional):
+                Output width and height is constrained to be multiple of this parameter.
+                Defaults to 1.
+            resize_method (str, optional):
+                "lower_bound": Output will be at least as large as the given size.
+                "upper_bound": Output will be at max as large as the given size. (Output size might be smaller than given size.)
+                "minimal": Scale as least as possible.  (Output size might be smaller than given size.)
+                Defaults to "lower_bound".
+        """
+        self.__width = width
+        self.__height = height
+
+        self.__resize_target = resize_target
+        self.__keep_aspect_ratio = keep_aspect_ratio
+        self.__multiple_of = ensure_multiple_of
+        self.__resize_method = resize_method
+        self.__image_interpolation_method = image_interpolation_method
+
+    def constrain_to_multiple_of(self, x, min_val=0, max_val=None):
+        y = (np.round(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if max_val is not None and y > max_val:
+            y = (np.floor(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if y < min_val:
+            y = (np.ceil(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        return y
+
+    def get_size(self, width, height):
+        # determine new height and width
+        scale_height = self.__height / height
+        scale_width = self.__width / width
+
+        if self.__keep_aspect_ratio:
+            if self.__resize_method == "lower_bound":
+                # scale such that output size is lower bound
+                if scale_width > scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "upper_bound":
+                # scale such that output size is upper bound
+                if scale_width < scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "minimal":
+                # scale as least as possbile
+                if abs(1 - scale_width) < abs(1 - scale_height):
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            else:
+                raise ValueError(
+                    f"resize_method {self.__resize_method} not implemented"
+                )
+
+        if self.__resize_method == "lower_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, min_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, min_val=self.__width
+            )
+        elif self.__resize_method == "upper_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, max_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, max_val=self.__width
+            )
+        elif self.__resize_method == "minimal":
+            new_height = self.constrain_to_multiple_of(scale_height * height)
+            new_width = self.constrain_to_multiple_of(scale_width * width)
+        else:
+            raise ValueError(f"resize_method {self.__resize_method} not implemented")
+
+        return (new_width, new_height)
+
+    def __call__(self, sample):
+        width, height = self.get_size(
+            sample["image"].shape[1], sample["image"].shape[0]
+        )
+
+        # resize sample
+        sample["image"] = cv2.resize(
+            sample["image"],
+            (width, height),
+            interpolation=self.__image_interpolation_method,
+        )
+
+        if self.__resize_target:
+            if "disparity" in sample:
+                sample["disparity"] = cv2.resize(
+                    sample["disparity"],
+                    (width, height),
+                    interpolation=cv2.INTER_NEAREST,
+                )
+
+            if "depth" in sample:
+                sample["depth"] = cv2.resize(
+                    sample["depth"], (width, height), interpolation=cv2.INTER_NEAREST
+                )
+
+            sample["mask"] = cv2.resize(
+                sample["mask"].astype(np.float32),
+                (width, height),
+                interpolation=cv2.INTER_NEAREST,
+            )
+            sample["mask"] = sample["mask"].astype(bool)
+
+        return sample
+
+
+class NormalizeImage(object):
+    """Normlize image by given mean and std.
+    """
+
+    def __init__(self, mean, std):
+        self.__mean = mean
+        self.__std = std
+
+    def __call__(self, sample):
+        sample["image"] = (sample["image"] - self.__mean) / self.__std
+
+        return sample
+
+
+class PrepareForNet(object):
+    """Prepare sample for usage as network input.
+    """
+
+    def __init__(self):
+        pass
+
+    def __call__(self, sample):
+        image = np.transpose(sample["image"], (2, 0, 1))
+        sample["image"] = np.ascontiguousarray(image).astype(np.float32)
+
+        if "mask" in sample:
+            sample["mask"] = sample["mask"].astype(np.float32)
+            sample["mask"] = np.ascontiguousarray(sample["mask"])
+
+        if "disparity" in sample:
+            disparity = sample["disparity"].astype(np.float32)
+            sample["disparity"] = np.ascontiguousarray(disparity)
+
+        if "depth" in sample:
+            depth = sample["depth"].astype(np.float32)
+            sample["depth"] = np.ascontiguousarray(depth)
+
+        return sample

ldm/modules/midas/midas/vit.py

@@ -0,0 +1,491 @@
+import torch
+import torch.nn as nn
+import timm
+import types
+import math
+import torch.nn.functional as F
+
+
+class Slice(nn.Module):
+    def __init__(self, start_index=1):
+        super(Slice, self).__init__()
+        self.start_index = start_index
+
+    def forward(self, x):
+        return x[:, self.start_index :]
+
+
+class AddReadout(nn.Module):
+    def __init__(self, start_index=1):
+        super(AddReadout, self).__init__()
+        self.start_index = start_index
+
+    def forward(self, x):
+        if self.start_index == 2:
+            readout = (x[:, 0] + x[:, 1]) / 2
+        else:
+            readout = x[:, 0]
+        return x[:, self.start_index :] + readout.unsqueeze(1)
+
+
+class ProjectReadout(nn.Module):
+    def __init__(self, in_features, start_index=1):
+        super(ProjectReadout, self).__init__()
+        self.start_index = start_index
+
+        self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU())
+
+    def forward(self, x):
+        readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :])
+        features = torch.cat((x[:, self.start_index :], readout), -1)
+
+        return self.project(features)
+
+
+class Transpose(nn.Module):
+    def __init__(self, dim0, dim1):
+        super(Transpose, self).__init__()
+        self.dim0 = dim0
+        self.dim1 = dim1
+
+    def forward(self, x):
+        x = x.transpose(self.dim0, self.dim1)
+        return x
+
+
+def forward_vit(pretrained, x):
+    b, c, h, w = x.shape
+
+    glob = pretrained.model.forward_flex(x)
+
+    layer_1 = pretrained.activations["1"]
+    layer_2 = pretrained.activations["2"]
+    layer_3 = pretrained.activations["3"]
+    layer_4 = pretrained.activations["4"]
+
+    layer_1 = pretrained.act_postprocess1[0:2](layer_1)
+    layer_2 = pretrained.act_postprocess2[0:2](layer_2)
+    layer_3 = pretrained.act_postprocess3[0:2](layer_3)
+    layer_4 = pretrained.act_postprocess4[0:2](layer_4)
+
+    unflatten = nn.Sequential(
+        nn.Unflatten(
+            2,
+            torch.Size(
+                [
+                    h // pretrained.model.patch_size[1],
+                    w // pretrained.model.patch_size[0],
+                ]
+            ),
+        )
+    )
+
+    if layer_1.ndim == 3:
+        layer_1 = unflatten(layer_1)
+    if layer_2.ndim == 3:
+        layer_2 = unflatten(layer_2)
+    if layer_3.ndim == 3:
+        layer_3 = unflatten(layer_3)
+    if layer_4.ndim == 3:
+        layer_4 = unflatten(layer_4)
+
+    layer_1 = pretrained.act_postprocess1[3 : len(pretrained.act_postprocess1)](layer_1)
+    layer_2 = pretrained.act_postprocess2[3 : len(pretrained.act_postprocess2)](layer_2)
+    layer_3 = pretrained.act_postprocess3[3 : len(pretrained.act_postprocess3)](layer_3)
+    layer_4 = pretrained.act_postprocess4[3 : len(pretrained.act_postprocess4)](layer_4)
+
+    return layer_1, layer_2, layer_3, layer_4
+
+
+def _resize_pos_embed(self, posemb, gs_h, gs_w):
+    posemb_tok, posemb_grid = (
+        posemb[:, : self.start_index],
+        posemb[0, self.start_index :],
+    )
+
+    gs_old = int(math.sqrt(len(posemb_grid)))
+
+    posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
+    posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear")
+    posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1)
+
+    posemb = torch.cat([posemb_tok, posemb_grid], dim=1)
+
+    return posemb
+
+
+def forward_flex(self, x):
+    b, c, h, w = x.shape
+
+    pos_embed = self._resize_pos_embed(
+        self.pos_embed, h // self.patch_size[1], w // self.patch_size[0]
+    )
+
+    B = x.shape[0]
+
+    if hasattr(self.patch_embed, "backbone"):
+        x = self.patch_embed.backbone(x)
+        if isinstance(x, (list, tuple)):
+            x = x[-1]  # last feature if backbone outputs list/tuple of features
+
+    x = self.patch_embed.proj(x).flatten(2).transpose(1, 2)
+
+    if getattr(self, "dist_token", None) is not None:
+        cls_tokens = self.cls_token.expand(
+            B, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        dist_token = self.dist_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, dist_token, x), dim=1)
+    else:
+        cls_tokens = self.cls_token.expand(
+            B, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+
+    x = x + pos_embed
+    x = self.pos_drop(x)
+
+    for blk in self.blocks:
+        x = blk(x)
+
+    x = self.norm(x)
+
+    return x
+
+
+activations = {}
+
+
+def get_activation(name):
+    def hook(model, input, output):
+        activations[name] = output
+
+    return hook
+
+
+def get_readout_oper(vit_features, features, use_readout, start_index=1):
+    if use_readout == "ignore":
+        readout_oper = [Slice(start_index)] * len(features)
+    elif use_readout == "add":
+        readout_oper = [AddReadout(start_index)] * len(features)
+    elif use_readout == "project":
+        readout_oper = [
+            ProjectReadout(vit_features, start_index) for out_feat in features
+        ]
+    else:
+        assert (
+            False
+        ), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'"
+
+    return readout_oper
+
+
+def _make_vit_b16_backbone(
+    model,
+    features=[96, 192, 384, 768],
+    size=[384, 384],
+    hooks=[2, 5, 8, 11],
+    vit_features=768,
+    use_readout="ignore",
+    start_index=1,
+):
+    pretrained = nn.Module()
+
+    pretrained.model = model
+    pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
+    pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
+    pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
+    pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
+
+    pretrained.activations = activations
+
+    readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
+
+    # 32, 48, 136, 384
+    pretrained.act_postprocess1 = nn.Sequential(
+        readout_oper[0],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[0],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.ConvTranspose2d(
+            in_channels=features[0],
+            out_channels=features[0],
+            kernel_size=4,
+            stride=4,
+            padding=0,
+            bias=True,
+            dilation=1,
+            groups=1,
+        ),
+    )
+
+    pretrained.act_postprocess2 = nn.Sequential(
+        readout_oper[1],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[1],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.ConvTranspose2d(
+            in_channels=features[1],
+            out_channels=features[1],
+            kernel_size=2,
+            stride=2,
+            padding=0,
+            bias=True,
+            dilation=1,
+            groups=1,
+        ),
+    )
+
+    pretrained.act_postprocess3 = nn.Sequential(
+        readout_oper[2],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[2],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+    )
+
+    pretrained.act_postprocess4 = nn.Sequential(
+        readout_oper[3],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[3],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.Conv2d(
+            in_channels=features[3],
+            out_channels=features[3],
+            kernel_size=3,
+            stride=2,
+            padding=1,
+        ),
+    )
+
+    pretrained.model.start_index = start_index
+    pretrained.model.patch_size = [16, 16]
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
+    pretrained.model._resize_pos_embed = types.MethodType(
+        _resize_pos_embed, pretrained.model
+    )
+
+    return pretrained
+
+
+def _make_pretrained_vitl16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_large_patch16_384", pretrained=pretrained)
+
+    hooks = [5, 11, 17, 23] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model,
+        features=[256, 512, 1024, 1024],
+        hooks=hooks,
+        vit_features=1024,
+        use_readout=use_readout,
+    )
+
+
+def _make_pretrained_vitb16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_base_patch16_384", pretrained=pretrained)
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
+    )
+
+
+def _make_pretrained_deitb16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_deit_base_patch16_384", pretrained=pretrained)
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
+    )
+
+
+def _make_pretrained_deitb16_distil_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model(
+        "vit_deit_base_distilled_patch16_384", pretrained=pretrained
+    )
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model,
+        features=[96, 192, 384, 768],
+        hooks=hooks,
+        use_readout=use_readout,
+        start_index=2,
+    )
+
+
+def _make_vit_b_rn50_backbone(
+    model,
+    features=[256, 512, 768, 768],
+    size=[384, 384],
+    hooks=[0, 1, 8, 11],
+    vit_features=768,
+    use_vit_only=False,
+    use_readout="ignore",
+    start_index=1,
+):
+    pretrained = nn.Module()
+
+    pretrained.model = model
+
+    if use_vit_only == True:
+        pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
+        pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
+    else:
+        pretrained.model.patch_embed.backbone.stages[0].register_forward_hook(
+            get_activation("1")
+        )
+        pretrained.model.patch_embed.backbone.stages[1].register_forward_hook(
+            get_activation("2")
+        )
+
+    pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
+    pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
+
+    pretrained.activations = activations
+
+    readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
+
+    if use_vit_only == True:
+        pretrained.act_postprocess1 = nn.Sequential(
+            readout_oper[0],
+            Transpose(1, 2),
+            nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+            nn.Conv2d(
+                in_channels=vit_features,
+                out_channels=features[0],
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=features[0],
+                out_channels=features[0],
+                kernel_size=4,
+                stride=4,
+                padding=0,
+                bias=True,
+                dilation=1,
+                groups=1,
+            ),
+        )
+
+        pretrained.act_postprocess2 = nn.Sequential(
+            readout_oper[1],
+            Transpose(1, 2),
+            nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+            nn.Conv2d(
+                in_channels=vit_features,
+                out_channels=features[1],
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=features[1],
+                out_channels=features[1],
+                kernel_size=2,
+                stride=2,
+                padding=0,
+                bias=True,
+                dilation=1,
+                groups=1,
+            ),
+        )
+    else:
+        pretrained.act_postprocess1 = nn.Sequential(
+            nn.Identity(), nn.Identity(), nn.Identity()
+        )
+        pretrained.act_postprocess2 = nn.Sequential(
+            nn.Identity(), nn.Identity(), nn.Identity()
+        )
+
+    pretrained.act_postprocess3 = nn.Sequential(
+        readout_oper[2],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[2],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+    )
+
+    pretrained.act_postprocess4 = nn.Sequential(
+        readout_oper[3],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[3],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.Conv2d(
+            in_channels=features[3],
+            out_channels=features[3],
+            kernel_size=3,
+            stride=2,
+            padding=1,
+        ),
+    )
+
+    pretrained.model.start_index = start_index
+    pretrained.model.patch_size = [16, 16]
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model._resize_pos_embed = types.MethodType(
+        _resize_pos_embed, pretrained.model
+    )
+
+    return pretrained
+
+
+def _make_pretrained_vitb_rn50_384(
+    pretrained, use_readout="ignore", hooks=None, use_vit_only=False
+):
+    model = timm.create_model("vit_base_resnet50_384", pretrained=pretrained)
+
+    hooks = [0, 1, 8, 11] if hooks == None else hooks
+    return _make_vit_b_rn50_backbone(
+        model,
+        features=[256, 512, 768, 768],
+        size=[384, 384],
+        hooks=hooks,
+        use_vit_only=use_vit_only,
+        use_readout=use_readout,
+    )

ldm/modules/midas/utils.py

@@ -0,0 +1,189 @@
+"""Utils for monoDepth."""
+import sys
+import re
+import numpy as np
+import cv2
+import torch
+
+
+def read_pfm(path):
+    """Read pfm file.
+
+    Args:
+        path (str): path to file
+
+    Returns:
+        tuple: (data, scale)
+    """
+    with open(path, "rb") as file:
+
+        color = None
+        width = None
+        height = None
+        scale = None
+        endian = None
+
+        header = file.readline().rstrip()
+        if header.decode("ascii") == "PF":
+            color = True
+        elif header.decode("ascii") == "Pf":
+            color = False
+        else:
+            raise Exception("Not a PFM file: " + path)
+
+        dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii"))
+        if dim_match:
+            width, height = list(map(int, dim_match.groups()))
+        else:
+            raise Exception("Malformed PFM header.")
+
+        scale = float(file.readline().decode("ascii").rstrip())
+        if scale < 0:
+            # little-endian
+            endian = "<"
+            scale = -scale
+        else:
+            # big-endian
+            endian = ">"
+
+        data = np.fromfile(file, endian + "f")
+        shape = (height, width, 3) if color else (height, width)
+
+        data = np.reshape(data, shape)
+        data = np.flipud(data)
+
+        return data, scale
+
+
+def write_pfm(path, image, scale=1):
+    """Write pfm file.
+
+    Args:
+        path (str): pathto file
+        image (array): data
+        scale (int, optional): Scale. Defaults to 1.
+    """
+
+    with open(path, "wb") as file:
+        color = None
+
+        if image.dtype.name != "float32":
+            raise Exception("Image dtype must be float32.")
+
+        image = np.flipud(image)
+
+        if len(image.shape) == 3 and image.shape[2] == 3:  # color image
+            color = True
+        elif (
+            len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1
+        ):  # greyscale
+            color = False
+        else:
+            raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.")
+
+        file.write("PF\n" if color else "Pf\n".encode())
+        file.write("%d %d\n".encode() % (image.shape[1], image.shape[0]))
+
+        endian = image.dtype.byteorder
+
+        if endian == "<" or endian == "=" and sys.byteorder == "little":
+            scale = -scale
+
+        file.write("%f\n".encode() % scale)
+
+        image.tofile(file)
+
+
+def read_image(path):
+    """Read image and output RGB image (0-1).
+
+    Args:
+        path (str): path to file
+
+    Returns:
+        array: RGB image (0-1)
+    """
+    img = cv2.imread(path)
+
+    if img.ndim == 2:
+        img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0
+
+    return img
+
+
+def resize_image(img):
+    """Resize image and make it fit for network.
+
+    Args:
+        img (array): image
+
+    Returns:
+        tensor: data ready for network
+    """
+    height_orig = img.shape[0]
+    width_orig = img.shape[1]
+
+    if width_orig > height_orig:
+        scale = width_orig / 384
+    else:
+        scale = height_orig / 384
+
+    height = (np.ceil(height_orig / scale / 32) * 32).astype(int)
+    width = (np.ceil(width_orig / scale / 32) * 32).astype(int)
+
+    img_resized = cv2.resize(img, (width, height), interpolation=cv2.INTER_AREA)
+
+    img_resized = (
+        torch.from_numpy(np.transpose(img_resized, (2, 0, 1))).contiguous().float()
+    )
+    img_resized = img_resized.unsqueeze(0)
+
+    return img_resized
+
+
+def resize_depth(depth, width, height):
+    """Resize depth map and bring to CPU (numpy).
+
+    Args:
+        depth (tensor): depth
+        width (int): image width
+        height (int): image height
+
+    Returns:
+        array: processed depth
+    """
+    depth = torch.squeeze(depth[0, :, :, :]).to("cpu")
+
+    depth_resized = cv2.resize(
+        depth.numpy(), (width, height), interpolation=cv2.INTER_CUBIC
+    )
+
+    return depth_resized
+
+def write_depth(path, depth, bits=1):
+    """Write depth map to pfm and png file.
+
+    Args:
+        path (str): filepath without extension
+        depth (array): depth
+    """
+    write_pfm(path + ".pfm", depth.astype(np.float32))
+
+    depth_min = depth.min()
+    depth_max = depth.max()
+
+    max_val = (2**(8*bits))-1
+
+    if depth_max - depth_min > np.finfo("float").eps:
+        out = max_val * (depth - depth_min) / (depth_max - depth_min)
+    else:
+        out = np.zeros(depth.shape, dtype=depth.type)
+
+    if bits == 1:
+        cv2.imwrite(path + ".png", out.astype("uint8"))
+    elif bits == 2:
+        cv2.imwrite(path + ".png", out.astype("uint16"))
+
+    return

ldm/util.py

@@ -0,0 +1,197 @@
+import importlib
+
+import torch
+from torch import optim
+import numpy as np
+
+from inspect import isfunction
+from PIL import Image, ImageDraw, ImageFont
+
+
+def log_txt_as_img(wh, xc, size=10):
+    # wh a tuple of (width, height)
+    # xc a list of captions to plot
+    b = len(xc)
+    txts = list()
+    for bi in range(b):
+        txt = Image.new("RGB", wh, color="white")
+        draw = ImageDraw.Draw(txt)
+        font = ImageFont.truetype('data/DejaVuSans.ttf', size=size)
+        nc = int(40 * (wh[0] / 256))
+        lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc))
+
+        try:
+            draw.text((0, 0), lines, fill="black", font=font)
+        except UnicodeEncodeError:
+            print("Cant encode string for logging. Skipping.")
+
+        txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0
+        txts.append(txt)
+    txts = np.stack(txts)
+    txts = torch.tensor(txts)
+    return txts
+
+
+def ismap(x):
+    if not isinstance(x, torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] > 3)
+
+
+def isimage(x):
+    if not isinstance(x,torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
+
+
+def exists(x):
+    return x is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def mean_flat(tensor):
+    """
+    https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def count_params(model, verbose=False):
+    total_params = sum(p.numel() for p in model.parameters())
+    if verbose:
+        print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.")
+    return total_params
+
+
+def instantiate_from_config(config):
+    if not "target" in config:
+        if config == '__is_first_stage__':
+            return None
+        elif config == "__is_unconditional__":
+            return None
+        raise KeyError("Expected key `target` to instantiate.")
+    return get_obj_from_str(config["target"])(**config.get("params", dict()))
+
+
+def get_obj_from_str(string, reload=False):
+    module, cls = string.rsplit(".", 1)
+    if reload:
+        module_imp = importlib.import_module(module)
+        importlib.reload(module_imp)
+    return getattr(importlib.import_module(module, package=None), cls)
+
+
+class AdamWwithEMAandWings(optim.Optimizer):
+    # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298
+    def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8,  # TODO: check hyperparameters before using
+                 weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999,   # ema decay to match previous code
+                 ema_power=1., param_names=()):
+        """AdamW that saves EMA versions of the parameters."""
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0.0 <= weight_decay:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0.0 <= ema_decay <= 1.0:
+            raise ValueError("Invalid ema_decay value: {}".format(ema_decay))
+        defaults = dict(lr=lr, betas=betas, eps=eps,
+                        weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay,
+                        ema_power=ema_power, param_names=param_names)
+        super().__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super().__setstate__(state)
+        for group in self.param_groups:
+            group.setdefault('amsgrad', False)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+        Args:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            params_with_grad = []
+            grads = []
+            exp_avgs = []
+            exp_avg_sqs = []
+            ema_params_with_grad = []
+            state_sums = []
+            max_exp_avg_sqs = []
+            state_steps = []
+            amsgrad = group['amsgrad']
+            beta1, beta2 = group['betas']
+            ema_decay = group['ema_decay']
+            ema_power = group['ema_power']
+
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+                params_with_grad.append(p)
+                if p.grad.is_sparse:
+                    raise RuntimeError('AdamW does not support sparse gradients')
+                grads.append(p.grad)
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state['step'] = 0
+                    # Exponential moving average of gradient values
+                    state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    # Exponential moving average of squared gradient values
+                    state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    if amsgrad:
+                        # Maintains max of all exp. moving avg. of sq. grad. values
+                        state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    # Exponential moving average of parameter values
+                    state['param_exp_avg'] = p.detach().float().clone()
+
+                exp_avgs.append(state['exp_avg'])
+                exp_avg_sqs.append(state['exp_avg_sq'])
+                ema_params_with_grad.append(state['param_exp_avg'])
+
+                if amsgrad:
+                    max_exp_avg_sqs.append(state['max_exp_avg_sq'])
+
+                # update the steps for each param group update
+                state['step'] += 1
+                # record the step after step update
+                state_steps.append(state['step'])
+
+            optim._functional.adamw(params_with_grad,
+                    grads,
+                    exp_avgs,
+                    exp_avg_sqs,
+                    max_exp_avg_sqs,
+                    state_steps,
+                    amsgrad=amsgrad,
+                    beta1=beta1,
+                    beta2=beta2,
+                    lr=group['lr'],
+                    weight_decay=group['weight_decay'],
+                    eps=group['eps'],
+                    maximize=False)
+
+            cur_ema_decay = min(ema_decay, 1 - state['step'] ** -ema_power)
+            for param, ema_param in zip(params_with_grad, ema_params_with_grad):
+                ema_param.mul_(cur_ema_decay).add_(param.float(), alpha=1 - cur_ema_decay)
+
+        return loss

modelcard.md

@@ -0,0 +1,140 @@
+# Stable Diffusion v2 Model Card
+This model card focuses on the model associated with the Stable Diffusion model, available [here](https://github.com/CompVis/stable-diffusion).
+
+## Model Details
+- **Developed by:** Robin Rombach, Patrick Esser
+- **Model type:** Diffusion-based text-to-image generation model
+- **Language(s):** English
+- **License:** CreativeML Open RAIL++-M License
+- **Model Description:** This is a model that can be used to generate and modify images based on text prompts. It is a [Latent Diffusion Model](https://arxiv.org/abs/2112.10752) that uses a fixed, pretrained text encoder ([OpenCLIP-ViT/H](https://github.com/mlfoundations/open_clip)).
+- **Resources for more information:** [GitHub Repository](https://github.com/Stability-AI/).
+- **Cite as:**
+
+      @InProceedings{Rombach_2022_CVPR,
+          author    = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
+          title     = {High-Resolution Image Synthesis With Latent Diffusion Models},
+          booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
+          month     = {June},
+          year      = {2022},
+          pages     = {10684-10695}
+      }
+
+# Uses
+
+## Direct Use 
+The model is intended for research purposes only. Possible research areas and tasks include
+
+- Safe deployment of models which have the potential to generate harmful content.
+- Probing and understanding the limitations and biases of generative models.
+- Generation of artworks and use in design and other artistic processes.
+- Applications in educational or creative tools.
+- Research on generative models.
+
+Excluded uses are described below.
+
+ ### Misuse, Malicious Use, and Out-of-Scope Use
+_Note: This section is originally taken from the [DALLE-MINI model card](https://huggingface.co/dalle-mini/dalle-mini), was used for Stable Diffusion v1, but applies in the same way to Stable Diffusion v2_.
+
+The model should not be used to intentionally create or disseminate images that create hostile or alienating environments for people. This includes generating images that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.
+
+#### Out-of-Scope Use
+The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.
+
+#### Misuse and Malicious Use
+Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:
+
+- Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc.
+- Intentionally promoting or propagating discriminatory content or harmful stereotypes.
+- Impersonating individuals without their consent.
+- Sexual content without consent of the people who might see it.
+- Mis- and disinformation
+- Representations of egregious violence and gore
+- Sharing of copyrighted or licensed material in violation of its terms of use.
+- Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use.
+
+## Limitations and Bias
+
+### Limitations
+
+- The model does not achieve perfect photorealism
+- The model cannot render legible text
+- The model does not perform well on more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”
+- Faces and people in general may not be generated properly.
+- The model was trained mainly with English captions and will not work as well in other languages.
+- The autoencoding part of the model is lossy
+- The model was trained on a subset of the large-scale dataset
+  [LAION-5B](https://laion.ai/blog/laion-5b/), which contains adult, violent and sexual content. To partially mitigate this, we have filtered the dataset using LAION's NFSW detector (see Training section).
+
+### Bias
+While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases. 
+Stable Diffusion vw was primarily trained on subsets of [LAION-2B(en)](https://laion.ai/blog/laion-5b/), 
+which consists of images that are limited to English descriptions. 
+Texts and images from communities and cultures that use other languages are likely to be insufficiently accounted for. 
+This affects the overall output of the model, as white and western cultures are often set as the default. Further, the 
+ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts.
+Stable Diffusion v2 mirrors and exacerbates biases to such a degree that viewer discretion must be advised irrespective of the input or its intent.
+
+
+## Training
+
+**Training Data**
+The model developers used the following dataset for training the model:
+
+- LAION-5B and subsets (details below). The training data is further filtered using LAION's NSFW detector, with a "p_unsafe" score of 0.1 (conservative). For more details, please refer to LAION-5B's [NeurIPS 2022](https://openreview.net/forum?id=M3Y74vmsMcY) paper and reviewer discussions on the topic.
+
+**Training Procedure**
+Stable Diffusion v2 is a latent diffusion model which combines an autoencoder with a diffusion model that is trained in the latent space of the autoencoder. During training, 
+
+- Images are encoded through an encoder, which turns images into latent representations. The autoencoder uses a relative downsampling factor of 8 and maps images of shape H x W x 3 to latents of shape H/f x W/f x 4
+- Text prompts are encoded through the OpenCLIP-ViT/H text-encoder.
+- The output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross-attention.
+- The loss is a reconstruction objective between the noise that was added to the latent and the prediction made by the UNet. We also use the so-called _v-objective_, see https://arxiv.org/abs/2202.00512.
+
+We currently provide the following checkpoints:
+
+- `512-base-ema.ckpt`: 550k steps at resolution `256x256` on a subset of [LAION-5B](https://laion.ai/blog/laion-5b/) filtered for explicit pornographic material, using the [LAION-NSFW classifier](https://github.com/LAION-AI/CLIP-based-NSFW-Detector) with `punsafe=0.1` and an [aesthetic score](https://github.com/christophschuhmann/improved-aesthetic-predictor) >= `4.5`.
+  850k steps at resolution `512x512` on the same dataset with resolution `>= 512x512`.
+- `768-v-ema.ckpt`: Resumed from `512-base-ema.ckpt` and trained for 150k steps using a [v-objective](https://arxiv.org/abs/2202.00512) on the same dataset. Resumed for another 140k steps on a `768x768` subset of our dataset.
+- `512-depth-ema.ckpt`: Resumed from `512-base-ema.ckpt` and finetuned for 200k steps. Added an extra input channel to process the (relative) depth prediction produced by [MiDaS](https://github.com/isl-org/MiDaS) (`dpt_hybrid`) which is used as an additional conditioning.
+The additional input channels of the U-Net which process this extra information were zero-initialized.
+- `512-inpainting-ema.ckpt`: Resumed from `512-base-ema.ckpt` and trained for another 200k steps. Follows the mask-generation strategy presented in [LAMA](https://github.com/saic-mdal/lama) which, in combination with the latent VAE representations of the masked image, are used as an additional conditioning.
+The additional input channels of the U-Net which process this extra information were zero-initialized. The same strategy was used to train the [1.5-inpainting checkpoint](https://github.com/saic-mdal/lama).
+- `x4-upscaling-ema.ckpt`: Trained for 1.25M steps on a 10M subset of LAION containing images `>2048x2048`. The model was trained on crops of size `512x512` and is a text-guided [latent upscaling diffusion model](https://arxiv.org/abs/2112.10752).
+In addition to the textual input, it receives a `noise_level` as an input parameter, which can be used to add noise to the low-resolution input according to a [predefined diffusion schedule](configs/stable-diffusion/x4-upscaling.yaml). 
+
+- **Hardware:** 32 x 8 x A100 GPUs
+- **Optimizer:** AdamW
+- **Gradient Accumulations**: 1
+- **Batch:** 32 x 8 x 2 x 4 = 2048
+- **Learning rate:** warmup to 0.0001 for 10,000 steps and then kept constant
+
+## Evaluation Results 
+Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,
+5.0, 6.0, 7.0, 8.0) and 50 steps DDIM sampling steps show the relative improvements of the checkpoints:
+
+![pareto](assets/model-variants.jpg) 
+
+Evaluated using 50 DDIM steps and 10000 random prompts from the COCO2017 validation set, evaluated at 512x512 resolution.  Not optimized for FID scores.
+
+## Environmental Impact
+
+**Stable Diffusion v1** **Estimated Emissions**
+Based on that information, we estimate the following CO2 emissions using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). The hardware, runtime, cloud provider, and compute region were utilized to estimate the carbon impact.
+
+- **Hardware Type:** A100 PCIe 40GB
+- **Hours used:** 200000
+- **Cloud Provider:** AWS
+- **Compute Region:** US-east
+- **Carbon Emitted (Power consumption x Time x Carbon produced based on location of power grid):** 15000 kg CO2 eq.
+
+## Citation
+    @InProceedings{Rombach_2022_CVPR,
+        author    = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
+        title     = {High-Resolution Image Synthesis With Latent Diffusion Models},
+        booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
+        month     = {June},
+        year      = {2022},
+        pages     = {10684-10695}
+    }
+
+*This model card was written by: Robin Rombach, Patrick Esser and David Ha and is based on the [Stable Diffusion v1](https://github.com/CompVis/stable-diffusion/blob/main/Stable_Diffusion_v1_Model_Card.md) and [DALL-E Mini model card](https://huggingface.co/dalle-mini/dalle-mini).*

requirements.txt

@@ -0,0 +1,16 @@
+albumentations==0.4.3
+opencv-python
+pudb==2019.2
+imageio==2.9.0
+imageio-ffmpeg==0.4.2
+pytorch-lightning==1.4.2
+torchmetrics==0.6
+omegaconf==2.1.1
+test-tube>=0.7.5
+streamlit>=0.73.1
+einops==0.3.0
+transformers==4.19.2
+webdataset==0.2.5
+open-clip-torch==2.7.0
+gradio==3.11
+-e .

scripts/gradio/inpainting.py

@@ -0,0 +1,195 @@
+import sys
+import cv2
+import torch
+import numpy as np
+import gradio as gr
+from PIL import Image
+from omegaconf import OmegaConf
+from einops import repeat
+from imwatermark import WatermarkEncoder
+from pathlib import Path
+
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.util import instantiate_from_config
+
+
+torch.set_grad_enabled(False)
+
+
+def put_watermark(img, wm_encoder=None):
+    if wm_encoder is not None:
+        img = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
+        img = wm_encoder.encode(img, 'dwtDct')
+        img = Image.fromarray(img[:, :, ::-1])
+    return img
+
+
+def initialize_model(config, ckpt):
+    config = OmegaConf.load(config)
+    model = instantiate_from_config(config.model)
+
+    model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False)
+
+    device = torch.device(
+        "cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = model.to(device)
+    sampler = DDIMSampler(model)
+
+    return sampler
+
+
+def make_batch_sd(
+        image,
+        mask,
+        txt,
+        device,
+        num_samples=1):
+    image = np.array(image.convert("RGB"))
+    image = image[None].transpose(0, 3, 1, 2)
+    image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+    mask = np.array(mask.convert("L"))
+    mask = mask.astype(np.float32) / 255.0
+    mask = mask[None, None]
+    mask[mask < 0.5] = 0
+    mask[mask >= 0.5] = 1
+    mask = torch.from_numpy(mask)
+
+    masked_image = image * (mask < 0.5)
+
+    batch = {
+        "image": repeat(image.to(device=device), "1 ... -> n ...", n=num_samples),
+        "txt": num_samples * [txt],
+        "mask": repeat(mask.to(device=device), "1 ... -> n ...", n=num_samples),
+        "masked_image": repeat(masked_image.to(device=device), "1 ... -> n ...", n=num_samples),
+    }
+    return batch
+
+
+def inpaint(sampler, image, mask, prompt, seed, scale, ddim_steps, num_samples=1, w=512, h=512):
+    device = torch.device(
+        "cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = sampler.model
+
+    print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...")
+    wm = "SDV2"
+    wm_encoder = WatermarkEncoder()
+    wm_encoder.set_watermark('bytes', wm.encode('utf-8'))
+
+    prng = np.random.RandomState(seed)
+    start_code = prng.randn(num_samples, 4, h // 8, w // 8)
+    start_code = torch.from_numpy(start_code).to(
+        device=device, dtype=torch.float32)
+
+    with torch.no_grad(), \
+            torch.autocast("cuda"):
+        batch = make_batch_sd(image, mask, txt=prompt,
+                              device=device, num_samples=num_samples)
+
+        c = model.cond_stage_model.encode(batch["txt"])
+
+        c_cat = list()
+        for ck in model.concat_keys:
+            cc = batch[ck].float()
+            if ck != model.masked_image_key:
+                bchw = [num_samples, 4, h // 8, w // 8]
+                cc = torch.nn.functional.interpolate(cc, size=bchw[-2:])
+            else:
+                cc = model.get_first_stage_encoding(
+                    model.encode_first_stage(cc))
+            c_cat.append(cc)
+        c_cat = torch.cat(c_cat, dim=1)
+
+        # cond
+        cond = {"c_concat": [c_cat], "c_crossattn": [c]}
+
+        # uncond cond
+        uc_cross = model.get_unconditional_conditioning(num_samples, "")
+        uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]}
+
+        shape = [model.channels, h // 8, w // 8]
+        samples_cfg, intermediates = sampler.sample(
+            ddim_steps,
+            num_samples,
+            shape,
+            cond,
+            verbose=False,
+            eta=1.0,
+            unconditional_guidance_scale=scale,
+            unconditional_conditioning=uc_full,
+            x_T=start_code,
+        )
+        x_samples_ddim = model.decode_first_stage(samples_cfg)
+
+        result = torch.clamp((x_samples_ddim + 1.0) / 2.0,
+                             min=0.0, max=1.0)
+
+        result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255
+    return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result]
+
+def pad_image(input_image):
+    pad_w, pad_h = np.max(((2, 2), np.ceil(
+        np.array(input_image.size) / 64).astype(int)), axis=0) * 64 - input_image.size
+    im_padded = Image.fromarray(
+        np.pad(np.array(input_image), ((0, pad_h), (0, pad_w), (0, 0)), mode='edge'))
+    return im_padded
+
+def predict(input_image, prompt, ddim_steps, num_samples, scale, seed):
+    init_image = input_image["image"].convert("RGB")
+    init_mask = input_image["mask"].convert("RGB")
+    image = pad_image(init_image) # resize to integer multiple of 32
+    mask = pad_image(init_mask) # resize to integer multiple of 32
+    width, height = image.size
+    print("Inpainting...", width, height)
+
+    result = inpaint(
+        sampler=sampler,
+        image=image,
+        mask=mask,
+        prompt=prompt,
+        seed=seed,
+        scale=scale,
+        ddim_steps=ddim_steps,
+        num_samples=num_samples,
+        h=height, w=width
+    )
+
+    return result
+
+
+sampler = initialize_model(sys.argv[1], sys.argv[2])
+
+block = gr.Blocks().queue()
+with block:
+    with gr.Row():
+        gr.Markdown("## Stable Diffusion Inpainting")
+
+    with gr.Row():
+        with gr.Column():
+            input_image = gr.Image(source='upload', tool='sketch', type="pil")
+            prompt = gr.Textbox(label="Prompt")
+            run_button = gr.Button(label="Run")
+            with gr.Accordion("Advanced options", open=False):
+                num_samples = gr.Slider(
+                    label="Images", minimum=1, maximum=4, value=4, step=1)
+                ddim_steps = gr.Slider(label="Steps", minimum=1,
+                                       maximum=50, value=45, step=1)
+                scale = gr.Slider(
+                    label="Guidance Scale", minimum=0.1, maximum=30.0, value=10, step=0.1
+                )
+                seed = gr.Slider(
+                    label="Seed",
+                    minimum=0,
+                    maximum=2147483647,
+                    step=1,
+                    randomize=True,
+                )
+        with gr.Column():
+            gallery = gr.Gallery(label="Generated images", show_label=False).style(
+                grid=[2], height="auto")
+
+    run_button.click(fn=predict, inputs=[
+                     input_image, prompt, ddim_steps, num_samples, scale, seed], outputs=[gallery])
+
+
+block.launch()

scripts/gradio/superresolution.py

@@ -0,0 +1,197 @@
+import sys
+import torch
+import numpy as np
+import gradio as gr
+from PIL import Image
+from omegaconf import OmegaConf
+from einops import repeat, rearrange
+from pytorch_lightning import seed_everything
+from imwatermark import WatermarkEncoder
+
+from scripts.txt2img import put_watermark
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.models.diffusion.ddpm import LatentUpscaleDiffusion, LatentUpscaleFinetuneDiffusion
+from ldm.util import exists, instantiate_from_config
+
+
+torch.set_grad_enabled(False)
+
+
+def initialize_model(config, ckpt):
+    config = OmegaConf.load(config)
+    model = instantiate_from_config(config.model)
+    model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False)
+
+    device = torch.device(
+        "cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = model.to(device)
+    sampler = DDIMSampler(model)
+    return sampler
+
+
+def make_batch_sd(
+        image,
+        txt,
+        device,
+        num_samples=1,
+):
+    image = np.array(image.convert("RGB"))
+    image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+    batch = {
+        "lr": rearrange(image, 'h w c -> 1 c h w'),
+        "txt": num_samples * [txt],
+    }
+    batch["lr"] = repeat(batch["lr"].to(device=device),
+                         "1 ... -> n ...", n=num_samples)
+    return batch
+
+
+def make_noise_augmentation(model, batch, noise_level=None):
+    x_low = batch[model.low_scale_key]
+    x_low = x_low.to(memory_format=torch.contiguous_format).float()
+    x_aug, noise_level = model.low_scale_model(x_low, noise_level)
+    return x_aug, noise_level
+
+
+def paint(sampler, image, prompt, seed, scale, h, w, steps, num_samples=1, callback=None, eta=0., noise_level=None):
+    device = torch.device(
+        "cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = sampler.model
+    seed_everything(seed)
+    prng = np.random.RandomState(seed)
+    start_code = prng.randn(num_samples, model.channels, h, w)
+    start_code = torch.from_numpy(start_code).to(
+        device=device, dtype=torch.float32)
+
+    print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...")
+    wm = "SDV2"
+    wm_encoder = WatermarkEncoder()
+    wm_encoder.set_watermark('bytes', wm.encode('utf-8'))
+    with torch.no_grad(),\
+            torch.autocast("cuda"):
+        batch = make_batch_sd(
+            image, txt=prompt, device=device, num_samples=num_samples)
+        c = model.cond_stage_model.encode(batch["txt"])
+        c_cat = list()
+        if isinstance(model, LatentUpscaleFinetuneDiffusion):
+            for ck in model.concat_keys:
+                cc = batch[ck]
+                if exists(model.reshuffle_patch_size):
+                    assert isinstance(model.reshuffle_patch_size, int)
+                    cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w',
+                                   p1=model.reshuffle_patch_size, p2=model.reshuffle_patch_size)
+                c_cat.append(cc)
+            c_cat = torch.cat(c_cat, dim=1)
+            # cond
+            cond = {"c_concat": [c_cat], "c_crossattn": [c]}
+            # uncond cond
+            uc_cross = model.get_unconditional_conditioning(num_samples, "")
+            uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]}
+        elif isinstance(model, LatentUpscaleDiffusion):
+            x_augment, noise_level = make_noise_augmentation(
+                model, batch, noise_level)
+            cond = {"c_concat": [x_augment],
+                    "c_crossattn": [c], "c_adm": noise_level}
+            # uncond cond
+            uc_cross = model.get_unconditional_conditioning(num_samples, "")
+            uc_full = {"c_concat": [x_augment], "c_crossattn": [
+                uc_cross], "c_adm": noise_level}
+        else:
+            raise NotImplementedError()
+
+        shape = [model.channels, h, w]
+        samples, intermediates = sampler.sample(
+            steps,
+            num_samples,
+            shape,
+            cond,
+            verbose=False,
+            eta=eta,
+            unconditional_guidance_scale=scale,
+            unconditional_conditioning=uc_full,
+            x_T=start_code,
+            callback=callback
+        )
+    with torch.no_grad():
+        x_samples_ddim = model.decode_first_stage(samples)
+    result = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
+    result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255
+    return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result]
+
+
+def pad_image(input_image):
+    pad_w, pad_h = np.max(((2, 2), np.ceil(
+        np.array(input_image.size) / 64).astype(int)), axis=0) * 64 - input_image.size
+    im_padded = Image.fromarray(
+        np.pad(np.array(input_image), ((0, pad_h), (0, pad_w), (0, 0)), mode='edge'))
+    return im_padded
+
+
+def predict(input_image, prompt, steps, num_samples, scale, seed, eta, noise_level):
+    init_image = input_image.convert("RGB")
+    image = pad_image(init_image)  # resize to integer multiple of 32
+    width, height = image.size
+
+    noise_level = torch.Tensor(
+        num_samples * [noise_level]).to(sampler.model.device).long()
+    sampler.make_schedule(steps, ddim_eta=eta, verbose=True)
+    result = paint(
+        sampler=sampler,
+        image=image,
+        prompt=prompt,
+        seed=seed,
+        scale=scale,
+        h=height, w=width, steps=steps,
+        num_samples=num_samples,
+        callback=None,
+        noise_level=noise_level
+    )
+    return result
+
+
+sampler = initialize_model(sys.argv[1], sys.argv[2])
+
+block = gr.Blocks().queue()
+with block:
+    with gr.Row():
+        gr.Markdown("## Stable Diffusion Upscaling")
+
+    with gr.Row():
+        with gr.Column():
+            input_image = gr.Image(source='upload', type="pil")
+            gr.Markdown(
+                "Tip: Add a description of the object that should be upscaled, e.g.: 'a professional photograph of a cat")
+            prompt = gr.Textbox(label="Prompt")
+            run_button = gr.Button(label="Run")
+            with gr.Accordion("Advanced options", open=False):
+                num_samples = gr.Slider(
+                    label="Number of Samples", minimum=1, maximum=4, value=1, step=1)
+                steps = gr.Slider(label="DDIM Steps", minimum=2,
+                                  maximum=200, value=75, step=1)
+                scale = gr.Slider(
+                    label="Scale", minimum=0.1, maximum=30.0, value=10, step=0.1
+                )
+                seed = gr.Slider(
+                    label="Seed",
+                    minimum=0,
+                    maximum=2147483647,
+                    step=1,
+                    randomize=True,
+                )
+                eta = gr.Number(label="eta (DDIM)",
+                                value=0.0, min=0.0, max=1.0)
+                noise_level = None
+                if isinstance(sampler.model, LatentUpscaleDiffusion):
+                    # TODO: make this work for all models
+                    noise_level = gr.Number(
+                        label="Noise Augmentation", min=0, max=350, value=20, step=1)
+
+        with gr.Column():
+            gallery = gr.Gallery(label="Generated images", show_label=False).style(
+                grid=[2], height="auto")
+
+    run_button.click(fn=predict, inputs=[
+                     input_image, prompt, steps, num_samples, scale, seed, eta, noise_level], outputs=[gallery])
+
+
+block.launch()

scripts/img2img.py

@@ -0,0 +1,279 @@
+"""make variations of input image"""
+
+import argparse, os
+import PIL
+import torch
+import numpy as np
+from omegaconf import OmegaConf
+from PIL import Image
+from tqdm import tqdm, trange
+from itertools import islice
+from einops import rearrange, repeat
+from torchvision.utils import make_grid
+from torch import autocast
+from contextlib import nullcontext
+from pytorch_lightning import seed_everything
+from imwatermark import WatermarkEncoder
+
+
+from scripts.txt2img import put_watermark
+from ldm.util import instantiate_from_config
+from ldm.models.diffusion.ddim import DDIMSampler
+
+
+def chunk(it, size):
+    it = iter(it)
+    return iter(lambda: tuple(islice(it, size)), ())
+
+
+def load_model_from_config(config, ckpt, verbose=False):
+    print(f"Loading model from {ckpt}")
+    pl_sd = torch.load(ckpt, map_location="cpu")
+    if "global_step" in pl_sd:
+        print(f"Global Step: {pl_sd['global_step']}")
+    sd = pl_sd["state_dict"]
+    model = instantiate_from_config(config.model)
+    m, u = model.load_state_dict(sd, strict=False)
+    if len(m) > 0 and verbose:
+        print("missing keys:")
+        print(m)
+    if len(u) > 0 and verbose:
+        print("unexpected keys:")
+        print(u)
+
+    model.cuda()
+    model.eval()
+    return model
+
+
+def load_img(path):
+    image = Image.open(path).convert("RGB")
+    w, h = image.size
+    print(f"loaded input image of size ({w}, {h}) from {path}")
+    w, h = map(lambda x: x - x % 64, (w, h))  # resize to integer multiple of 64
+    image = image.resize((w, h), resample=PIL.Image.LANCZOS)
+    image = np.array(image).astype(np.float32) / 255.0
+    image = image[None].transpose(0, 3, 1, 2)
+    image = torch.from_numpy(image)
+    return 2. * image - 1.
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--prompt",
+        type=str,
+        nargs="?",
+        default="a painting of a virus monster playing guitar",
+        help="the prompt to render"
+    )
+
+    parser.add_argument(
+        "--init-img",
+        type=str,
+        nargs="?",
+        help="path to the input image"
+    )
+
+    parser.add_argument(
+        "--outdir",
+        type=str,
+        nargs="?",
+        help="dir to write results to",
+        default="outputs/img2img-samples"
+    )
+
+    parser.add_argument(
+        "--ddim_steps",
+        type=int,
+        default=50,
+        help="number of ddim sampling steps",
+    )
+
+    parser.add_argument(
+        "--fixed_code",
+        action='store_true',
+        help="if enabled, uses the same starting code across all samples ",
+    )
+
+    parser.add_argument(
+        "--ddim_eta",
+        type=float,
+        default=0.0,
+        help="ddim eta (eta=0.0 corresponds to deterministic sampling",
+    )
+    parser.add_argument(
+        "--n_iter",
+        type=int,
+        default=1,
+        help="sample this often",
+    )
+
+    parser.add_argument(
+        "--C",
+        type=int,
+        default=4,
+        help="latent channels",
+    )
+    parser.add_argument(
+        "--f",
+        type=int,
+        default=8,
+        help="downsampling factor, most often 8 or 16",
+    )
+
+    parser.add_argument(
+        "--n_samples",
+        type=int,
+        default=2,
+        help="how many samples to produce for each given prompt. A.k.a batch size",
+    )
+
+    parser.add_argument(
+        "--n_rows",
+        type=int,
+        default=0,
+        help="rows in the grid (default: n_samples)",
+    )
+
+    parser.add_argument(
+        "--scale",
+        type=float,
+        default=9.0,
+        help="unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))",
+    )
+
+    parser.add_argument(
+        "--strength",
+        type=float,
+        default=0.8,
+        help="strength for noising/unnoising. 1.0 corresponds to full destruction of information in init image",
+    )
+
+    parser.add_argument(
+        "--from-file",
+        type=str,
+        help="if specified, load prompts from this file",
+    )
+    parser.add_argument(
+        "--config",
+        type=str,
+        default="configs/stable-diffusion/v2-inference.yaml",
+        help="path to config which constructs model",
+    )
+    parser.add_argument(
+        "--ckpt",
+        type=str,
+        help="path to checkpoint of model",
+    )
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="the seed (for reproducible sampling)",
+    )
+    parser.add_argument(
+        "--precision",
+        type=str,
+        help="evaluate at this precision",
+        choices=["full", "autocast"],
+        default="autocast"
+    )
+
+    opt = parser.parse_args()
+    seed_everything(opt.seed)
+
+    config = OmegaConf.load(f"{opt.config}")
+    model = load_model_from_config(config, f"{opt.ckpt}")
+
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = model.to(device)
+
+    sampler = DDIMSampler(model)
+
+    os.makedirs(opt.outdir, exist_ok=True)
+    outpath = opt.outdir
+
+    print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...")
+    wm = "SDV2"
+    wm_encoder = WatermarkEncoder()
+    wm_encoder.set_watermark('bytes', wm.encode('utf-8'))
+
+    batch_size = opt.n_samples
+    n_rows = opt.n_rows if opt.n_rows > 0 else batch_size
+    if not opt.from_file:
+        prompt = opt.prompt
+        assert prompt is not None
+        data = [batch_size * [prompt]]
+
+    else:
+        print(f"reading prompts from {opt.from_file}")
+        with open(opt.from_file, "r") as f:
+            data = f.read().splitlines()
+            data = list(chunk(data, batch_size))
+
+    sample_path = os.path.join(outpath, "samples")
+    os.makedirs(sample_path, exist_ok=True)
+    base_count = len(os.listdir(sample_path))
+    grid_count = len(os.listdir(outpath)) - 1
+
+    assert os.path.isfile(opt.init_img)
+    init_image = load_img(opt.init_img).to(device)
+    init_image = repeat(init_image, '1 ... -> b ...', b=batch_size)
+    init_latent = model.get_first_stage_encoding(model.encode_first_stage(init_image))  # move to latent space
+
+    sampler.make_schedule(ddim_num_steps=opt.ddim_steps, ddim_eta=opt.ddim_eta, verbose=False)
+
+    assert 0. <= opt.strength <= 1., 'can only work with strength in [0.0, 1.0]'
+    t_enc = int(opt.strength * opt.ddim_steps)
+    print(f"target t_enc is {t_enc} steps")
+
+    precision_scope = autocast if opt.precision == "autocast" else nullcontext
+    with torch.no_grad():
+        with precision_scope("cuda"):
+            with model.ema_scope():
+                all_samples = list()
+                for n in trange(opt.n_iter, desc="Sampling"):
+                    for prompts in tqdm(data, desc="data"):
+                        uc = None
+                        if opt.scale != 1.0:
+                            uc = model.get_learned_conditioning(batch_size * [""])
+                        if isinstance(prompts, tuple):
+                            prompts = list(prompts)
+                        c = model.get_learned_conditioning(prompts)
+
+                        # encode (scaled latent)
+                        z_enc = sampler.stochastic_encode(init_latent, torch.tensor([t_enc] * batch_size).to(device))
+                        # decode it
+                        samples = sampler.decode(z_enc, c, t_enc, unconditional_guidance_scale=opt.scale,
+                                                 unconditional_conditioning=uc, )
+
+                        x_samples = model.decode_first_stage(samples)
+                        x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0)
+
+                        for x_sample in x_samples:
+                            x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
+                            img = Image.fromarray(x_sample.astype(np.uint8))
+                            img = put_watermark(img, wm_encoder)
+                            img.save(os.path.join(sample_path, f"{base_count:05}.png"))
+                            base_count += 1
+                        all_samples.append(x_samples)
+
+                # additionally, save as grid
+                grid = torch.stack(all_samples, 0)
+                grid = rearrange(grid, 'n b c h w -> (n b) c h w')
+                grid = make_grid(grid, nrow=n_rows)
+
+                # to image
+                grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy()
+                grid = Image.fromarray(grid.astype(np.uint8))
+                grid = put_watermark(grid, wm_encoder)
+                grid.save(os.path.join(outpath, f'grid-{grid_count:04}.png'))
+                grid_count += 1
+
+    print(f"Your samples are ready and waiting for you here: \n{outpath} \nEnjoy.")
+
+
+if __name__ == "__main__":
+    main()

scripts/streamlit/depth2img.py

@@ -0,0 +1,157 @@
+import sys
+import torch
+import numpy as np
+import streamlit as st
+from PIL import Image
+from omegaconf import OmegaConf
+from einops import repeat, rearrange
+from pytorch_lightning import seed_everything
+from imwatermark import WatermarkEncoder
+
+from scripts.txt2img import put_watermark
+from ldm.util import instantiate_from_config
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.data.util import AddMiDaS
+
+torch.set_grad_enabled(False)
+
+
+@st.cache(allow_output_mutation=True)
+def initialize_model(config, ckpt):
+    config = OmegaConf.load(config)
+    model = instantiate_from_config(config.model)
+    model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False)
+
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = model.to(device)
+    sampler = DDIMSampler(model)
+    return sampler
+
+
+def make_batch_sd(
+        image,
+        txt,
+        device,
+        num_samples=1,
+        model_type="dpt_hybrid"
+):
+    image = np.array(image.convert("RGB"))
+    image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+    # sample['jpg'] is tensor hwc in [-1, 1] at this point
+    midas_trafo = AddMiDaS(model_type=model_type)
+    batch = {
+        "jpg": image,
+        "txt": num_samples * [txt],
+    }
+    batch = midas_trafo(batch)
+    batch["jpg"] = rearrange(batch["jpg"], 'h w c -> 1 c h w')
+    batch["jpg"] = repeat(batch["jpg"].to(device=device), "1 ... -> n ...", n=num_samples)
+    batch["midas_in"] = repeat(torch.from_numpy(batch["midas_in"][None, ...]).to(device=device), "1 ... -> n ...", n=num_samples)
+    return batch
+
+
+def paint(sampler, image, prompt, t_enc, seed, scale, num_samples=1, callback=None,
+          do_full_sample=False):
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = sampler.model
+    seed_everything(seed)
+
+    print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...")
+    wm = "SDV2"
+    wm_encoder = WatermarkEncoder()
+    wm_encoder.set_watermark('bytes', wm.encode('utf-8'))
+
+    with torch.no_grad(),\
+            torch.autocast("cuda"):
+        batch = make_batch_sd(image, txt=prompt, device=device, num_samples=num_samples)
+        z = model.get_first_stage_encoding(model.encode_first_stage(batch[model.first_stage_key]))  # move to latent space
+        c = model.cond_stage_model.encode(batch["txt"])
+        c_cat = list()
+        for ck in model.concat_keys:
+            cc = batch[ck]
+            cc = model.depth_model(cc)
+            depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3],
+                                                                                           keepdim=True)
+            display_depth = (cc - depth_min) / (depth_max - depth_min)
+            st.image(Image.fromarray((display_depth[0, 0, ...].cpu().numpy() * 255.).astype(np.uint8)))
+            cc = torch.nn.functional.interpolate(
+                cc,
+                size=z.shape[2:],
+                mode="bicubic",
+                align_corners=False,
+            )
+            depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3],
+                                                                                           keepdim=True)
+            cc = 2. * (cc - depth_min) / (depth_max - depth_min) - 1.
+            c_cat.append(cc)
+        c_cat = torch.cat(c_cat, dim=1)
+        # cond
+        cond = {"c_concat": [c_cat], "c_crossattn": [c]}
+
+        # uncond cond
+        uc_cross = model.get_unconditional_conditioning(num_samples, "")
+        uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]}
+        if not do_full_sample:
+            # encode (scaled latent)
+            z_enc = sampler.stochastic_encode(z, torch.tensor([t_enc] * num_samples).to(model.device))
+        else:
+            z_enc = torch.randn_like(z)
+        # decode it
+        samples = sampler.decode(z_enc, cond, t_enc, unconditional_guidance_scale=scale,
+                                 unconditional_conditioning=uc_full, callback=callback)
+        x_samples_ddim = model.decode_first_stage(samples)
+        result = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
+        result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255
+    return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result]
+
+
+def run():
+    st.title("Stable Diffusion Depth2Img")
+    # run via streamlit run scripts/demo/depth2img.py <path-tp-config> <path-to-ckpt>
+    sampler = initialize_model(sys.argv[1], sys.argv[2])
+
+    image = st.file_uploader("Image", ["jpg", "png"])
+    if image:
+        image = Image.open(image)
+        w, h = image.size
+        st.text(f"loaded input image of size ({w}, {h})")
+        width, height = map(lambda x: x - x % 64, (w, h))  # resize to integer multiple of 64
+        image = image.resize((width, height))
+        st.text(f"resized input image to size ({width}, {height} (w, h))")
+        st.image(image)
+
+        prompt = st.text_input("Prompt")
+
+        seed = st.number_input("Seed", min_value=0, max_value=1000000, value=0)
+        num_samples = st.number_input("Number of Samples", min_value=1, max_value=64, value=1)
+        scale = st.slider("Scale", min_value=0.1, max_value=30.0, value=9.0, step=0.1)
+        steps = st.slider("DDIM Steps", min_value=0, max_value=50, value=50, step=1)
+        strength = st.slider("Strength", min_value=0., max_value=1., value=0.9)
+
+        t_progress = st.progress(0)
+        def t_callback(t):
+            t_progress.progress(min((t + 1) / t_enc, 1.))
+
+        assert 0. <= strength <= 1., 'can only work with strength in [0.0, 1.0]'
+        do_full_sample = strength == 1.
+        t_enc = min(int(strength * steps), steps-1)
+        sampler.make_schedule(steps, ddim_eta=0., verbose=True)
+        if st.button("Sample"):
+            result = paint(
+                sampler=sampler,
+                image=image,
+                prompt=prompt,
+                t_enc=t_enc,
+                seed=seed,
+                scale=scale,
+                num_samples=num_samples,
+                callback=t_callback,
+                do_full_sample=do_full_sample,
+            )
+            st.write("Result")
+            for image in result:
+                st.image(image, output_format='PNG')
+
+
+if __name__ == "__main__":
+    run()

scripts/streamlit/inpainting.py

@@ -0,0 +1,195 @@
+import sys
+import cv2
+import torch
+import numpy as np
+import streamlit as st
+from PIL import Image
+from omegaconf import OmegaConf
+from einops import repeat
+from streamlit_drawable_canvas import st_canvas
+from imwatermark import WatermarkEncoder
+
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.util import instantiate_from_config
+
+
+torch.set_grad_enabled(False)
+
+
+def put_watermark(img, wm_encoder=None):
+    if wm_encoder is not None:
+        img = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
+        img = wm_encoder.encode(img, 'dwtDct')
+        img = Image.fromarray(img[:, :, ::-1])
+    return img
+
+
+@st.cache(allow_output_mutation=True)
+def initialize_model(config, ckpt):
+    config = OmegaConf.load(config)
+    model = instantiate_from_config(config.model)
+
+    model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False)
+
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = model.to(device)
+    sampler = DDIMSampler(model)
+
+    return sampler
+
+
+def make_batch_sd(
+        image,
+        mask,
+        txt,
+        device,
+        num_samples=1):
+    image = np.array(image.convert("RGB"))
+    image = image[None].transpose(0, 3, 1, 2)
+    image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+    mask = np.array(mask.convert("L"))
+    mask = mask.astype(np.float32) / 255.0
+    mask = mask[None, None]
+    mask[mask < 0.5] = 0
+    mask[mask >= 0.5] = 1
+    mask = torch.from_numpy(mask)
+
+    masked_image = image * (mask < 0.5)
+
+    batch = {
+        "image": repeat(image.to(device=device), "1 ... -> n ...", n=num_samples),
+        "txt": num_samples * [txt],
+        "mask": repeat(mask.to(device=device), "1 ... -> n ...", n=num_samples),
+        "masked_image": repeat(masked_image.to(device=device), "1 ... -> n ...", n=num_samples),
+    }
+    return batch
+
+
+def inpaint(sampler, image, mask, prompt, seed, scale, ddim_steps, num_samples=1, w=512, h=512, eta=1.):
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = sampler.model
+
+    print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...")
+    wm = "SDV2"
+    wm_encoder = WatermarkEncoder()
+    wm_encoder.set_watermark('bytes', wm.encode('utf-8'))
+
+    prng = np.random.RandomState(seed)
+    start_code = prng.randn(num_samples, 4, h // 8, w // 8)
+    start_code = torch.from_numpy(start_code).to(device=device, dtype=torch.float32)
+
+    with torch.no_grad(), \
+            torch.autocast("cuda"):
+            batch = make_batch_sd(image, mask, txt=prompt, device=device, num_samples=num_samples)
+
+            c = model.cond_stage_model.encode(batch["txt"])
+
+            c_cat = list()
+            for ck in model.concat_keys:
+                cc = batch[ck].float()
+                if ck != model.masked_image_key:
+                    bchw = [num_samples, 4, h // 8, w // 8]
+                    cc = torch.nn.functional.interpolate(cc, size=bchw[-2:])
+                else:
+                    cc = model.get_first_stage_encoding(model.encode_first_stage(cc))
+                c_cat.append(cc)
+            c_cat = torch.cat(c_cat, dim=1)
+
+            # cond
+            cond = {"c_concat": [c_cat], "c_crossattn": [c]}
+
+            # uncond cond
+            uc_cross = model.get_unconditional_conditioning(num_samples, "")
+            uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]}
+
+            shape = [model.channels, h // 8, w // 8]
+            samples_cfg, intermediates = sampler.sample(
+                ddim_steps,
+                num_samples,
+                shape,
+                cond,
+                verbose=False,
+                eta=eta,
+                unconditional_guidance_scale=scale,
+                unconditional_conditioning=uc_full,
+                x_T=start_code,
+            )
+            x_samples_ddim = model.decode_first_stage(samples_cfg)
+
+            result = torch.clamp((x_samples_ddim + 1.0) / 2.0,
+                                 min=0.0, max=1.0)
+
+            result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255
+    return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result]
+
+
+def run():
+    st.title("Stable Diffusion Inpainting")
+
+    sampler = initialize_model(sys.argv[1], sys.argv[2])
+
+    image = st.file_uploader("Image", ["jpg", "png"])
+    if image:
+        image = Image.open(image)
+        w, h = image.size
+        print(f"loaded input image of size ({w}, {h})")
+        width, height = map(lambda x: x - x % 64, (w, h))  # resize to integer multiple of 32
+        image = image.resize((width, height))
+
+        prompt = st.text_input("Prompt")
+
+        seed = st.number_input("Seed", min_value=0, max_value=1000000, value=0)
+        num_samples = st.number_input("Number of Samples", min_value=1, max_value=64, value=1)
+        scale = st.slider("Scale", min_value=0.1, max_value=30.0, value=10., step=0.1)
+        ddim_steps = st.slider("DDIM Steps", min_value=0, max_value=50, value=50, step=1)
+        eta = st.sidebar.number_input("eta (DDIM)", value=0., min_value=0., max_value=1.)
+
+        fill_color = "rgba(255, 255, 255, 0.0)"
+        stroke_width = st.number_input("Brush Size",
+                                       value=64,
+                                       min_value=1,
+                                       max_value=100)
+        stroke_color = "rgba(255, 255, 255, 1.0)"
+        bg_color = "rgba(0, 0, 0, 1.0)"
+        drawing_mode = "freedraw"
+
+        st.write("Canvas")
+        st.caption(
+            "Draw a mask to inpaint, then click the 'Send to Streamlit' button (bottom left, with an arrow on it).")
+        canvas_result = st_canvas(
+            fill_color=fill_color,
+            stroke_width=stroke_width,
+            stroke_color=stroke_color,
+            background_color=bg_color,
+            background_image=image,
+            update_streamlit=False,
+            height=height,
+            width=width,
+            drawing_mode=drawing_mode,
+            key="canvas",
+        )
+        if canvas_result:
+            mask = canvas_result.image_data
+            mask = mask[:, :, -1] > 0
+            if mask.sum() > 0:
+                mask = Image.fromarray(mask)
+
+                result = inpaint(
+                    sampler=sampler,
+                    image=image,
+                    mask=mask,
+                    prompt=prompt,
+                    seed=seed,
+                    scale=scale,
+                    ddim_steps=ddim_steps,
+                    num_samples=num_samples,
+                    h=height, w=width, eta=eta
+                )
+                st.write("Inpainted")
+                for image in result:
+                    st.image(image, output_format='PNG')
+
+
+if __name__ == "__main__":
+    run()

scripts/streamlit/superresolution.py

@@ -0,0 +1,170 @@
+import sys
+import torch
+import numpy as np
+import streamlit as st
+from PIL import Image
+from omegaconf import OmegaConf
+from einops import repeat, rearrange
+from pytorch_lightning import seed_everything
+from imwatermark import WatermarkEncoder
+
+from scripts.txt2img import put_watermark
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.models.diffusion.ddpm import LatentUpscaleDiffusion, LatentUpscaleFinetuneDiffusion
+from ldm.util import exists, instantiate_from_config
+
+
+torch.set_grad_enabled(False)
+
+
+@st.cache(allow_output_mutation=True)
+def initialize_model(config, ckpt):
+    config = OmegaConf.load(config)
+    model = instantiate_from_config(config.model)
+    model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False)
+
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = model.to(device)
+    sampler = DDIMSampler(model)
+    return sampler
+
+
+def make_batch_sd(
+        image,
+        txt,
+        device,
+        num_samples=1,
+):
+    image = np.array(image.convert("RGB"))
+    image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+    batch = {
+        "lr": rearrange(image, 'h w c -> 1 c h w'),
+        "txt": num_samples * [txt],
+    }
+    batch["lr"] = repeat(batch["lr"].to(device=device), "1 ... -> n ...", n=num_samples)
+    return batch
+
+
+def make_noise_augmentation(model, batch, noise_level=None):
+    x_low = batch[model.low_scale_key]
+    x_low = x_low.to(memory_format=torch.contiguous_format).float()
+    x_aug, noise_level = model.low_scale_model(x_low, noise_level)
+    return x_aug, noise_level
+
+
+def paint(sampler, image, prompt, seed, scale, h, w, steps, num_samples=1, callback=None, eta=0., noise_level=None):
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = sampler.model
+    seed_everything(seed)
+    prng = np.random.RandomState(seed)
+    start_code = prng.randn(num_samples, model.channels, h , w)
+    start_code = torch.from_numpy(start_code).to(device=device, dtype=torch.float32)
+
+    print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...")
+    wm = "SDV2"
+    wm_encoder = WatermarkEncoder()
+    wm_encoder.set_watermark('bytes', wm.encode('utf-8'))
+    with torch.no_grad(),\
+            torch.autocast("cuda"):
+        batch = make_batch_sd(image, txt=prompt, device=device, num_samples=num_samples)
+        c = model.cond_stage_model.encode(batch["txt"])
+        c_cat = list()
+        if isinstance(model, LatentUpscaleFinetuneDiffusion):
+            for ck in model.concat_keys:
+                cc = batch[ck]
+                if exists(model.reshuffle_patch_size):
+                    assert isinstance(model.reshuffle_patch_size, int)
+                    cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w',
+                                   p1=model.reshuffle_patch_size, p2=model.reshuffle_patch_size)
+                c_cat.append(cc)
+            c_cat = torch.cat(c_cat, dim=1)
+            # cond
+            cond = {"c_concat": [c_cat], "c_crossattn": [c]}
+            # uncond cond
+            uc_cross = model.get_unconditional_conditioning(num_samples, "")
+            uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]}
+        elif isinstance(model, LatentUpscaleDiffusion):
+            x_augment, noise_level = make_noise_augmentation(model, batch, noise_level)
+            cond = {"c_concat": [x_augment], "c_crossattn": [c], "c_adm": noise_level}
+            # uncond cond
+            uc_cross = model.get_unconditional_conditioning(num_samples, "")
+            uc_full = {"c_concat": [x_augment], "c_crossattn": [uc_cross], "c_adm": noise_level}
+        else:
+            raise NotImplementedError()
+
+        shape = [model.channels, h, w]
+        samples, intermediates = sampler.sample(
+            steps,
+            num_samples,
+            shape,
+            cond,
+            verbose=False,
+            eta=eta,
+            unconditional_guidance_scale=scale,
+            unconditional_conditioning=uc_full,
+            x_T=start_code,
+            callback=callback
+        )
+    with torch.no_grad():
+        x_samples_ddim = model.decode_first_stage(samples)
+    result = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
+    result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255
+    st.text(f"upscaled image shape: {result.shape}")
+    return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result]
+
+
+def run():
+    st.title("Stable Diffusion Upscaling")
+    # run via streamlit run scripts/demo/depth2img.py <path-tp-config> <path-to-ckpt>
+    sampler = initialize_model(sys.argv[1], sys.argv[2])
+
+    image = st.file_uploader("Image", ["jpg", "png"])
+    if image:
+        image = Image.open(image)
+        w, h = image.size
+        st.text(f"loaded input image of size ({w}, {h})")
+        width, height = map(lambda x: x - x % 64, (w, h))  # resize to integer multiple of 64
+        image = image.resize((width, height))
+        st.text(f"resized input image to size ({width}, {height} (w, h))")
+        st.image(image)
+
+        st.write(f"\n Tip: Add a description of the object that should be upscaled, e.g.: 'a professional photograph of a cat'")
+        prompt = st.text_input("Prompt", "a high quality professional photograph")
+
+        seed = st.number_input("Seed", min_value=0, max_value=1000000, value=0)
+        num_samples = st.number_input("Number of Samples", min_value=1, max_value=64, value=1)
+        scale = st.slider("Scale", min_value=0.1, max_value=30.0, value=9.0, step=0.1)
+        steps = st.slider("DDIM Steps", min_value=2, max_value=250, value=50, step=1)
+        eta = st.sidebar.number_input("eta (DDIM)", value=0., min_value=0., max_value=1.)
+
+        noise_level = None
+        if isinstance(sampler.model, LatentUpscaleDiffusion):
+            # TODO: make this work for all models
+            noise_level = st.sidebar.number_input("Noise Augmentation", min_value=0, max_value=350, value=20)
+            noise_level = torch.Tensor(num_samples * [noise_level]).to(sampler.model.device).long()
+
+        t_progress = st.progress(0)
+        def t_callback(t):
+            t_progress.progress(min((t + 1) / steps, 1.))
+
+        sampler.make_schedule(steps, ddim_eta=eta, verbose=True)
+        if st.button("Sample"):
+            result = paint(
+                sampler=sampler,
+                image=image,
+                prompt=prompt,
+                seed=seed,
+                scale=scale,
+                h=height, w=width, steps=steps,
+                num_samples=num_samples,
+                callback=t_callback,
+                noise_level=noise_level,
+                eta=eta
+            )
+            st.write("Result")
+            for image in result:
+                st.image(image, output_format='PNG')
+
+
+if __name__ == "__main__":
+    run()

scripts/tests/test_watermark.py

@@ -0,0 +1,18 @@
+import cv2
+import fire
+from imwatermark import WatermarkDecoder
+
+
+def testit(img_path):
+    bgr = cv2.imread(img_path)
+    decoder = WatermarkDecoder('bytes', 136)
+    watermark = decoder.decode(bgr, 'dwtDct')
+    try:
+        dec = watermark.decode('utf-8')
+    except:
+        dec = "null"
+    print(dec)
+
+
+if __name__ == "__main__":
+    fire.Fire(testit)

scripts/txt2img.py

@@ -0,0 +1,289 @@
+import argparse, os
+import cv2
+import torch
+import numpy as np
+from omegaconf import OmegaConf
+from PIL import Image
+from tqdm import tqdm, trange
+from itertools import islice
+from einops import rearrange
+from torchvision.utils import make_grid
+from pytorch_lightning import seed_everything
+from torch import autocast
+from contextlib import nullcontext
+from imwatermark import WatermarkEncoder
+
+from ldm.util import instantiate_from_config
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.models.diffusion.plms import PLMSSampler
+from ldm.models.diffusion.dpm_solver import DPMSolverSampler
+
+torch.set_grad_enabled(False)
+
+def chunk(it, size):
+    it = iter(it)
+    return iter(lambda: tuple(islice(it, size)), ())
+
+
+def load_model_from_config(config, ckpt, verbose=False):
+    print(f"Loading model from {ckpt}")
+    pl_sd = torch.load(ckpt, map_location="cpu")
+    if "global_step" in pl_sd:
+        print(f"Global Step: {pl_sd['global_step']}")
+    sd = pl_sd["state_dict"]
+    model = instantiate_from_config(config.model)
+    m, u = model.load_state_dict(sd, strict=False)
+    if len(m) > 0 and verbose:
+        print("missing keys:")
+        print(m)
+    if len(u) > 0 and verbose:
+        print("unexpected keys:")
+        print(u)
+
+    model.cuda()
+    model.eval()
+    return model
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--prompt",
+        type=str,
+        nargs="?",
+        default="a professional photograph of an astronaut riding a triceratops",
+        help="the prompt to render"
+    )
+    parser.add_argument(
+        "--outdir",
+        type=str,
+        nargs="?",
+        help="dir to write results to",
+        default="outputs/txt2img-samples"
+    )
+    parser.add_argument(
+        "--steps",
+        type=int,
+        default=50,
+        help="number of ddim sampling steps",
+    )
+    parser.add_argument(
+        "--plms",
+        action='store_true',
+        help="use plms sampling",
+    )
+    parser.add_argument(
+        "--dpm",
+        action='store_true',
+        help="use DPM (2) sampler",
+    )
+    parser.add_argument(
+        "--fixed_code",
+        action='store_true',
+        help="if enabled, uses the same starting code across all samples ",
+    )
+    parser.add_argument(
+        "--ddim_eta",
+        type=float,
+        default=0.0,
+        help="ddim eta (eta=0.0 corresponds to deterministic sampling",
+    )
+    parser.add_argument(
+        "--n_iter",
+        type=int,
+        default=3,
+        help="sample this often",
+    )
+    parser.add_argument(
+        "--H",
+        type=int,
+        default=512,
+        help="image height, in pixel space",
+    )
+    parser.add_argument(
+        "--W",
+        type=int,
+        default=512,
+        help="image width, in pixel space",
+    )
+    parser.add_argument(
+        "--C",
+        type=int,
+        default=4,
+        help="latent channels",
+    )
+    parser.add_argument(
+        "--f",
+        type=int,
+        default=8,
+        help="downsampling factor, most often 8 or 16",
+    )
+    parser.add_argument(
+        "--n_samples",
+        type=int,
+        default=3,
+        help="how many samples to produce for each given prompt. A.k.a batch size",
+    )
+    parser.add_argument(
+        "--n_rows",
+        type=int,
+        default=0,
+        help="rows in the grid (default: n_samples)",
+    )
+    parser.add_argument(
+        "--scale",
+        type=float,
+        default=9.0,
+        help="unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))",
+    )
+    parser.add_argument(
+        "--from-file",
+        type=str,
+        help="if specified, load prompts from this file, separated by newlines",
+    )
+    parser.add_argument(
+        "--config",
+        type=str,
+        default="configs/stable-diffusion/v2-inference.yaml",
+        help="path to config which constructs model",
+    )
+    parser.add_argument(
+        "--ckpt",
+        type=str,
+        help="path to checkpoint of model",
+    )
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="the seed (for reproducible sampling)",
+    )
+    parser.add_argument(
+        "--precision",
+        type=str,
+        help="evaluate at this precision",
+        choices=["full", "autocast"],
+        default="autocast"
+    )
+    parser.add_argument(
+        "--repeat",
+        type=int,
+        default=1,
+        help="repeat each prompt in file this often",
+    )
+    opt = parser.parse_args()
+    return opt
+
+
+def put_watermark(img, wm_encoder=None):
+    if wm_encoder is not None:
+        img = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
+        img = wm_encoder.encode(img, 'dwtDct')
+        img = Image.fromarray(img[:, :, ::-1])
+    return img
+
+
+def main(opt):
+    seed_everything(opt.seed)
+
+    config = OmegaConf.load(f"{opt.config}")
+    model = load_model_from_config(config, f"{opt.ckpt}")
+
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = model.to(device)
+
+    if opt.plms:
+        sampler = PLMSSampler(model)
+    elif opt.dpm:
+        sampler = DPMSolverSampler(model)
+    else:
+        sampler = DDIMSampler(model)
+
+    os.makedirs(opt.outdir, exist_ok=True)
+    outpath = opt.outdir
+
+    print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...")
+    wm = "SDV2"
+    wm_encoder = WatermarkEncoder()
+    wm_encoder.set_watermark('bytes', wm.encode('utf-8'))
+
+    batch_size = opt.n_samples
+    n_rows = opt.n_rows if opt.n_rows > 0 else batch_size
+    if not opt.from_file:
+        prompt = opt.prompt
+        assert prompt is not None
+        data = [batch_size * [prompt]]
+
+    else:
+        print(f"reading prompts from {opt.from_file}")
+        with open(opt.from_file, "r") as f:
+            data = f.read().splitlines()
+            data = [p for p in data for i in range(opt.repeat)]
+            data = list(chunk(data, batch_size))
+
+    sample_path = os.path.join(outpath, "samples")
+    os.makedirs(sample_path, exist_ok=True)
+    sample_count = 0
+    base_count = len(os.listdir(sample_path))
+    grid_count = len(os.listdir(outpath)) - 1
+
+    start_code = None
+    if opt.fixed_code:
+        start_code = torch.randn([opt.n_samples, opt.C, opt.H // opt.f, opt.W // opt.f], device=device)
+
+    precision_scope = autocast if opt.precision == "autocast" else nullcontext
+    with torch.no_grad(), \
+        precision_scope("cuda"), \
+        model.ema_scope():
+            all_samples = list()
+            for n in trange(opt.n_iter, desc="Sampling"):
+                for prompts in tqdm(data, desc="data"):
+                    uc = None
+                    if opt.scale != 1.0:
+                        uc = model.get_learned_conditioning(batch_size * [""])
+                    if isinstance(prompts, tuple):
+                        prompts = list(prompts)
+                    c = model.get_learned_conditioning(prompts)
+                    shape = [opt.C, opt.H // opt.f, opt.W // opt.f]
+                    samples, _ = sampler.sample(S=opt.steps,
+                                                     conditioning=c,
+                                                     batch_size=opt.n_samples,
+                                                     shape=shape,
+                                                     verbose=False,
+                                                     unconditional_guidance_scale=opt.scale,
+                                                     unconditional_conditioning=uc,
+                                                     eta=opt.ddim_eta,
+                                                     x_T=start_code)
+
+                    x_samples = model.decode_first_stage(samples)
+                    x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0)
+
+                    for x_sample in x_samples:
+                        x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
+                        img = Image.fromarray(x_sample.astype(np.uint8))
+                        img = put_watermark(img, wm_encoder)
+                        img.save(os.path.join(sample_path, f"{base_count:05}.png"))
+                        base_count += 1
+                        sample_count += 1
+
+                    all_samples.append(x_samples)
+
+            # additionally, save as grid
+            grid = torch.stack(all_samples, 0)
+            grid = rearrange(grid, 'n b c h w -> (n b) c h w')
+            grid = make_grid(grid, nrow=n_rows)
+
+            # to image
+            grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy()
+            grid = Image.fromarray(grid.astype(np.uint8))
+            grid = put_watermark(grid, wm_encoder)
+            grid.save(os.path.join(outpath, f'grid-{grid_count:04}.png'))
+            grid_count += 1
+
+    print(f"Your samples are ready and waiting for you here: \n{outpath} \n"
+          f" \nEnjoy.")
+
+
+if __name__ == "__main__":
+    opt = parse_args()
+    main(opt)

setup.py

@@ -0,0 +1,13 @@
+from setuptools import setup, find_packages
+
+setup(
+    name='stable-diffusion',
+    version='0.0.1',
+    description='',
+    packages=find_packages(),
+    install_requires=[
+        'torch',
+        'numpy',
+        'tqdm',
+    ],
+)