init

app.py

@@ -3,6 +3,7 @@ import sys
 
 depth_directory = os.path.abspath(os.path.join(os.path.dirname(__file__), 'depth'))
 sys.path.append(depth_directory)
+sys.path.append('./latent-diffusion')
 os.chdir(depth_directory)
 
 import cv2

stable-diffusion/LICENSE

@@ -0,0 +1,82 @@
+Copyright (c) 2022 Robin Rombach and Patrick Esser and contributors
+
+CreativeML Open RAIL-M
+dated August 22, 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, update the Model through electronic means, or modify the Output of the Model based on updates. You shall undertake reasonable efforts to use the latest version of the Model.
+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).

stable-diffusion/README.md

@@ -0,0 +1,215 @@
+# Stable Diffusion
+*Stable Diffusion was made possible thanks to a collaboration with [Stability AI](https://stability.ai/) and [Runway](https://runwayml.com/) and builds upon our previous 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/)_
+
+![txt2img-stable2](assets/stable-samples/txt2img/merged-0006.png)
+[Stable Diffusion](#stable-diffusion-v1) is a latent text-to-image diffusion
+model.
+Thanks to a generous compute donation from [Stability AI](https://stability.ai/) and support from [LAION](https://laion.ai/), we were able to train a Latent Diffusion Model on 512x512 images from a subset of the [LAION-5B](https://laion.ai/blog/laion-5b/) database. 
+Similar to Google's [Imagen](https://arxiv.org/abs/2205.11487), 
+this model uses a frozen CLIP ViT-L/14 text encoder to condition the model on text prompts.
+With its 860M UNet and 123M text encoder, the model is relatively lightweight and runs on a GPU with at least 10GB VRAM.
+See [this section](#stable-diffusion-v1) below and the [model card](https://huggingface.co/CompVis/stable-diffusion).
+
+  
+## Requirements
+A suitable [conda](https://conda.io/) environment named `ldm` can be created
+and activated with:
+
+```
+conda env create -f environment.yaml
+conda activate ldm
+```
+
+You can also update an existing [latent diffusion](https://github.com/CompVis/latent-diffusion) environment by running
+
+```
+conda install pytorch torchvision -c pytorch
+pip install transformers==4.19.2 diffusers invisible-watermark
+pip install -e .
+``` 
+
+
+## Stable Diffusion v1
+
+Stable Diffusion v1 refers to a specific configuration of the model
+architecture that uses a downsampling-factor 8 autoencoder with an 860M UNet
+and CLIP ViT-L/14 text encoder for the diffusion model. The model was pretrained on 256x256 images and 
+then finetuned on 512x512 images.
+
+*Note: Stable Diffusion v1 is a general text-to-image diffusion model and therefore mirrors biases and (mis-)conceptions that are present
+in its training data. 
+Details on the training procedure and data, as well as the intended use of the model can be found in the corresponding [model card](Stable_Diffusion_v1_Model_Card.md).*
+
+The weights are available via [the CompVis organization at Hugging Face](https://huggingface.co/CompVis) under [a license which contains specific use-based restrictions to prevent misuse and harm as informed by the model card, but otherwise remains permissive](LICENSE). While commercial use is permitted under the terms of the license, **we do not recommend using the provided weights for services or products without additional safety mechanisms and considerations**, since there are [known limitations and biases](Stable_Diffusion_v1_Model_Card.md#limitations-and-bias) of the weights, and research on safe and ethical deployment of general text-to-image models is an ongoing effort. **The weights are research artifacts and should be treated as such.**
+
+[The CreativeML OpenRAIL M license](LICENSE) is an [Open RAIL M license](https://www.licenses.ai/blog/2022/8/18/naming-convention-of-responsible-ai-licenses), adapted from the work that [BigScience](https://bigscience.huggingface.co/) and [the RAIL Initiative](https://www.licenses.ai/) are jointly carrying in the area of responsible AI licensing. See also [the article about the BLOOM Open RAIL license](https://bigscience.huggingface.co/blog/the-bigscience-rail-license) on which our license is based.
+
+### Weights
+
+We currently provide the following checkpoints:
+
+- `sd-v1-1.ckpt`: 237k steps at resolution `256x256` on [laion2B-en](https://huggingface.co/datasets/laion/laion2B-en).
+  194k steps at resolution `512x512` on [laion-high-resolution](https://huggingface.co/datasets/laion/laion-high-resolution) (170M examples from LAION-5B with resolution `>= 1024x1024`).
+- `sd-v1-2.ckpt`: Resumed from `sd-v1-1.ckpt`.
+  515k steps at resolution `512x512` on [laion-aesthetics v2 5+](https://laion.ai/blog/laion-aesthetics/) (a subset of laion2B-en with estimated aesthetics score `> 5.0`, and additionally
+filtered to images with an original size `>= 512x512`, and an estimated watermark probability `< 0.5`. The watermark estimate is from the [LAION-5B](https://laion.ai/blog/laion-5b/) metadata, the aesthetics score is estimated using the [LAION-Aesthetics Predictor V2](https://github.com/christophschuhmann/improved-aesthetic-predictor)).
+- `sd-v1-3.ckpt`: Resumed from `sd-v1-2.ckpt`. 195k steps at resolution `512x512` on "laion-aesthetics v2 5+" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
+- `sd-v1-4.ckpt`: Resumed from `sd-v1-2.ckpt`. 225k steps at resolution `512x512` on "laion-aesthetics v2 5+" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
+
+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 PLMS sampling
+steps show the relative improvements of the checkpoints:
+![sd evaluation results](assets/v1-variants-scores.jpg)
+
+
+
+### Text-to-Image with Stable Diffusion
+![txt2img-stable2](assets/stable-samples/txt2img/merged-0005.png)
+![txt2img-stable2](assets/stable-samples/txt2img/merged-0007.png)
+
+Stable Diffusion is a latent diffusion model conditioned on the (non-pooled) text embeddings of a CLIP ViT-L/14 text encoder.
+We provide a [reference script for sampling](#reference-sampling-script), but
+there also exists a [diffusers integration](#diffusers-integration), which we
+expect to see more active community development.
+
+#### Reference Sampling Script
+
+We provide a reference sampling script, which incorporates
+
+- a [Safety Checker Module](https://github.com/CompVis/stable-diffusion/pull/36),
+  to reduce the probability of explicit outputs,
+- 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).
+
+After [obtaining the `stable-diffusion-v1-*-original` weights](#weights), link them
+```
+mkdir -p models/ldm/stable-diffusion-v1/
+ln -s <path/to/model.ckpt> models/ldm/stable-diffusion-v1/model.ckpt 
+```
+and sample with
+```
+python scripts/txt2img.py --prompt "a photograph of an astronaut riding a horse" --plms 
+```
+
+By default, this uses a guidance scale of `--scale 7.5`, [Katherine Crowson's implementation](https://github.com/CompVis/latent-diffusion/pull/51) of the [PLMS](https://arxiv.org/abs/2202.09778) sampler, 
+and renders images of size 512x512 (which it was trained on) in 50 steps. All supported arguments are listed below (type `python scripts/txt2img.py --help`).
+
+
+```commandline
+usage: txt2img.py [-h] [--prompt [PROMPT]] [--outdir [OUTDIR]] [--skip_grid] [--skip_save] [--ddim_steps DDIM_STEPS] [--plms] [--laion400m] [--fixed_code] [--ddim_eta DDIM_ETA]
+                  [--n_iter N_ITER] [--H H] [--W W] [--C C] [--f F] [--n_samples N_SAMPLES] [--n_rows N_ROWS] [--scale SCALE] [--from-file FROM_FILE] [--config CONFIG] [--ckpt CKPT]
+                  [--seed SEED] [--precision {full,autocast}]
+
+optional arguments:
+  -h, --help            show this help message and exit
+  --prompt [PROMPT]     the prompt to render
+  --outdir [OUTDIR]     dir to write results to
+  --skip_grid           do not save a grid, only individual samples. Helpful when evaluating lots of samples
+  --skip_save           do not save individual samples. For speed measurements.
+  --ddim_steps DDIM_STEPS
+                        number of ddim sampling steps
+  --plms                use plms sampling
+  --laion400m           uses the LAION400M model
+  --fixed_code          if enabled, uses the same starting code across samples
+  --ddim_eta DDIM_ETA   ddim eta (eta=0.0 corresponds to deterministic sampling
+  --n_iter N_ITER       sample this often
+  --H H                 image height, in pixel space
+  --W W                 image width, in pixel space
+  --C C                 latent channels
+  --f F                 downsampling factor
+  --n_samples N_SAMPLES
+                        how many samples to produce for each given prompt. A.k.a. batch size
+  --n_rows N_ROWS       rows in the grid (default: n_samples)
+  --scale SCALE         unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))
+  --from-file FROM_FILE
+                        if specified, load prompts from this file
+  --config CONFIG       path to config which constructs model
+  --ckpt CKPT           path to checkpoint of model
+  --seed SEED           the seed (for reproducible sampling)
+  --precision {full,autocast}
+                        evaluate at this precision
+```
+Note: The inference config for all v1 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. If you want to examine the effect of EMA vs no EMA, we provide "full" checkpoints
+which contain both types of weights. For these, `use_ema=False` will load and use the non-EMA weights.
+
+
+#### Diffusers Integration
+
+A simple way to download and sample Stable Diffusion is by using the [diffusers library](https://github.com/huggingface/diffusers/tree/main#new--stable-diffusion-is-now-fully-compatible-with-diffusers):
+```py
+# make sure you're logged in with `huggingface-cli login`
+from torch import autocast
+from diffusers import StableDiffusionPipeline
+
+pipe = StableDiffusionPipeline.from_pretrained(
+	"CompVis/stable-diffusion-v1-4", 
+	use_auth_token=True
+).to("cuda")
+
+prompt = "a photo of an astronaut riding a horse on mars"
+with autocast("cuda"):
+    image = pipe(prompt)["sample"][0]  
+    
+image.save("astronaut_rides_horse.png")
+```
+
+
+### Image Modification with Stable Diffusion
+
+By using a diffusion-denoising mechanism as first proposed by [SDEdit](https://arxiv.org/abs/2108.01073), the model can be used for different 
+tasks such as text-guided image-to-image translation and upscaling. Similar to the txt2img sampling script, 
+we provide a script to perform image modification with Stable Diffusion.  
+
+The following describes an example where a rough sketch made in [Pinta](https://www.pinta-project.com/) is converted into a detailed artwork.
+```
+python scripts/img2img.py --prompt "A fantasy landscape, trending on artstation" --init-img <path-to-img.jpg> --strength 0.8
+```
+Here, strength is a value between 0.0 and 1.0, that controls the amount of noise that is added to the input image. 
+Values that approach 1.0 allow for lots of variations but will also produce images that are not semantically consistent with the input. See the following example.
+
+**Input**
+
+![sketch-in](assets/stable-samples/img2img/sketch-mountains-input.jpg)
+
+**Outputs**
+
+![out3](assets/stable-samples/img2img/mountains-3.png)
+![out2](assets/stable-samples/img2img/mountains-2.png)
+
+This procedure can, for example, also be used to upscale samples from the base model.
+
+
+## Comments 
+
+- 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!
+
+- The implementation of the transformer encoder is from [x-transformers](https://github.com/lucidrains/x-transformers) by [lucidrains](https://github.com/lucidrains?tab=repositories). 
+
+
+## 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}
+}
+```
+
+

stable-diffusion/Stable_Diffusion_v1_Model_Card.md

@@ -0,0 +1,144 @@
+# Stable Diffusion v1 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:** [Proprietary](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 ([CLIP ViT-L/14](https://arxiv.org/abs/2103.00020)) as suggested in the [Imagen paper](https://arxiv.org/abs/2205.11487).
+- **Resources for more information:** [GitHub Repository](https://github.com/CompVis/stable-diffusion), [Paper](https://arxiv.org/abs/2112.10752).
+- **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 taken from the [DALLE-MINI model card](https://huggingface.co/dalle-mini/dalle-mini), but applies in the same way to Stable Diffusion v1_.
+
+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 large-scale dataset
+  [LAION-5B](https://laion.ai/blog/laion-5b/) which contains adult material
+  and is not fit for product use without additional safety mechanisms and
+  considerations.
+- No additional measures were used to deduplicate the dataset. As a result, we observe some degree of memorization for images that are duplicated in the training data.
+  The training data can be searched at [https://rom1504.github.io/clip-retrieval/](https://rom1504.github.io/clip-retrieval/) to possibly assist in the detection of memorized images.
+
+### Bias
+While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases. 
+Stable Diffusion v1 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 v1 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 thereof (see next section)
+
+**Training Procedure**
+Stable Diffusion v1 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 a ViT-L/14 text-encoder.
+- The non-pooled 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 currently provide the following checkpoints:
+
+- `sd-v1-1.ckpt`: 237k steps at resolution `256x256` on [laion2B-en](https://huggingface.co/datasets/laion/laion2B-en).
+  194k steps at resolution `512x512` on [laion-high-resolution](https://huggingface.co/datasets/laion/laion-high-resolution) (170M examples from LAION-5B with resolution `>= 1024x1024`).
+- `sd-v1-2.ckpt`: Resumed from `sd-v1-1.ckpt`.
+  515k steps at resolution `512x512` on [laion-aesthetics v2 5+](https://laion.ai/blog/laion-aesthetics/) (a subset of laion2B-en with estimated aesthetics score `> 5.0`, and additionally
+filtered to images with an original size `>= 512x512`, and an estimated watermark probability `< 0.5`. The watermark estimate is from the [LAION-5B](https://laion.ai/blog/laion-5b/) metadata, the aesthetics score is estimated using the [LAION-Aesthetics Predictor V2](https://github.com/christophschuhmann/improved-aesthetic-predictor)).
+- `sd-v1-3.ckpt`: Resumed from `sd-v1-2.ckpt`. 195k steps at resolution `512x512` on "laion-aesthetics v2 5+" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
+- `sd-v1-4.ckpt`: Resumed from `sd-v1-2.ckpt`. 225k steps at resolution `512x512` on "laion-aesthetics v2 5+" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
+
+- **Hardware:** 32 x 8 x A100 GPUs
+- **Optimizer:** AdamW
+- **Gradient Accumulations**: 2
+- **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 PLMS sampling
+steps show the relative improvements of the checkpoints:
+
+![pareto](assets/v1-variants-scores.jpg) 
+
+Evaluated using 50 PLMS 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:** 150000
+- **Cloud Provider:** AWS
+- **Compute Region:** US-east
+- **Carbon Emitted (Power consumption x Time x Carbon produced based on location of power grid):** 11250 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 and Patrick Esser and is based on the [DALL-E Mini model card](https://huggingface.co/dalle-mini/dalle-mini).*

stable-diffusion/configs/autoencoder/autoencoder_kl_16x16x16.yaml

@@ -0,0 +1,54 @@
+model:
+  base_learning_rate: 4.5e-6
+  target: ldm.models.autoencoder.AutoencoderKL
+  params:
+    monitor: "val/rec_loss"
+    embed_dim: 16
+    lossconfig:
+      target: ldm.modules.losses.LPIPSWithDiscriminator
+      params:
+        disc_start: 50001
+        kl_weight: 0.000001
+        disc_weight: 0.5
+
+    ddconfig:
+      double_z: True
+      z_channels: 16
+      resolution: 256
+      in_channels: 3
+      out_ch: 3
+      ch: 128
+      ch_mult: [ 1,1,2,2,4]  # num_down = len(ch_mult)-1
+      num_res_blocks: 2
+      attn_resolutions: [16]
+      dropout: 0.0
+
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 12
+    wrap: True
+    train:
+      target: ldm.data.imagenet.ImageNetSRTrain
+      params:
+        size: 256
+        degradation: pil_nearest
+    validation:
+      target: ldm.data.imagenet.ImageNetSRValidation
+      params:
+        size: 256
+        degradation: pil_nearest
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 1000
+        max_images: 8
+        increase_log_steps: True
+
+  trainer:
+    benchmark: True
+    accumulate_grad_batches: 2

stable-diffusion/configs/autoencoder/autoencoder_kl_32x32x4.yaml

@@ -0,0 +1,53 @@
+model:
+  base_learning_rate: 4.5e-6
+  target: ldm.models.autoencoder.AutoencoderKL
+  params:
+    monitor: "val/rec_loss"
+    embed_dim: 4
+    lossconfig:
+      target: ldm.modules.losses.LPIPSWithDiscriminator
+      params:
+        disc_start: 50001
+        kl_weight: 0.000001
+        disc_weight: 0.5
+
+    ddconfig:
+      double_z: True
+      z_channels: 4
+      resolution: 256
+      in_channels: 3
+      out_ch: 3
+      ch: 128
+      ch_mult: [ 1,2,4,4 ]  # num_down = len(ch_mult)-1
+      num_res_blocks: 2
+      attn_resolutions: [ ]
+      dropout: 0.0
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 12
+    wrap: True
+    train:
+      target: ldm.data.imagenet.ImageNetSRTrain
+      params:
+        size: 256
+        degradation: pil_nearest
+    validation:
+      target: ldm.data.imagenet.ImageNetSRValidation
+      params:
+        size: 256
+        degradation: pil_nearest
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 1000
+        max_images: 8
+        increase_log_steps: True
+
+  trainer:
+    benchmark: True
+    accumulate_grad_batches: 2

stable-diffusion/configs/autoencoder/autoencoder_kl_64x64x3.yaml

@@ -0,0 +1,54 @@
+model:
+  base_learning_rate: 4.5e-6
+  target: ldm.models.autoencoder.AutoencoderKL
+  params:
+    monitor: "val/rec_loss"
+    embed_dim: 3
+    lossconfig:
+      target: ldm.modules.losses.LPIPSWithDiscriminator
+      params:
+        disc_start: 50001
+        kl_weight: 0.000001
+        disc_weight: 0.5
+
+    ddconfig:
+      double_z: True
+      z_channels: 3
+      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
+
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 12
+    wrap: True
+    train:
+      target: ldm.data.imagenet.ImageNetSRTrain
+      params:
+        size: 256
+        degradation: pil_nearest
+    validation:
+      target: ldm.data.imagenet.ImageNetSRValidation
+      params:
+        size: 256
+        degradation: pil_nearest
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 1000
+        max_images: 8
+        increase_log_steps: True
+
+  trainer:
+    benchmark: True
+    accumulate_grad_batches: 2

stable-diffusion/configs/autoencoder/autoencoder_kl_8x8x64.yaml

@@ -0,0 +1,53 @@
+model:
+  base_learning_rate: 4.5e-6
+  target: ldm.models.autoencoder.AutoencoderKL
+  params:
+    monitor: "val/rec_loss"
+    embed_dim: 64
+    lossconfig:
+      target: ldm.modules.losses.LPIPSWithDiscriminator
+      params:
+        disc_start: 50001
+        kl_weight: 0.000001
+        disc_weight: 0.5
+
+    ddconfig:
+      double_z: True
+      z_channels: 64
+      resolution: 256
+      in_channels: 3
+      out_ch: 3
+      ch: 128
+      ch_mult: [ 1,1,2,2,4,4]  # num_down = len(ch_mult)-1
+      num_res_blocks: 2
+      attn_resolutions: [16,8]
+      dropout: 0.0
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 12
+    wrap: True
+    train:
+      target: ldm.data.imagenet.ImageNetSRTrain
+      params:
+        size: 256
+        degradation: pil_nearest
+    validation:
+      target: ldm.data.imagenet.ImageNetSRValidation
+      params:
+        size: 256
+        degradation: pil_nearest
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 1000
+        max_images: 8
+        increase_log_steps: True
+
+  trainer:
+    benchmark: True
+    accumulate_grad_batches: 2

stable-diffusion/configs/latent-diffusion/celebahq-ldm-vq-4.yaml

@@ -0,0 +1,86 @@
+model:
+  base_learning_rate: 2.0e-06
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.0015
+    linear_end: 0.0195
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    image_size: 64
+    channels: 3
+    monitor: val/loss_simple_ema
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 64
+        in_channels: 3
+        out_channels: 3
+        model_channels: 224
+        attention_resolutions:
+        # note: this isn\t actually the resolution but
+        # the downsampling factor, i.e. this corresnponds to
+        # attention on spatial resolution 8,16,32, as the
+        # spatial reolution of the latents is 64 for f4
+        - 8
+        - 4
+        - 2
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 3
+        - 4
+        num_head_channels: 32
+    first_stage_config:
+      target: ldm.models.autoencoder.VQModelInterface
+      params:
+        embed_dim: 3
+        n_embed: 8192
+        ckpt_path: models/first_stage_models/vq-f4/model.ckpt
+        ddconfig:
+          double_z: false
+          z_channels: 3
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config: __is_unconditional__
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 48
+    num_workers: 5
+    wrap: false
+    train:
+      target: taming.data.faceshq.CelebAHQTrain
+      params:
+        size: 256
+    validation:
+      target: taming.data.faceshq.CelebAHQValidation
+      params:
+        size: 256
+
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 5000
+        max_images: 8
+        increase_log_steps: False
+
+  trainer:
+    benchmark: True

stable-diffusion/configs/latent-diffusion/cin-ldm-vq-f8.yaml

@@ -0,0 +1,98 @@
+model:
+  base_learning_rate: 1.0e-06
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.0015
+    linear_end: 0.0195
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    cond_stage_key: class_label
+    image_size: 32
+    channels: 4
+    cond_stage_trainable: true
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32
+        in_channels: 4
+        out_channels: 4
+        model_channels: 256
+        attention_resolutions:
+        #note: this isn\t actually the resolution but
+        # the downsampling factor, i.e. this corresnponds to
+        # attention on spatial resolution 8,16,32, as the
+        # spatial reolution of the latents is 32 for f8
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        num_head_channels: 32
+        use_spatial_transformer: true
+        transformer_depth: 1
+        context_dim: 512
+    first_stage_config:
+      target: ldm.models.autoencoder.VQModelInterface
+      params:
+        embed_dim: 4
+        n_embed: 16384
+        ckpt_path: configs/first_stage_models/vq-f8/model.yaml
+        ddconfig:
+          double_z: false
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 2
+          - 4
+          num_res_blocks: 2
+          attn_resolutions:
+          - 32
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.ClassEmbedder
+      params:
+        embed_dim: 512
+        key: class_label
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 64
+    num_workers: 12
+    wrap: false
+    train:
+      target: ldm.data.imagenet.ImageNetTrain
+      params:
+        config:
+          size: 256
+    validation:
+      target: ldm.data.imagenet.ImageNetValidation
+      params:
+        config:
+          size: 256
+
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 5000
+        max_images: 8
+        increase_log_steps: False
+
+  trainer:
+    benchmark: True

stable-diffusion/configs/latent-diffusion/cin256-v2.yaml

@@ -0,0 +1,68 @@
+model:
+  base_learning_rate: 0.0001
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.0015
+    linear_end: 0.0195
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    cond_stage_key: class_label
+    image_size: 64
+    channels: 3
+    cond_stage_trainable: true
+    conditioning_key: crossattn
+    monitor: val/loss
+    use_ema: False
+    
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 64
+        in_channels: 3
+        out_channels: 3
+        model_channels: 192
+        attention_resolutions:
+        - 8
+        - 4
+        - 2
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 3
+        - 5
+        num_heads: 1
+        use_spatial_transformer: true
+        transformer_depth: 1
+        context_dim: 512
+    
+    first_stage_config:
+      target: ldm.models.autoencoder.VQModelInterface
+      params:
+        embed_dim: 3
+        n_embed: 8192
+        ddconfig:
+          double_z: false
+          z_channels: 3
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.ClassEmbedder
+      params:
+        n_classes: 1001
+        embed_dim: 512
+        key: class_label

stable-diffusion/configs/latent-diffusion/ffhq-ldm-vq-4.yaml

@@ -0,0 +1,85 @@
+model:
+  base_learning_rate: 2.0e-06
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.0015
+    linear_end: 0.0195
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    image_size: 64
+    channels: 3
+    monitor: val/loss_simple_ema
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 64
+        in_channels: 3
+        out_channels: 3
+        model_channels: 224
+        attention_resolutions:
+        # note: this isn\t actually the resolution but
+        # the downsampling factor, i.e. this corresnponds to
+        # attention on spatial resolution 8,16,32, as the
+        # spatial reolution of the latents is 64 for f4
+        - 8
+        - 4
+        - 2
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 3
+        - 4
+        num_head_channels: 32
+    first_stage_config:
+      target: ldm.models.autoencoder.VQModelInterface
+      params:
+        embed_dim: 3
+        n_embed: 8192
+        ckpt_path: configs/first_stage_models/vq-f4/model.yaml
+        ddconfig:
+          double_z: false
+          z_channels: 3
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config: __is_unconditional__
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 42
+    num_workers: 5
+    wrap: false
+    train:
+      target: taming.data.faceshq.FFHQTrain
+      params:
+        size: 256
+    validation:
+      target: taming.data.faceshq.FFHQValidation
+      params:
+        size: 256
+
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 5000
+        max_images: 8
+        increase_log_steps: False
+
+  trainer:
+    benchmark: True

stable-diffusion/configs/latent-diffusion/lsun_bedrooms-ldm-vq-4.yaml

@@ -0,0 +1,85 @@
+model:
+  base_learning_rate: 2.0e-06
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.0015
+    linear_end: 0.0195
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    image_size: 64
+    channels: 3
+    monitor: val/loss_simple_ema
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 64
+        in_channels: 3
+        out_channels: 3
+        model_channels: 224
+        attention_resolutions:
+        # note: this isn\t actually the resolution but
+        # the downsampling factor, i.e. this corresnponds to
+        # attention on spatial resolution 8,16,32, as the
+        # spatial reolution of the latents is 64 for f4
+        - 8
+        - 4
+        - 2
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 3
+        - 4
+        num_head_channels: 32
+    first_stage_config:
+      target: ldm.models.autoencoder.VQModelInterface
+      params:
+        ckpt_path: configs/first_stage_models/vq-f4/model.yaml
+        embed_dim: 3
+        n_embed: 8192
+        ddconfig:
+          double_z: false
+          z_channels: 3
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config: __is_unconditional__
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 48
+    num_workers: 5
+    wrap: false
+    train:
+      target: ldm.data.lsun.LSUNBedroomsTrain
+      params:
+        size: 256
+    validation:
+      target: ldm.data.lsun.LSUNBedroomsValidation
+      params:
+        size: 256
+
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 5000
+        max_images: 8
+        increase_log_steps: False
+
+  trainer:
+    benchmark: True

stable-diffusion/configs/latent-diffusion/lsun_churches-ldm-kl-8.yaml

@@ -0,0 +1,91 @@
+model:
+  base_learning_rate: 5.0e-5   # set to target_lr by starting main.py with '--scale_lr False'
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.0015
+    linear_end: 0.0155
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    loss_type: l1
+    first_stage_key: "image"
+    cond_stage_key: "image"
+    image_size: 32
+    channels: 4
+    cond_stage_trainable: False
+    concat_mode: False
+    scale_by_std: True
+    monitor: 'val/loss_simple_ema'
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [10000]
+        cycle_lengths: [10000000000000]
+        f_start: [1.e-6]
+        f_max: [1.]
+        f_min: [ 1.]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32
+        in_channels: 4
+        out_channels: 4
+        model_channels: 192
+        attention_resolutions: [ 1, 2, 4, 8 ]   # 32, 16, 8, 4
+        num_res_blocks: 2
+        channel_mult: [ 1,2,2,4,4 ]  # 32, 16, 8, 4, 2
+        num_heads: 8
+        use_scale_shift_norm: True
+        resblock_updown: True
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: "val/rec_loss"
+        ckpt_path: "models/first_stage_models/kl-f8/model.ckpt"
+        ddconfig:
+          double_z: True
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [ 1,2,4,4 ]  # num_down = len(ch_mult)-1
+          num_res_blocks: 2
+          attn_resolutions: [ ]
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config: "__is_unconditional__"
+
+data:
+  target: main.DataModuleFromConfig
+  params:
+    batch_size: 96
+    num_workers: 5
+    wrap: False
+    train:
+      target: ldm.data.lsun.LSUNChurchesTrain
+      params:
+        size: 256
+    validation:
+      target: ldm.data.lsun.LSUNChurchesValidation
+      params:
+        size: 256
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 5000
+        max_images: 8
+        increase_log_steps: False
+
+
+  trainer:
+    benchmark: True

stable-diffusion/configs/latent-diffusion/txt2img-1p4B-eval.yaml

@@ -0,0 +1,71 @@
+model:
+  base_learning_rate: 5.0e-05
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    cond_stage_key: caption
+    image_size: 32
+    channels: 4
+    cond_stage_trainable: true
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32
+        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_heads: 8
+        use_spatial_transformer: true
+        transformer_depth: 1
+        context_dim: 1280
+        use_checkpoint: true
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          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.BERTEmbedder
+      params:
+        n_embed: 1280
+        n_layer: 32

stable-diffusion/configs/retrieval-augmented-diffusion/768x768.yaml

@@ -0,0 +1,68 @@
+model:
+  base_learning_rate: 0.0001
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.0015
+    linear_end: 0.015
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: jpg
+    cond_stage_key: nix
+    image_size: 48
+    channels: 16
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_by_std: false
+    scale_factor: 0.22765929
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 48
+        in_channels: 16
+        out_channels: 16
+        model_channels: 448
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 3
+        - 4
+        use_scale_shift_norm: false
+        resblock_updown: false
+        num_head_channels: 32
+        use_spatial_transformer: true
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: true
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        monitor: val/rec_loss
+        embed_dim: 16
+        ddconfig:
+          double_z: true
+          z_channels: 16
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 1
+          - 2
+          - 2
+          - 4
+          num_res_blocks: 2
+          attn_resolutions:
+          - 16
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config:
+      target: torch.nn.Identity

stable-diffusion/configs/stable-diffusion/v1-inference.yaml

@@ -0,0 +1,70 @@
+model:
+  base_learning_rate: 1.0e-04
+  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   # Note: different from the one we trained before
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 10000 ]
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1. ]
+        f_min: [ 1. ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        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_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: True
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          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.FrozenCLIPEmbedder

stable-diffusion/environment.yaml

@@ -0,0 +1,31 @@
+name: ldm
+channels:
+  - pytorch
+  - defaults
+dependencies:
+  - python=3.8.5
+  - pip=20.3
+  - cudatoolkit=11.3
+  - pytorch=1.11.0
+  - torchvision=0.12.0
+  - numpy=1.19.2
+  - pip:
+    - albumentations==0.4.3
+    - diffusers
+    - opencv-python==4.1.2.30
+    - pudb==2019.2
+    - invisible-watermark
+    - imageio==2.9.0
+    - imageio-ffmpeg==0.4.2
+    - pytorch-lightning==1.4.2
+    - omegaconf==2.1.1
+    - test-tube>=0.7.5
+    - streamlit>=0.73.1
+    - einops==0.3.0
+    - torch-fidelity==0.3.0
+    - transformers==4.19.2
+    - torchmetrics==0.6.0
+    - kornia==0.6
+    - -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
+    - -e git+https://github.com/openai/CLIP.git@main#egg=clip
+    - -e .

stable-diffusion/ldm/data/base.py

@@ -0,0 +1,23 @@
+from abc import abstractmethod
+from torch.utils.data import Dataset, ConcatDataset, ChainDataset, IterableDataset
+
+
+class Txt2ImgIterableBaseDataset(IterableDataset):
+    '''
+    Define an interface to make the IterableDatasets for text2img data chainable
+    '''
+    def __init__(self, num_records=0, valid_ids=None, size=256):
+        super().__init__()
+        self.num_records = num_records
+        self.valid_ids = valid_ids
+        self.sample_ids = valid_ids
+        self.size = size
+
+        print(f'{self.__class__.__name__} dataset contains {self.__len__()} examples.')
+
+    def __len__(self):
+        return self.num_records
+
+    @abstractmethod
+    def __iter__(self):
+        pass

stable-diffusion/ldm/data/imagenet.py

@@ -0,0 +1,394 @@
+import os, yaml, pickle, shutil, tarfile, glob
+import cv2
+import albumentations
+import PIL
+import numpy as np
+import torchvision.transforms.functional as TF
+from omegaconf import OmegaConf
+from functools import partial
+from PIL import Image
+from tqdm import tqdm
+from torch.utils.data import Dataset, Subset
+
+import taming.data.utils as tdu
+from taming.data.imagenet import str_to_indices, give_synsets_from_indices, download, retrieve
+from taming.data.imagenet import ImagePaths
+
+from ldm.modules.image_degradation import degradation_fn_bsr, degradation_fn_bsr_light
+
+
+def synset2idx(path_to_yaml="data/index_synset.yaml"):
+    with open(path_to_yaml) as f:
+        di2s = yaml.load(f)
+    return dict((v,k) for k,v in di2s.items())
+
+
+class ImageNetBase(Dataset):
+    def __init__(self, config=None):
+        self.config = config or OmegaConf.create()
+        if not type(self.config)==dict:
+            self.config = OmegaConf.to_container(self.config)
+        self.keep_orig_class_label = self.config.get("keep_orig_class_label", False)
+        self.process_images = True  # if False we skip loading & processing images and self.data contains filepaths
+        self._prepare()
+        self._prepare_synset_to_human()
+        self._prepare_idx_to_synset()
+        self._prepare_human_to_integer_label()
+        self._load()
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, i):
+        return self.data[i]
+
+    def _prepare(self):
+        raise NotImplementedError()
+
+    def _filter_relpaths(self, relpaths):
+        ignore = set([
+            "n06596364_9591.JPEG",
+        ])
+        relpaths = [rpath for rpath in relpaths if not rpath.split("/")[-1] in ignore]
+        if "sub_indices" in self.config:
+            indices = str_to_indices(self.config["sub_indices"])
+            synsets = give_synsets_from_indices(indices, path_to_yaml=self.idx2syn)  # returns a list of strings
+            self.synset2idx = synset2idx(path_to_yaml=self.idx2syn)
+            files = []
+            for rpath in relpaths:
+                syn = rpath.split("/")[0]
+                if syn in synsets:
+                    files.append(rpath)
+            return files
+        else:
+            return relpaths
+
+    def _prepare_synset_to_human(self):
+        SIZE = 2655750
+        URL = "https://heibox.uni-heidelberg.de/f/9f28e956cd304264bb82/?dl=1"
+        self.human_dict = os.path.join(self.root, "synset_human.txt")
+        if (not os.path.exists(self.human_dict) or
+                not os.path.getsize(self.human_dict)==SIZE):
+            download(URL, self.human_dict)
+
+    def _prepare_idx_to_synset(self):
+        URL = "https://heibox.uni-heidelberg.de/f/d835d5b6ceda4d3aa910/?dl=1"
+        self.idx2syn = os.path.join(self.root, "index_synset.yaml")
+        if (not os.path.exists(self.idx2syn)):
+            download(URL, self.idx2syn)
+
+    def _prepare_human_to_integer_label(self):
+        URL = "https://heibox.uni-heidelberg.de/f/2362b797d5be43b883f6/?dl=1"
+        self.human2integer = os.path.join(self.root, "imagenet1000_clsidx_to_labels.txt")
+        if (not os.path.exists(self.human2integer)):
+            download(URL, self.human2integer)
+        with open(self.human2integer, "r") as f:
+            lines = f.read().splitlines()
+            assert len(lines) == 1000
+            self.human2integer_dict = dict()
+            for line in lines:
+                value, key = line.split(":")
+                self.human2integer_dict[key] = int(value)
+
+    def _load(self):
+        with open(self.txt_filelist, "r") as f:
+            self.relpaths = f.read().splitlines()
+            l1 = len(self.relpaths)
+            self.relpaths = self._filter_relpaths(self.relpaths)
+            print("Removed {} files from filelist during filtering.".format(l1 - len(self.relpaths)))
+
+        self.synsets = [p.split("/")[0] for p in self.relpaths]
+        self.abspaths = [os.path.join(self.datadir, p) for p in self.relpaths]
+
+        unique_synsets = np.unique(self.synsets)
+        class_dict = dict((synset, i) for i, synset in enumerate(unique_synsets))
+        if not self.keep_orig_class_label:
+            self.class_labels = [class_dict[s] for s in self.synsets]
+        else:
+            self.class_labels = [self.synset2idx[s] for s in self.synsets]
+
+        with open(self.human_dict, "r") as f:
+            human_dict = f.read().splitlines()
+            human_dict = dict(line.split(maxsplit=1) for line in human_dict)
+
+        self.human_labels = [human_dict[s] for s in self.synsets]
+
+        labels = {
+            "relpath": np.array(self.relpaths),
+            "synsets": np.array(self.synsets),
+            "class_label": np.array(self.class_labels),
+            "human_label": np.array(self.human_labels),
+        }
+
+        if self.process_images:
+            self.size = retrieve(self.config, "size", default=256)
+            self.data = ImagePaths(self.abspaths,
+                                   labels=labels,
+                                   size=self.size,
+                                   random_crop=self.random_crop,
+                                   )
+        else:
+            self.data = self.abspaths
+
+
+class ImageNetTrain(ImageNetBase):
+    NAME = "ILSVRC2012_train"
+    URL = "http://www.image-net.org/challenges/LSVRC/2012/"
+    AT_HASH = "a306397ccf9c2ead27155983c254227c0fd938e2"
+    FILES = [
+        "ILSVRC2012_img_train.tar",
+    ]
+    SIZES = [
+        147897477120,
+    ]
+
+    def __init__(self, process_images=True, data_root=None, **kwargs):
+        self.process_images = process_images
+        self.data_root = data_root
+        super().__init__(**kwargs)
+
+    def _prepare(self):
+        if self.data_root:
+            self.root = os.path.join(self.data_root, self.NAME)
+        else:
+            cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
+            self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
+
+        self.datadir = os.path.join(self.root, "data")
+        self.txt_filelist = os.path.join(self.root, "filelist.txt")
+        self.expected_length = 1281167
+        self.random_crop = retrieve(self.config, "ImageNetTrain/random_crop",
+                                    default=True)
+        if not tdu.is_prepared(self.root):
+            # prep
+            print("Preparing dataset {} in {}".format(self.NAME, self.root))
+
+            datadir = self.datadir
+            if not os.path.exists(datadir):
+                path = os.path.join(self.root, self.FILES[0])
+                if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
+                    import academictorrents as at
+                    atpath = at.get(self.AT_HASH, datastore=self.root)
+                    assert atpath == path
+
+                print("Extracting {} to {}".format(path, datadir))
+                os.makedirs(datadir, exist_ok=True)
+                with tarfile.open(path, "r:") as tar:
+                    tar.extractall(path=datadir)
+
+                print("Extracting sub-tars.")
+                subpaths = sorted(glob.glob(os.path.join(datadir, "*.tar")))
+                for subpath in tqdm(subpaths):
+                    subdir = subpath[:-len(".tar")]
+                    os.makedirs(subdir, exist_ok=True)
+                    with tarfile.open(subpath, "r:") as tar:
+                        tar.extractall(path=subdir)
+
+            filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
+            filelist = [os.path.relpath(p, start=datadir) for p in filelist]
+            filelist = sorted(filelist)
+            filelist = "\n".join(filelist)+"\n"
+            with open(self.txt_filelist, "w") as f:
+                f.write(filelist)
+
+            tdu.mark_prepared(self.root)
+
+
+class ImageNetValidation(ImageNetBase):
+    NAME = "ILSVRC2012_validation"
+    URL = "http://www.image-net.org/challenges/LSVRC/2012/"
+    AT_HASH = "5d6d0df7ed81efd49ca99ea4737e0ae5e3a5f2e5"
+    VS_URL = "https://heibox.uni-heidelberg.de/f/3e0f6e9c624e45f2bd73/?dl=1"
+    FILES = [
+        "ILSVRC2012_img_val.tar",
+        "validation_synset.txt",
+    ]
+    SIZES = [
+        6744924160,
+        1950000,
+    ]
+
+    def __init__(self, process_images=True, data_root=None, **kwargs):
+        self.data_root = data_root
+        self.process_images = process_images
+        super().__init__(**kwargs)
+
+    def _prepare(self):
+        if self.data_root:
+            self.root = os.path.join(self.data_root, self.NAME)
+        else:
+            cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache"))
+            self.root = os.path.join(cachedir, "autoencoders/data", self.NAME)
+        self.datadir = os.path.join(self.root, "data")
+        self.txt_filelist = os.path.join(self.root, "filelist.txt")
+        self.expected_length = 50000
+        self.random_crop = retrieve(self.config, "ImageNetValidation/random_crop",
+                                    default=False)
+        if not tdu.is_prepared(self.root):
+            # prep
+            print("Preparing dataset {} in {}".format(self.NAME, self.root))
+
+            datadir = self.datadir
+            if not os.path.exists(datadir):
+                path = os.path.join(self.root, self.FILES[0])
+                if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]:
+                    import academictorrents as at
+                    atpath = at.get(self.AT_HASH, datastore=self.root)
+                    assert atpath == path
+
+                print("Extracting {} to {}".format(path, datadir))
+                os.makedirs(datadir, exist_ok=True)
+                with tarfile.open(path, "r:") as tar:
+                    tar.extractall(path=datadir)
+
+                vspath = os.path.join(self.root, self.FILES[1])
+                if not os.path.exists(vspath) or not os.path.getsize(vspath)==self.SIZES[1]:
+                    download(self.VS_URL, vspath)
+
+                with open(vspath, "r") as f:
+                    synset_dict = f.read().splitlines()
+                    synset_dict = dict(line.split() for line in synset_dict)
+
+                print("Reorganizing into synset folders")
+                synsets = np.unique(list(synset_dict.values()))
+                for s in synsets:
+                    os.makedirs(os.path.join(datadir, s), exist_ok=True)
+                for k, v in synset_dict.items():
+                    src = os.path.join(datadir, k)
+                    dst = os.path.join(datadir, v)
+                    shutil.move(src, dst)
+
+            filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG"))
+            filelist = [os.path.relpath(p, start=datadir) for p in filelist]
+            filelist = sorted(filelist)
+            filelist = "\n".join(filelist)+"\n"
+            with open(self.txt_filelist, "w") as f:
+                f.write(filelist)
+
+            tdu.mark_prepared(self.root)
+
+
+
+class ImageNetSR(Dataset):
+    def __init__(self, size=None,
+                 degradation=None, downscale_f=4, min_crop_f=0.5, max_crop_f=1.,
+                 random_crop=True):
+        """
+        Imagenet Superresolution Dataloader
+        Performs following ops in order:
+        1.  crops a crop of size s from image either as random or center crop
+        2.  resizes crop to size with cv2.area_interpolation
+        3.  degrades resized crop with degradation_fn
+
+        :param size: resizing to size after cropping
+        :param degradation: degradation_fn, e.g. cv_bicubic or bsrgan_light
+        :param downscale_f: Low Resolution Downsample factor
+        :param min_crop_f: determines crop size s,
+          where s = c * min_img_side_len with c sampled from interval (min_crop_f, max_crop_f)
+        :param max_crop_f: ""
+        :param data_root:
+        :param random_crop:
+        """
+        self.base = self.get_base()
+        assert size
+        assert (size / downscale_f).is_integer()
+        self.size = size
+        self.LR_size = int(size / downscale_f)
+        self.min_crop_f = min_crop_f
+        self.max_crop_f = max_crop_f
+        assert(max_crop_f <= 1.)
+        self.center_crop = not random_crop
+
+        self.image_rescaler = albumentations.SmallestMaxSize(max_size=size, interpolation=cv2.INTER_AREA)
+
+        self.pil_interpolation = False # gets reset later if incase interp_op is from pillow
+
+        if degradation == "bsrgan":
+            self.degradation_process = partial(degradation_fn_bsr, sf=downscale_f)
+
+        elif degradation == "bsrgan_light":
+            self.degradation_process = partial(degradation_fn_bsr_light, sf=downscale_f)
+
+        else:
+            interpolation_fn = {
+            "cv_nearest": cv2.INTER_NEAREST,
+            "cv_bilinear": cv2.INTER_LINEAR,
+            "cv_bicubic": cv2.INTER_CUBIC,
+            "cv_area": cv2.INTER_AREA,
+            "cv_lanczos": cv2.INTER_LANCZOS4,
+            "pil_nearest": PIL.Image.NEAREST,
+            "pil_bilinear": PIL.Image.BILINEAR,
+            "pil_bicubic": PIL.Image.BICUBIC,
+            "pil_box": PIL.Image.BOX,
+            "pil_hamming": PIL.Image.HAMMING,
+            "pil_lanczos": PIL.Image.LANCZOS,
+            }[degradation]
+
+            self.pil_interpolation = degradation.startswith("pil_")
+
+            if self.pil_interpolation:
+                self.degradation_process = partial(TF.resize, size=self.LR_size, interpolation=interpolation_fn)
+
+            else:
+                self.degradation_process = albumentations.SmallestMaxSize(max_size=self.LR_size,
+                                                                          interpolation=interpolation_fn)
+
+    def __len__(self):
+        return len(self.base)
+
+    def __getitem__(self, i):
+        example = self.base[i]
+        image = Image.open(example["file_path_"])
+
+        if not image.mode == "RGB":
+            image = image.convert("RGB")
+
+        image = np.array(image).astype(np.uint8)
+
+        min_side_len = min(image.shape[:2])
+        crop_side_len = min_side_len * np.random.uniform(self.min_crop_f, self.max_crop_f, size=None)
+        crop_side_len = int(crop_side_len)
+
+        if self.center_crop:
+            self.cropper = albumentations.CenterCrop(height=crop_side_len, width=crop_side_len)
+
+        else:
+            self.cropper = albumentations.RandomCrop(height=crop_side_len, width=crop_side_len)
+
+        image = self.cropper(image=image)["image"]
+        image = self.image_rescaler(image=image)["image"]
+
+        if self.pil_interpolation:
+            image_pil = PIL.Image.fromarray(image)
+            LR_image = self.degradation_process(image_pil)
+            LR_image = np.array(LR_image).astype(np.uint8)
+
+        else:
+            LR_image = self.degradation_process(image=image)["image"]
+
+        example["image"] = (image/127.5 - 1.0).astype(np.float32)
+        example["LR_image"] = (LR_image/127.5 - 1.0).astype(np.float32)
+
+        return example
+
+
+class ImageNetSRTrain(ImageNetSR):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def get_base(self):
+        with open("data/imagenet_train_hr_indices.p", "rb") as f:
+            indices = pickle.load(f)
+        dset = ImageNetTrain(process_images=False,)
+        return Subset(dset, indices)
+
+
+class ImageNetSRValidation(ImageNetSR):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def get_base(self):
+        with open("data/imagenet_val_hr_indices.p", "rb") as f:
+            indices = pickle.load(f)
+        dset = ImageNetValidation(process_images=False,)
+        return Subset(dset, indices)

stable-diffusion/ldm/data/lsun.py

@@ -0,0 +1,92 @@
+import os
+import numpy as np
+import PIL
+from PIL import Image
+from torch.utils.data import Dataset
+from torchvision import transforms
+
+
+class LSUNBase(Dataset):
+    def __init__(self,
+                 txt_file,
+                 data_root,
+                 size=None,
+                 interpolation="bicubic",
+                 flip_p=0.5
+                 ):
+        self.data_paths = txt_file
+        self.data_root = data_root
+        with open(self.data_paths, "r") as f:
+            self.image_paths = f.read().splitlines()
+        self._length = len(self.image_paths)
+        self.labels = {
+            "relative_file_path_": [l for l in self.image_paths],
+            "file_path_": [os.path.join(self.data_root, l)
+                           for l in self.image_paths],
+        }
+
+        self.size = size
+        self.interpolation = {"linear": PIL.Image.LINEAR,
+                              "bilinear": PIL.Image.BILINEAR,
+                              "bicubic": PIL.Image.BICUBIC,
+                              "lanczos": PIL.Image.LANCZOS,
+                              }[interpolation]
+        self.flip = transforms.RandomHorizontalFlip(p=flip_p)
+
+    def __len__(self):
+        return self._length
+
+    def __getitem__(self, i):
+        example = dict((k, self.labels[k][i]) for k in self.labels)
+        image = Image.open(example["file_path_"])
+        if not image.mode == "RGB":
+            image = image.convert("RGB")
+
+        # default to score-sde preprocessing
+        img = np.array(image).astype(np.uint8)
+        crop = min(img.shape[0], img.shape[1])
+        h, w, = img.shape[0], img.shape[1]
+        img = img[(h - crop) // 2:(h + crop) // 2,
+              (w - crop) // 2:(w + crop) // 2]
+
+        image = Image.fromarray(img)
+        if self.size is not None:
+            image = image.resize((self.size, self.size), resample=self.interpolation)
+
+        image = self.flip(image)
+        image = np.array(image).astype(np.uint8)
+        example["image"] = (image / 127.5 - 1.0).astype(np.float32)
+        return example
+
+
+class LSUNChurchesTrain(LSUNBase):
+    def __init__(self, **kwargs):
+        super().__init__(txt_file="data/lsun/church_outdoor_train.txt", data_root="data/lsun/churches", **kwargs)
+
+
+class LSUNChurchesValidation(LSUNBase):
+    def __init__(self, flip_p=0., **kwargs):
+        super().__init__(txt_file="data/lsun/church_outdoor_val.txt", data_root="data/lsun/churches",
+                         flip_p=flip_p, **kwargs)
+
+
+class LSUNBedroomsTrain(LSUNBase):
+    def __init__(self, **kwargs):
+        super().__init__(txt_file="data/lsun/bedrooms_train.txt", data_root="data/lsun/bedrooms", **kwargs)
+
+
+class LSUNBedroomsValidation(LSUNBase):
+    def __init__(self, flip_p=0.0, **kwargs):
+        super().__init__(txt_file="data/lsun/bedrooms_val.txt", data_root="data/lsun/bedrooms",
+                         flip_p=flip_p, **kwargs)
+
+
+class LSUNCatsTrain(LSUNBase):
+    def __init__(self, **kwargs):
+        super().__init__(txt_file="data/lsun/cat_train.txt", data_root="data/lsun/cats", **kwargs)
+
+
+class LSUNCatsValidation(LSUNBase):
+    def __init__(self, flip_p=0., **kwargs):
+        super().__init__(txt_file="data/lsun/cat_val.txt", data_root="data/lsun/cats",
+                         flip_p=flip_p, **kwargs)

stable-diffusion/ldm/lr_scheduler.py

@@ -0,0 +1,98 @@
+import numpy as np
+
+
+class LambdaWarmUpCosineScheduler:
+    """
+    note: use with a base_lr of 1.0
+    """
+    def __init__(self, warm_up_steps, lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0):
+        self.lr_warm_up_steps = warm_up_steps
+        self.lr_start = lr_start
+        self.lr_min = lr_min
+        self.lr_max = lr_max
+        self.lr_max_decay_steps = max_decay_steps
+        self.last_lr = 0.
+        self.verbosity_interval = verbosity_interval
+
+    def schedule(self, n, **kwargs):
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_lr}")
+        if n < self.lr_warm_up_steps:
+            lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start
+            self.last_lr = lr
+            return lr
+        else:
+            t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps)
+            t = min(t, 1.0)
+            lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * (
+                    1 + np.cos(t * np.pi))
+            self.last_lr = lr
+            return lr
+
+    def __call__(self, n, **kwargs):
+        return self.schedule(n,**kwargs)
+
+
+class LambdaWarmUpCosineScheduler2:
+    """
+    supports repeated iterations, configurable via lists
+    note: use with a base_lr of 1.0.
+    """
+    def __init__(self, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0):
+        assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths)
+        self.lr_warm_up_steps = warm_up_steps
+        self.f_start = f_start
+        self.f_min = f_min
+        self.f_max = f_max
+        self.cycle_lengths = cycle_lengths
+        self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths))
+        self.last_f = 0.
+        self.verbosity_interval = verbosity_interval
+
+    def find_in_interval(self, n):
+        interval = 0
+        for cl in self.cum_cycles[1:]:
+            if n <= cl:
+                return interval
+            interval += 1
+
+    def schedule(self, n, **kwargs):
+        cycle = self.find_in_interval(n)
+        n = n - self.cum_cycles[cycle]
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
+                                                       f"current cycle {cycle}")
+        if n < self.lr_warm_up_steps[cycle]:
+            f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
+            self.last_f = f
+            return f
+        else:
+            t = (n - self.lr_warm_up_steps[cycle]) / (self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle])
+            t = min(t, 1.0)
+            f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * (
+                    1 + np.cos(t * np.pi))
+            self.last_f = f
+            return f
+
+    def __call__(self, n, **kwargs):
+        return self.schedule(n, **kwargs)
+
+
+class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2):
+
+    def schedule(self, n, **kwargs):
+        cycle = self.find_in_interval(n)
+        n = n - self.cum_cycles[cycle]
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
+                                                       f"current cycle {cycle}")
+
+        if n < self.lr_warm_up_steps[cycle]:
+            f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
+            self.last_f = f
+            return f
+        else:
+            f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle])
+            self.last_f = f
+            return f
+

stable-diffusion/ldm/models/autoencoder.py

@@ -0,0 +1,443 @@
+import torch
+import pytorch_lightning as pl
+import torch.nn.functional as F
+from contextlib import contextmanager
+
+from taming.modules.vqvae.quantize import VectorQuantizer2 as VectorQuantizer
+
+from ldm.modules.diffusionmodules.model import Encoder, Decoder
+from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
+from ldm.util import instantiate_from_config
+
+
+class VQModel(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 n_embed,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 batch_resize_range=None,
+                 scheduler_config=None,
+                 lr_g_factor=1.0,
+                 remap=None,
+                 sane_index_shape=False, # tell vector quantizer to return indices as bhw
+                 use_ema=False
+                 ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.n_embed = n_embed
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
+                                        remap=remap,
+                                        sane_index_shape=sane_index_shape)
+        self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        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.batch_resize_range = batch_resize_range
+        if self.batch_resize_range is not None:
+            print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
+
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self)
+            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)
+        self.scheduler_config = scheduler_config
+        self.lr_g_factor = lr_g_factor
+
+    @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 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]
+        missing, unexpected = self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self)
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        quant, emb_loss, info = self.quantize(h)
+        return quant, emb_loss, info
+
+    def encode_to_prequant(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, quant):
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+    def decode_code(self, code_b):
+        quant_b = self.quantize.embed_code(code_b)
+        dec = self.decode(quant_b)
+        return dec
+
+    def forward(self, input, return_pred_indices=False):
+        quant, diff, (_,_,ind) = self.encode(input)
+        dec = self.decode(quant)
+        if return_pred_indices:
+            return dec, diff, ind
+        return dec, diff
+
+    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()
+        if self.batch_resize_range is not None:
+            lower_size = self.batch_resize_range[0]
+            upper_size = self.batch_resize_range[1]
+            if self.global_step <= 4:
+                # do the first few batches with max size to avoid later oom
+                new_resize = upper_size
+            else:
+                new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
+            if new_resize != x.shape[2]:
+                x = F.interpolate(x, size=new_resize, mode="bicubic")
+            x = x.detach()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        # https://github.com/pytorch/pytorch/issues/37142
+        # try not to fool the heuristics
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+
+        if optimizer_idx == 0:
+            # autoencode
+            aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train",
+                                            predicted_indices=ind)
+
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # discriminator
+            discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            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, suffix="_ema")
+        return log_dict
+
+    def _validation_step(self, batch, batch_idx, suffix=""):
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+        aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
+                                        self.global_step,
+                                        last_layer=self.get_last_layer(),
+                                        split="val"+suffix,
+                                        predicted_indices=ind
+                                        )
+
+        discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
+                                            self.global_step,
+                                            last_layer=self.get_last_layer(),
+                                            split="val"+suffix,
+                                            predicted_indices=ind
+                                            )
+        rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log(f"val{suffix}/rec_loss", rec_loss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        self.log(f"val{suffix}/aeloss", aeloss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        if version.parse(pl.__version__) >= version.parse('1.4.0'):
+            del log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def configure_optimizers(self):
+        lr_d = self.learning_rate
+        lr_g = self.lr_g_factor*self.learning_rate
+        print("lr_d", lr_d)
+        print("lr_g", lr_g)
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quantize.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr_g, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr_d, betas=(0.5, 0.9))
+
+        if self.scheduler_config is not None:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+                {
+                    'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+            ]
+            return [opt_ae, opt_disc], scheduler
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if only_inputs:
+            log["inputs"] = x
+            return log
+        xrec, _ = 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["inputs"] = x
+        log["reconstructions"] = xrec
+        if plot_ema:
+            with self.ema_scope():
+                xrec_ema, _ = self(x)
+                if x.shape[1] > 3: xrec_ema = self.to_rgb(xrec_ema)
+                log["reconstructions_ema"] = xrec_ema
+        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 VQModelInterface(VQModel):
+    def __init__(self, embed_dim, *args, **kwargs):
+        super().__init__(embed_dim=embed_dim, *args, **kwargs)
+        self.embed_dim = embed_dim
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, h, force_not_quantize=False):
+        # also go through quantization layer
+        if not force_not_quantize:
+            quant, emb_loss, info = self.quantize(h)
+        else:
+            quant = h
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+
+class AutoencoderKL(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 ):
+        super().__init__()
+        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
+        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}")
+
+    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):
+        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")
+
+        discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
+                                            last_layer=self.get_last_layer(), split="val")
+
+        self.log("val/rec_loss", log_dict_ae["val/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
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  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, **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
+        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  # TODO: Should be true by default but check to not break older stuff
+        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

stable-diffusion/ldm/models/diffusion/classifier.py

@@ -0,0 +1,267 @@
+import os
+import torch
+import pytorch_lightning as pl
+from omegaconf import OmegaConf
+from torch.nn import functional as F
+from torch.optim import AdamW
+from torch.optim.lr_scheduler import LambdaLR
+from copy import deepcopy
+from einops import rearrange
+from glob import glob
+from natsort import natsorted
+
+from ldm.modules.diffusionmodules.openaimodel import EncoderUNetModel, UNetModel
+from ldm.util import log_txt_as_img, default, ismap, instantiate_from_config
+
+__models__ = {
+    'class_label': EncoderUNetModel,
+    'segmentation': UNetModel
+}
+
+
+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 NoisyLatentImageClassifier(pl.LightningModule):
+
+    def __init__(self,
+                 diffusion_path,
+                 num_classes,
+                 ckpt_path=None,
+                 pool='attention',
+                 label_key=None,
+                 diffusion_ckpt_path=None,
+                 scheduler_config=None,
+                 weight_decay=1.e-2,
+                 log_steps=10,
+                 monitor='val/loss',
+                 *args,
+                 **kwargs):
+        super().__init__(*args, **kwargs)
+        self.num_classes = num_classes
+        # get latest config of diffusion model
+        diffusion_config = natsorted(glob(os.path.join(diffusion_path, 'configs', '*-project.yaml')))[-1]
+        self.diffusion_config = OmegaConf.load(diffusion_config).model
+        self.diffusion_config.params.ckpt_path = diffusion_ckpt_path
+        self.load_diffusion()
+
+        self.monitor = monitor
+        self.numd = self.diffusion_model.first_stage_model.encoder.num_resolutions - 1
+        self.log_time_interval = self.diffusion_model.num_timesteps // log_steps
+        self.log_steps = log_steps
+
+        self.label_key = label_key if not hasattr(self.diffusion_model, 'cond_stage_key') \
+            else self.diffusion_model.cond_stage_key
+
+        assert self.label_key is not None, 'label_key neither in diffusion model nor in model.params'
+
+        if self.label_key not in __models__:
+            raise NotImplementedError()
+
+        self.load_classifier(ckpt_path, pool)
+
+        self.scheduler_config = scheduler_config
+        self.use_scheduler = self.scheduler_config is not None
+        self.weight_decay = weight_decay
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["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]
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    def load_diffusion(self):
+        model = instantiate_from_config(self.diffusion_config)
+        self.diffusion_model = model.eval()
+        self.diffusion_model.train = disabled_train
+        for param in self.diffusion_model.parameters():
+            param.requires_grad = False
+
+    def load_classifier(self, ckpt_path, pool):
+        model_config = deepcopy(self.diffusion_config.params.unet_config.params)
+        model_config.in_channels = self.diffusion_config.params.unet_config.params.out_channels
+        model_config.out_channels = self.num_classes
+        if self.label_key == 'class_label':
+            model_config.pool = pool
+
+        self.model = __models__[self.label_key](**model_config)
+        if ckpt_path is not None:
+            print('#####################################################################')
+            print(f'load from ckpt "{ckpt_path}"')
+            print('#####################################################################')
+            self.init_from_ckpt(ckpt_path)
+
+    @torch.no_grad()
+    def get_x_noisy(self, x, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x))
+        continuous_sqrt_alpha_cumprod = None
+        if self.diffusion_model.use_continuous_noise:
+            continuous_sqrt_alpha_cumprod = self.diffusion_model.sample_continuous_noise_level(x.shape[0], t + 1)
+            # todo: make sure t+1 is correct here
+
+        return self.diffusion_model.q_sample(x_start=x, t=t, noise=noise,
+                                             continuous_sqrt_alpha_cumprod=continuous_sqrt_alpha_cumprod)
+
+    def forward(self, x_noisy, t, *args, **kwargs):
+        return self.model(x_noisy, t)
+
+    @torch.no_grad()
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = rearrange(x, 'b h w c -> b c h w')
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    @torch.no_grad()
+    def get_conditioning(self, batch, k=None):
+        if k is None:
+            k = self.label_key
+        assert k is not None, 'Needs to provide label key'
+
+        targets = batch[k].to(self.device)
+
+        if self.label_key == 'segmentation':
+            targets = rearrange(targets, 'b h w c -> b c h w')
+            for down in range(self.numd):
+                h, w = targets.shape[-2:]
+                targets = F.interpolate(targets, size=(h // 2, w // 2), mode='nearest')
+
+            # targets = rearrange(targets,'b c h w -> b h w c')
+
+        return targets
+
+    def compute_top_k(self, logits, labels, k, reduction="mean"):
+        _, top_ks = torch.topk(logits, k, dim=1)
+        if reduction == "mean":
+            return (top_ks == labels[:, None]).float().sum(dim=-1).mean().item()
+        elif reduction == "none":
+            return (top_ks == labels[:, None]).float().sum(dim=-1)
+
+    def on_train_epoch_start(self):
+        # save some memory
+        self.diffusion_model.model.to('cpu')
+
+    @torch.no_grad()
+    def write_logs(self, loss, logits, targets):
+        log_prefix = 'train' if self.training else 'val'
+        log = {}
+        log[f"{log_prefix}/loss"] = loss.mean()
+        log[f"{log_prefix}/acc@1"] = self.compute_top_k(
+            logits, targets, k=1, reduction="mean"
+        )
+        log[f"{log_prefix}/acc@5"] = self.compute_top_k(
+            logits, targets, k=5, reduction="mean"
+        )
+
+        self.log_dict(log, prog_bar=False, logger=True, on_step=self.training, on_epoch=True)
+        self.log('loss', log[f"{log_prefix}/loss"], prog_bar=True, logger=False)
+        self.log('global_step', self.global_step, logger=False, on_epoch=False, prog_bar=True)
+        lr = self.optimizers().param_groups[0]['lr']
+        self.log('lr_abs', lr, on_step=True, logger=True, on_epoch=False, prog_bar=True)
+
+    def shared_step(self, batch, t=None):
+        x, *_ = self.diffusion_model.get_input(batch, k=self.diffusion_model.first_stage_key)
+        targets = self.get_conditioning(batch)
+        if targets.dim() == 4:
+            targets = targets.argmax(dim=1)
+        if t is None:
+            t = torch.randint(0, self.diffusion_model.num_timesteps, (x.shape[0],), device=self.device).long()
+        else:
+            t = torch.full(size=(x.shape[0],), fill_value=t, device=self.device).long()
+        x_noisy = self.get_x_noisy(x, t)
+        logits = self(x_noisy, t)
+
+        loss = F.cross_entropy(logits, targets, reduction='none')
+
+        self.write_logs(loss.detach(), logits.detach(), targets.detach())
+
+        loss = loss.mean()
+        return loss, logits, x_noisy, targets
+
+    def training_step(self, batch, batch_idx):
+        loss, *_ = self.shared_step(batch)
+        return loss
+
+    def reset_noise_accs(self):
+        self.noisy_acc = {t: {'acc@1': [], 'acc@5': []} for t in
+                          range(0, self.diffusion_model.num_timesteps, self.diffusion_model.log_every_t)}
+
+    def on_validation_start(self):
+        self.reset_noise_accs()
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        loss, *_ = self.shared_step(batch)
+
+        for t in self.noisy_acc:
+            _, logits, _, targets = self.shared_step(batch, t)
+            self.noisy_acc[t]['acc@1'].append(self.compute_top_k(logits, targets, k=1, reduction='mean'))
+            self.noisy_acc[t]['acc@5'].append(self.compute_top_k(logits, targets, k=5, reduction='mean'))
+
+        return loss
+
+    def configure_optimizers(self):
+        optimizer = AdamW(self.model.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay)
+
+        if self.use_scheduler:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [optimizer], scheduler
+
+        return optimizer
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, *args, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.diffusion_model.first_stage_key)
+        log['inputs'] = x
+
+        y = self.get_conditioning(batch)
+
+        if self.label_key == 'class_label':
+            y = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+            log['labels'] = y
+
+        if ismap(y):
+            log['labels'] = self.diffusion_model.to_rgb(y)
+
+            for step in range(self.log_steps):
+                current_time = step * self.log_time_interval
+
+                _, logits, x_noisy, _ = self.shared_step(batch, t=current_time)
+
+                log[f'inputs@t{current_time}'] = x_noisy
+
+                pred = F.one_hot(logits.argmax(dim=1), num_classes=self.num_classes)
+                pred = rearrange(pred, 'b h w c -> b c h w')
+
+                log[f'pred@t{current_time}'] = self.diffusion_model.to_rgb(pred)
+
+        for key in log:
+            log[key] = log[key][:N]
+
+        return log

stable-diffusion/ldm/models/diffusion/ddim.py

@@ -0,0 +1,241 @@
+"""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, \
+    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, ...
+               **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 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,
+                                                    )
+        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,):
+        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
+
+            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)
+            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):
+        b, *_, device = *x.shape, x.device
+
+        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)
+
+        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
+        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)
+        # 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 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):
+
+        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)
+        return x_dec

stable-diffusion/ldm/models/diffusion/ddpm.py

@@ -0,0 +1,1445 @@
+"""
+wild mixture of
+https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
+https://github.com/CompVis/taming-transformers
+-- merci
+"""
+
+import torch
+import torch.nn as nn
+import numpy as np
+import pytorch_lightning as pl
+from torch.optim.lr_scheduler import LambdaLR
+from einops import rearrange, repeat
+from contextlib import contextmanager
+from functools import partial
+from tqdm import tqdm
+from torchvision.utils import make_grid
+from pytorch_lightning.utilities.distributed import rank_zero_only
+
+from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
+from ldm.modules.ema import LitEma
+from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
+from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
+from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
+from ldm.models.diffusion.ddim import DDIMSampler
+
+
+__conditioning_keys__ = {'concat': 'c_concat',
+                         'crossattn': 'c_crossattn',
+                         'adm': 'y'}
+
+
+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 uniform_on_device(r1, r2, shape, device):
+    return (r1 - r2) * torch.rand(*shape, device=device) + r2
+
+
+class DDPM(pl.LightningModule):
+    # classic DDPM with Gaussian diffusion, in image space
+    def __init__(self,
+                 unet_config,
+                 timesteps=1000,
+                 beta_schedule="linear",
+                 loss_type="l2",
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 load_only_unet=False,
+                 monitor="val/loss",
+                 use_ema=True,
+                 first_stage_key="image",
+                 image_size=256,
+                 channels=3,
+                 log_every_t=100,
+                 clip_denoised=True,
+                 linear_start=1e-4,
+                 linear_end=2e-2,
+                 cosine_s=8e-3,
+                 given_betas=None,
+                 original_elbo_weight=0.,
+                 v_posterior=0.,  # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
+                 l_simple_weight=1.,
+                 conditioning_key=None,
+                 parameterization="eps",  # all assuming fixed variance schedules
+                 scheduler_config=None,
+                 use_positional_encodings=False,
+                 learn_logvar=False,
+                 logvar_init=0.,
+                 ):
+        super().__init__()
+        assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
+        self.parameterization = parameterization
+        print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
+        self.cond_stage_model = None
+        self.clip_denoised = clip_denoised
+        self.log_every_t = log_every_t
+        self.first_stage_key = first_stage_key
+        self.image_size = image_size  # try conv?
+        self.channels = channels
+        self.use_positional_encodings = use_positional_encodings
+        self.model = DiffusionWrapper(unet_config, conditioning_key)
+        count_params(self.model, verbose=True)
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self.model)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        self.use_scheduler = scheduler_config is not None
+        if self.use_scheduler:
+            self.scheduler_config = scheduler_config
+
+        self.v_posterior = v_posterior
+        self.original_elbo_weight = original_elbo_weight
+        self.l_simple_weight = l_simple_weight
+
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
+
+        self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
+                               linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
+
+        self.loss_type = loss_type
+
+        self.learn_logvar = learn_logvar
+        self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
+        if self.learn_logvar:
+            self.logvar = nn.Parameter(self.logvar, requires_grad=True)
+
+
+    def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        if exists(given_betas):
+            betas = given_betas
+        else:
+            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)))
+
+        # calculations for posterior q(x_{t-1} | x_t, x_0)
+        posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
+                    1. - alphas_cumprod) + self.v_posterior * betas
+        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
+        self.register_buffer('posterior_variance', to_torch(posterior_variance))
+        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
+        self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
+        self.register_buffer('posterior_mean_coef1', to_torch(
+            betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
+        self.register_buffer('posterior_mean_coef2', to_torch(
+            (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
+
+        if self.parameterization == "eps":
+            lvlb_weights = self.betas ** 2 / (
+                        2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
+        elif self.parameterization == "x0":
+            lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
+        else:
+            raise NotImplementedError("mu not supported")
+        # TODO how to choose this term
+        lvlb_weights[0] = lvlb_weights[1]
+        self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
+        assert not torch.isnan(self.lvlb_weights).all()
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.model.parameters())
+            self.model_ema.copy_to(self.model)
+            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.model.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["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]
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    def q_mean_variance(self, x_start, t):
+        """
+        Get the distribution q(x_t | x_0).
+        :param x_start: the [N x C x ...] tensor of noiseless inputs.
+        :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
+        :return: A tuple (mean, variance, log_variance), all of x_start's shape.
+        """
+        mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
+        variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
+        log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
+        return mean, variance, log_variance
+
+    def predict_start_from_noise(self, x_t, t, noise):
+        return (
+                extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
+                extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
+        )
+
+    def q_posterior(self, x_start, x_t, t):
+        posterior_mean = (
+                extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
+                extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
+        )
+        posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
+        posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
+        return posterior_mean, posterior_variance, posterior_log_variance_clipped
+
+    def p_mean_variance(self, x, t, clip_denoised: bool):
+        model_out = self.model(x, t)
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
+        b, *_, device = *x.shape, x.device
+        model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
+        noise = noise_like(x.shape, device, repeat_noise)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, shape, return_intermediates=False):
+        device = self.betas.device
+        b = shape[0]
+        img = torch.randn(shape, device=device)
+        intermediates = [img]
+        for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
+            img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
+                                clip_denoised=self.clip_denoised)
+            if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
+                intermediates.append(img)
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, batch_size=16, return_intermediates=False):
+        image_size = self.image_size
+        channels = self.channels
+        return self.p_sample_loop((batch_size, channels, image_size, image_size),
+                                  return_intermediates=return_intermediates)
+
+    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 get_loss(self, pred, target, mean=True):
+        if self.loss_type == 'l1':
+            loss = (target - pred).abs()
+            if mean:
+                loss = loss.mean()
+        elif self.loss_type == 'l2':
+            if mean:
+                loss = torch.nn.functional.mse_loss(target, pred)
+            else:
+                loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
+        else:
+            raise NotImplementedError("unknown loss type '{loss_type}'")
+
+        return loss
+
+    def p_losses(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_out = self.model(x_noisy, t)
+
+        loss_dict = {}
+        if self.parameterization == "eps":
+            target = noise
+        elif self.parameterization == "x0":
+            target = x_start
+        else:
+            raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
+
+        loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
+
+        log_prefix = 'train' if self.training else 'val'
+
+        loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
+        loss_simple = loss.mean() * self.l_simple_weight
+
+        loss_vlb = (self.lvlb_weights[t] * loss).mean()
+        loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
+
+        loss = loss_simple + self.original_elbo_weight * loss_vlb
+
+        loss_dict.update({f'{log_prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def forward(self, x, *args, **kwargs):
+        # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+        # assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        return self.p_losses(x, t, *args, **kwargs)
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = rearrange(x, 'b h w c -> b c h w')
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def shared_step(self, batch):
+        x = self.get_input(batch, self.first_stage_key)
+        loss, loss_dict = self(x)
+        return loss, loss_dict
+
+    def training_step(self, batch, batch_idx):
+        loss, loss_dict = self.shared_step(batch)
+
+        self.log_dict(loss_dict, prog_bar=True,
+                      logger=True, on_step=True, on_epoch=True)
+
+        self.log("global_step", self.global_step,
+                 prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        if self.use_scheduler:
+            lr = self.optimizers().param_groups[0]['lr']
+            self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        return loss
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        _, loss_dict_no_ema = self.shared_step(batch)
+        with self.ema_scope():
+            _, loss_dict_ema = self.shared_step(batch)
+            loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
+        self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+        self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self.model)
+
+    def _get_rows_from_list(self, samples):
+        n_imgs_per_row = len(samples)
+        denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.first_stage_key)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        x = x.to(self.device)[:N]
+        log["inputs"] = x
+
+        # get diffusion row
+        diffusion_row = list()
+        x_start = x[:n_row]
+
+        for t in range(self.num_timesteps):
+            if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                t = t.to(self.device).long()
+                noise = torch.randn_like(x_start)
+                x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+                diffusion_row.append(x_noisy)
+
+        log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
+
+            log["samples"] = samples
+            log["denoise_row"] = self._get_rows_from_list(denoise_row)
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.learn_logvar:
+            params = params + [self.logvar]
+        opt = torch.optim.AdamW(params, lr=lr)
+        return opt
+
+
+class LatentDiffusion(DDPM):
+    """main class"""
+    def __init__(self,
+                 first_stage_config,
+                 cond_stage_config,
+                 num_timesteps_cond=None,
+                 cond_stage_key="image",
+                 cond_stage_trainable=False,
+                 concat_mode=True,
+                 cond_stage_forward=None,
+                 conditioning_key=None,
+                 scale_factor=1.0,
+                 scale_by_std=False,
+                 *args, **kwargs):
+        self.num_timesteps_cond = default(num_timesteps_cond, 1)
+        self.scale_by_std = scale_by_std
+        assert self.num_timesteps_cond <= kwargs['timesteps']
+        # for backwards compatibility after implementation of DiffusionWrapper
+        if conditioning_key is None:
+            conditioning_key = 'concat' if concat_mode else 'crossattn'
+        if cond_stage_config == '__is_unconditional__':
+            conditioning_key = None
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", [])
+        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
+        self.concat_mode = concat_mode
+        self.cond_stage_trainable = cond_stage_trainable
+        self.cond_stage_key = cond_stage_key
+        try:
+            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
+        except:
+            self.num_downs = 0
+        if not scale_by_std:
+            self.scale_factor = scale_factor
+        else:
+            self.register_buffer('scale_factor', torch.tensor(scale_factor))
+        self.instantiate_first_stage(first_stage_config)
+        self.instantiate_cond_stage(cond_stage_config)
+        self.cond_stage_forward = cond_stage_forward
+        self.clip_denoised = False
+        self.bbox_tokenizer = None  
+
+        self.restarted_from_ckpt = False
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys)
+            self.restarted_from_ckpt = True
+
+    def make_cond_schedule(self, ):
+        self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
+        ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
+        self.cond_ids[:self.num_timesteps_cond] = ids
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
+            assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+            x = super().get_input(batch, self.first_stage_key)
+            x = x.to(self.device)
+            encoder_posterior = self.encode_first_stage(x)
+            z = self.get_first_stage_encoding(encoder_posterior).detach()
+            del self.scale_factor
+            self.register_buffer('scale_factor', 1. / z.flatten().std())
+            print(f"setting self.scale_factor to {self.scale_factor}")
+            print("### USING STD-RESCALING ###")
+
+    def register_schedule(self,
+                          given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
+
+        self.shorten_cond_schedule = self.num_timesteps_cond > 1
+        if self.shorten_cond_schedule:
+            self.make_cond_schedule()
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+    def instantiate_cond_stage(self, config):
+        if not self.cond_stage_trainable:
+            if config == "__is_first_stage__":
+                print("Using first stage also as cond stage.")
+                self.cond_stage_model = self.first_stage_model
+            elif config == "__is_unconditional__":
+                print(f"Training {self.__class__.__name__} as an unconditional model.")
+                self.cond_stage_model = None
+                # self.be_unconditional = True
+            else:
+                model = instantiate_from_config(config)
+                self.cond_stage_model = model.eval()
+                self.cond_stage_model.train = disabled_train
+                for param in self.cond_stage_model.parameters():
+                    param.requires_grad = False
+        else:
+            assert config != '__is_first_stage__'
+            assert config != '__is_unconditional__'
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model
+
+    def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
+        denoise_row = []
+        for zd in tqdm(samples, desc=desc):
+            denoise_row.append(self.decode_first_stage(zd.to(self.device),
+                                                            force_not_quantize=force_no_decoder_quantization))
+        n_imgs_per_row = len(denoise_row)
+        denoise_row = torch.stack(denoise_row)  # n_log_step, n_row, C, H, W
+        denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    def get_first_stage_encoding(self, encoder_posterior):
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample()
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
+        return self.scale_factor * z
+
+    def get_learned_conditioning(self, c):
+        if self.cond_stage_forward is None:
+            if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
+                c = self.cond_stage_model.encode(c)
+                if isinstance(c, DiagonalGaussianDistribution):
+                    c = c.mode()
+            else:
+                c = self.cond_stage_model(c)
+        else:
+            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+        return c
+
+    def meshgrid(self, h, w):
+        y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
+        x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
+
+        arr = torch.cat([y, x], dim=-1)
+        return arr
+
+    def delta_border(self, h, w):
+        """
+        :param h: height
+        :param w: width
+        :return: normalized distance to image border,
+         wtith min distance = 0 at border and max dist = 0.5 at image center
+        """
+        lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
+        arr = self.meshgrid(h, w) / lower_right_corner
+        dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
+        dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
+        edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
+        return edge_dist
+
+    def get_weighting(self, h, w, Ly, Lx, device):
+        weighting = self.delta_border(h, w)
+        weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
+                               self.split_input_params["clip_max_weight"], )
+        weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
+
+        if self.split_input_params["tie_braker"]:
+            L_weighting = self.delta_border(Ly, Lx)
+            L_weighting = torch.clip(L_weighting,
+                                     self.split_input_params["clip_min_tie_weight"],
+                                     self.split_input_params["clip_max_tie_weight"])
+
+            L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
+            weighting = weighting * L_weighting
+        return weighting
+
+    def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1):  # todo load once not every time, shorten code
+        """
+        :param x: img of size (bs, c, h, w)
+        :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
+        """
+        bs, nc, h, w = x.shape
+
+        # number of crops in image
+        Ly = (h - kernel_size[0]) // stride[0] + 1
+        Lx = (w - kernel_size[1]) // stride[1] + 1
+
+        if uf == 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
+
+            weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h, w)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
+
+        elif uf > 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
+                                dilation=1, padding=0,
+                                stride=(stride[0] * uf, stride[1] * uf))
+            fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h * uf, w * uf)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
+
+        elif df > 1 and uf == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
+                                dilation=1, padding=0,
+                                stride=(stride[0] // df, stride[1] // df))
+            fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h // df, w // df)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
+
+        else:
+            raise NotImplementedError
+
+        return fold, unfold, normalization, weighting
+
+    @torch.no_grad()
+    def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
+                  cond_key=None, return_original_cond=False, bs=None):
+        x = super().get_input(batch, k)
+        if bs is not None:
+            x = x[:bs]
+        x = x.to(self.device)
+        encoder_posterior = self.encode_first_stage(x)
+        z = self.get_first_stage_encoding(encoder_posterior).detach()
+
+        if self.model.conditioning_key is not None:
+            if cond_key is None:
+                cond_key = self.cond_stage_key
+            if cond_key != self.first_stage_key:
+                if cond_key in ['caption', 'coordinates_bbox']:
+                    xc = batch[cond_key]
+                elif cond_key == 'class_label':
+                    xc = batch
+                else:
+                    xc = super().get_input(batch, cond_key).to(self.device)
+            else:
+                xc = x
+            if not self.cond_stage_trainable or force_c_encode:
+                if isinstance(xc, dict) or isinstance(xc, list):
+                    # import pudb; pudb.set_trace()
+                    c = self.get_learned_conditioning(xc)
+                else:
+                    c = self.get_learned_conditioning(xc.to(self.device))
+            else:
+                c = xc
+            if bs is not None:
+                c = c[:bs]
+
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                ckey = __conditioning_keys__[self.model.conditioning_key]
+                c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
+
+        else:
+            c = None
+            xc = None
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                c = {'pos_x': pos_x, 'pos_y': pos_y}
+        out = [z, c]
+        if return_first_stage_outputs:
+            xrec = self.decode_first_stage(z)
+            out.extend([x, xrec])
+        if return_original_cond:
+            out.append(xc)
+        return out
+
+    @torch.no_grad()
+    def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    # same as above but without decorator
+    def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):  
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                df = self.split_input_params["vqf"]
+                self.split_input_params['original_image_size'] = x.shape[-2:]
+                bs, nc, h, w = x.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
+                z = unfold(x)  # (bn, nc * prod(**ks), L)
+                # Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
+                               for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)
+                o = o * weighting
+
+                # Reverse reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization
+                return decoded
+
+            else:
+                return self.first_stage_model.encode(x)
+        else:
+            return self.first_stage_model.encode(x)
+
+    def shared_step(self, batch, **kwargs):
+        x, c = self.get_input(batch, self.first_stage_key)
+        loss = self(x, c)
+        return loss
+
+    def forward(self, x, c, *args, **kwargs):
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        if self.model.conditioning_key is not None:
+            assert c is not None
+            if self.cond_stage_trainable:
+                c = self.get_learned_conditioning(c)
+            if self.shorten_cond_schedule:  # TODO: drop this option
+                tc = self.cond_ids[t].to(self.device)
+                c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
+        return self.p_losses(x, c, t, *args, **kwargs)
+
+    def _rescale_annotations(self, bboxes, crop_coordinates):  # TODO: move to dataset
+        def rescale_bbox(bbox):
+            x0 = clamp((bbox[0] - crop_coordinates[0]) / crop_coordinates[2])
+            y0 = clamp((bbox[1] - crop_coordinates[1]) / crop_coordinates[3])
+            w = min(bbox[2] / crop_coordinates[2], 1 - x0)
+            h = min(bbox[3] / crop_coordinates[3], 1 - y0)
+            return x0, y0, w, h
+
+        return [rescale_bbox(b) for b in bboxes]
+
+    def apply_model(self, x_noisy, t, cond, return_ids=False):
+
+        if isinstance(cond, dict):
+            # hybrid case, cond is exptected to be a dict
+            pass
+        else:
+            if not isinstance(cond, list):
+                cond = [cond]
+            key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
+            cond = {key: cond}
+
+        if hasattr(self, "split_input_params"):
+            assert len(cond) == 1  # todo can only deal with one conditioning atm
+            assert not return_ids  
+            ks = self.split_input_params["ks"]  # eg. (128, 128)
+            stride = self.split_input_params["stride"]  # eg. (64, 64)
+
+            h, w = x_noisy.shape[-2:]
+
+            fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
+
+            z = unfold(x_noisy)  # (bn, nc * prod(**ks), L)
+            # Reshape to img shape
+            z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+            z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
+
+            if self.cond_stage_key in ["image", "LR_image", "segmentation",
+                                       'bbox_img'] and self.model.conditioning_key:  # todo check for completeness
+                c_key = next(iter(cond.keys()))  # get key
+                c = next(iter(cond.values()))  # get value
+                assert (len(c) == 1)  # todo extend to list with more than one elem
+                c = c[0]  # get element
+
+                c = unfold(c)
+                c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
+
+            elif self.cond_stage_key == 'coordinates_bbox':
+                assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
+
+                # assuming padding of unfold is always 0 and its dilation is always 1
+                n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
+                full_img_h, full_img_w = self.split_input_params['original_image_size']
+                # as we are operating on latents, we need the factor from the original image size to the
+                # spatial latent size to properly rescale the crops for regenerating the bbox annotations
+                num_downs = self.first_stage_model.encoder.num_resolutions - 1
+                rescale_latent = 2 ** (num_downs)
+
+                # get top left postions of patches as conforming for the bbbox tokenizer, therefore we
+                # need to rescale the tl patch coordinates to be in between (0,1)
+                tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
+                                         rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
+                                        for patch_nr in range(z.shape[-1])]
+
+                # patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
+                patch_limits = [(x_tl, y_tl,
+                                 rescale_latent * ks[0] / full_img_w,
+                                 rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
+                # patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
+
+                # tokenize crop coordinates for the bounding boxes of the respective patches
+                patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
+                                      for bbox in patch_limits]  # list of length l with tensors of shape (1, 2)
+                print(patch_limits_tknzd[0].shape)
+                # cut tknzd crop position from conditioning
+                assert isinstance(cond, dict), 'cond must be dict to be fed into model'
+                cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
+                print(cut_cond.shape)
+
+                adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
+                adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
+                print(adapted_cond.shape)
+                adapted_cond = self.get_learned_conditioning(adapted_cond)
+                print(adapted_cond.shape)
+                adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
+                print(adapted_cond.shape)
+
+                cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
+
+            else:
+                cond_list = [cond for i in range(z.shape[-1])]  # Todo make this more efficient
+
+            # apply model by loop over crops
+            output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
+            assert not isinstance(output_list[0],
+                                  tuple)  # todo cant deal with multiple model outputs check this never happens
+
+            o = torch.stack(output_list, axis=-1)
+            o = o * weighting
+            # Reverse reshape to img shape
+            o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+            # stitch crops together
+            x_recon = fold(o) / normalization
+
+        else:
+            x_recon = self.model(x_noisy, t, **cond)
+
+        if isinstance(x_recon, tuple) and not return_ids:
+            return x_recon[0]
+        else:
+            return x_recon
+
+    def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
+        return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
+               extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+
+    def _prior_bpd(self, x_start):
+        """
+        Get the prior KL term for the variational lower-bound, measured in
+        bits-per-dim.
+        This term can't be optimized, as it only depends on the encoder.
+        :param x_start: the [N x C x ...] tensor of inputs.
+        :return: a batch of [N] KL values (in bits), one per batch element.
+        """
+        batch_size = x_start.shape[0]
+        t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
+        qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
+        kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
+        return mean_flat(kl_prior) / np.log(2.0)
+
+    def p_losses(self, x_start, cond, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_output = self.apply_model(x_noisy, t, cond)
+
+        loss_dict = {}
+        prefix = 'train' if self.training else 'val'
+
+        if self.parameterization == "x0":
+            target = x_start
+        elif self.parameterization == "eps":
+            target = noise
+        else:
+            raise NotImplementedError()
+
+        loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
+        loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
+
+        logvar_t = self.logvar[t].to(self.device)
+        loss = loss_simple / torch.exp(logvar_t) + logvar_t
+        # loss = loss_simple / torch.exp(self.logvar) + self.logvar
+        if self.learn_logvar:
+            loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
+            loss_dict.update({'logvar': self.logvar.data.mean()})
+
+        loss = self.l_simple_weight * loss.mean()
+
+        loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
+        loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
+        loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
+        loss += (self.original_elbo_weight * loss_vlb)
+        loss_dict.update({f'{prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
+                        return_x0=False, score_corrector=None, corrector_kwargs=None):
+        t_in = t
+        model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
+
+        if score_corrector is not None:
+            assert self.parameterization == "eps"
+            model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
+
+        if return_codebook_ids:
+            model_out, logits = model_out
+
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        else:
+            raise NotImplementedError()
+
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+        if quantize_denoised:
+            x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        if return_codebook_ids:
+            return model_mean, posterior_variance, posterior_log_variance, logits
+        elif return_x0:
+            return model_mean, posterior_variance, posterior_log_variance, x_recon
+        else:
+            return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
+                 return_codebook_ids=False, quantize_denoised=False, return_x0=False,
+                 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
+        b, *_, device = *x.shape, x.device
+        outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
+                                       return_codebook_ids=return_codebook_ids,
+                                       quantize_denoised=quantize_denoised,
+                                       return_x0=return_x0,
+                                       score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+        if return_codebook_ids:
+            raise DeprecationWarning("Support dropped.")
+            model_mean, _, model_log_variance, logits = outputs
+        elif return_x0:
+            model_mean, _, model_log_variance, x0 = outputs
+        else:
+            model_mean, _, model_log_variance = outputs
+
+        noise = noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+
+        if return_codebook_ids:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
+        if return_x0:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
+        else:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
+                              img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
+                              score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
+                              log_every_t=None):
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        timesteps = self.num_timesteps
+        if batch_size is not None:
+            b = batch_size if batch_size is not None else shape[0]
+            shape = [batch_size] + list(shape)
+        else:
+            b = batch_size = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=self.device)
+        else:
+            img = x_T
+        intermediates = []
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
+                        total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+        if type(temperature) == float:
+            temperature = [temperature] * timesteps
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=self.device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img, x0_partial = self.p_sample(img, cond, ts,
+                                            clip_denoised=self.clip_denoised,
+                                            quantize_denoised=quantize_denoised, return_x0=True,
+                                            temperature=temperature[i], noise_dropout=noise_dropout,
+                                            score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(x0_partial)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_loop(self, cond, shape, return_intermediates=False,
+                      x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, start_T=None,
+                      log_every_t=None):
+
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        device = self.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        intermediates = [img]
+        if timesteps is None:
+            timesteps = self.num_timesteps
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+
+        if mask is not None:
+            assert x0 is not None
+            assert x0.shape[2:3] == mask.shape[2:3]  # spatial size has to match
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img = self.p_sample(img, cond, ts,
+                                clip_denoised=self.clip_denoised,
+                                quantize_denoised=quantize_denoised)
+            if mask is not None:
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(img)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
+               verbose=True, timesteps=None, quantize_denoised=False,
+               mask=None, x0=None, shape=None,**kwargs):
+        if shape is None:
+            shape = (batch_size, self.channels, self.image_size, self.image_size)
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+        return self.p_sample_loop(cond,
+                                  shape,
+                                  return_intermediates=return_intermediates, x_T=x_T,
+                                  verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
+                                  mask=mask, x0=x0)
+
+    @torch.no_grad()
+    def sample_log(self,cond,batch_size,ddim, ddim_steps,**kwargs):
+
+        if ddim:
+            ddim_sampler = DDIMSampler(self)
+            shape = (self.channels, self.image_size, self.image_size)
+            samples, intermediates =ddim_sampler.sample(ddim_steps,batch_size,
+                                                        shape,cond,verbose=False,**kwargs)
+
+        else:
+            samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
+                                                 return_intermediates=True,**kwargs)
+
+        return samples, intermediates
+
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
+                   plot_diffusion_rows=True, **kwargs):
+
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
+                                           return_first_stage_outputs=True,
+                                           force_c_encode=True,
+                                           return_original_cond=True,
+                                           bs=N)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if plot_diffusion_rows:
+            # get diffusion row
+            diffusion_row = list()
+            z_start = z[:n_row]
+            for t in range(self.num_timesteps):
+                if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                    t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                    t = t.to(self.device).long()
+                    noise = torch.randn_like(z_start)
+                    z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
+                    diffusion_row.append(self.decode_first_stage(z_noisy))
+
+            diffusion_row = torch.stack(diffusion_row)  # n_log_step, n_row, C, H, W
+            diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
+            diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
+            diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
+            log["diffusion_row"] = diffusion_grid
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                         ddim_steps=ddim_steps,eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+            if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
+                    self.first_stage_model, IdentityFirstStage):
+                # also display when quantizing x0 while sampling
+                with self.ema_scope("Plotting Quantized Denoised"):
+                    samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                             ddim_steps=ddim_steps,eta=ddim_eta,
+                                                             quantize_denoised=True)
+                    # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
+                    #                                      quantize_denoised=True)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_x0_quantized"] = x_samples
+
+            if inpaint:
+                # make a simple center square
+                b, h, w = z.shape[0], z.shape[2], z.shape[3]
+                mask = torch.ones(N, h, w).to(self.device)
+                # zeros will be filled in
+                mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
+                mask = mask[:, None, ...]
+                with self.ema_scope("Plotting Inpaint"):
+
+                    samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_inpainting"] = x_samples
+                log["mask"] = mask
+
+                # outpaint
+                with self.ema_scope("Plotting Outpaint"):
+                    samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_outpainting"] = x_samples
+
+        if plot_progressive_rows:
+            with self.ema_scope("Plotting Progressives"):
+                img, progressives = self.progressive_denoising(c,
+                                                               shape=(self.channels, self.image_size, self.image_size),
+                                                               batch_size=N)
+            prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
+            log["progressive_row"] = prog_row
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.cond_stage_trainable:
+            print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
+            params = params + list(self.cond_stage_model.parameters())
+        if self.learn_logvar:
+            print('Diffusion model optimizing logvar')
+            params.append(self.logvar)
+        opt = torch.optim.AdamW(params, lr=lr)
+        if self.use_scheduler:
+            assert 'target' in self.scheduler_config
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [opt], scheduler
+        return opt
+
+    @torch.no_grad()
+    def to_rgb(self, x):
+        x = x.float()
+        if not hasattr(self, "colorize"):
+            self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
+        x = nn.functional.conv2d(x, weight=self.colorize)
+        x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
+        return x
+
+
+class DiffusionWrapper(pl.LightningModule):
+    def __init__(self, diff_model_config, conditioning_key):
+        super().__init__()
+        self.diffusion_model = instantiate_from_config(diff_model_config)
+        self.conditioning_key = conditioning_key
+        assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm']
+
+    def forward(self, x, t, c_concat: list = None, c_crossattn: list = None):
+        if self.conditioning_key is None:
+            out = self.diffusion_model(x, t)
+        elif self.conditioning_key == 'concat':
+            xc = torch.cat([x] + c_concat, dim=1)
+            out = self.diffusion_model(xc, t)
+        elif self.conditioning_key == 'crossattn':
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(x, t, context=cc)
+        elif self.conditioning_key == 'hybrid':
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc)
+        elif self.conditioning_key == 'adm':
+            cc = c_crossattn[0]
+            out = self.diffusion_model(x, t, y=cc)
+        else:
+            raise NotImplementedError()
+
+        return out
+
+
+class Layout2ImgDiffusion(LatentDiffusion):
+    # TODO: move all layout-specific hacks to this class
+    def __init__(self, cond_stage_key, *args, **kwargs):
+        assert cond_stage_key == 'coordinates_bbox', 'Layout2ImgDiffusion only for cond_stage_key="coordinates_bbox"'
+        super().__init__(cond_stage_key=cond_stage_key, *args, **kwargs)
+
+    def log_images(self, batch, N=8, *args, **kwargs):
+        logs = super().log_images(batch=batch, N=N, *args, **kwargs)
+
+        key = 'train' if self.training else 'validation'
+        dset = self.trainer.datamodule.datasets[key]
+        mapper = dset.conditional_builders[self.cond_stage_key]
+
+        bbox_imgs = []
+        map_fn = lambda catno: dset.get_textual_label(dset.get_category_id(catno))
+        for tknzd_bbox in batch[self.cond_stage_key][:N]:
+            bboximg = mapper.plot(tknzd_bbox.detach().cpu(), map_fn, (256, 256))
+            bbox_imgs.append(bboximg)
+
+        cond_img = torch.stack(bbox_imgs, dim=0)
+        logs['bbox_image'] = cond_img
+        return logs

stable-diffusion/ldm/models/diffusion/dpm_solver/__init__.py

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

stable-diffusion/ldm/models/diffusion/dpm_solver/dpm_solver.py

@@ -0,0 +1,1184 @@
+import torch
+import torch.nn.functional as F
+import math
+
+
+class NoiseScheduleVP:
+    def __init__(
+            self,
+            schedule='discrete',
+            betas=None,
+            alphas_cumprod=None,
+            continuous_beta_0=0.1,
+            continuous_beta_1=20.,
+        ):
+        """Create a wrapper class for the forward SDE (VP type).
+
+        ***
+        Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
+                We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
+        ***
+
+        The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
+        We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
+        Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
+
+            log_alpha_t = self.marginal_log_mean_coeff(t)
+            sigma_t = self.marginal_std(t)
+            lambda_t = self.marginal_lambda(t)
+
+        Moreover, as lambda(t) is an invertible function, we also support its inverse function:
+
+            t = self.inverse_lambda(lambda_t)
+
+        ===============================================================
+
+        We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
+
+        1. For discrete-time DPMs:
+
+            For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
+                t_i = (i + 1) / N
+            e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
+            We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
+
+            Args:
+                betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
+                alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
+
+            Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
+
+            **Important**:  Please pay special attention for the args for `alphas_cumprod`:
+                The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
+                    q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
+                Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
+                    alpha_{t_n} = \sqrt{\hat{alpha_n}},
+                and
+                    log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
+
+
+        2. For continuous-time DPMs:
+
+            We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
+            schedule are the default settings in DDPM and improved-DDPM:
+
+            Args:
+                beta_min: A `float` number. The smallest beta for the linear schedule.
+                beta_max: A `float` number. The largest beta for the linear schedule.
+                cosine_s: A `float` number. The hyperparameter in the cosine schedule.
+                cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
+                T: A `float` number. The ending time of the forward process.
+
+        ===============================================================
+
+        Args:
+            schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
+                    'linear' or 'cosine' for continuous-time DPMs.
+        Returns:
+            A wrapper object of the forward SDE (VP type).
+        
+        ===============================================================
+
+        Example:
+
+        # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', betas=betas)
+
+        # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
+
+        # For continuous-time DPMs (VPSDE), linear schedule:
+        >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
+
+        """
+
+        if schedule not in ['discrete', 'linear', 'cosine']:
+            raise ValueError("Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(schedule))
+
+        self.schedule = schedule
+        if schedule == 'discrete':
+            if betas is not None:
+                log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
+            else:
+                assert alphas_cumprod is not None
+                log_alphas = 0.5 * torch.log(alphas_cumprod)
+            self.total_N = len(log_alphas)
+            self.T = 1.
+            self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1))
+            self.log_alpha_array = log_alphas.reshape((1, -1,))
+        else:
+            self.total_N = 1000
+            self.beta_0 = continuous_beta_0
+            self.beta_1 = continuous_beta_1
+            self.cosine_s = 0.008
+            self.cosine_beta_max = 999.
+            self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * (1. + self.cosine_s) / math.pi - self.cosine_s
+            self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.))
+            self.schedule = schedule
+            if schedule == 'cosine':
+                # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
+                # Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
+                self.T = 0.9946
+            else:
+                self.T = 1.
+
+    def marginal_log_mean_coeff(self, t):
+        """
+        Compute log(alpha_t) of a given continuous-time label t in [0, T].
+        """
+        if self.schedule == 'discrete':
+            return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device), self.log_alpha_array.to(t.device)).reshape((-1))
+        elif self.schedule == 'linear':
+            return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
+        elif self.schedule == 'cosine':
+            log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.))
+            log_alpha_t =  log_alpha_fn(t) - self.cosine_log_alpha_0
+            return log_alpha_t
+
+    def marginal_alpha(self, t):
+        """
+        Compute alpha_t of a given continuous-time label t in [0, T].
+        """
+        return torch.exp(self.marginal_log_mean_coeff(t))
+
+    def marginal_std(self, t):
+        """
+        Compute sigma_t of a given continuous-time label t in [0, T].
+        """
+        return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t)))
+
+    def marginal_lambda(self, t):
+        """
+        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
+        """
+        log_mean_coeff = self.marginal_log_mean_coeff(t)
+        log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff))
+        return log_mean_coeff - log_std
+
+    def inverse_lambda(self, lamb):
+        """
+        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
+        """
+        if self.schedule == 'linear':
+            tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
+            Delta = self.beta_0**2 + tmp
+            return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
+        elif self.schedule == 'discrete':
+            log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb)
+            t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]), torch.flip(self.t_array.to(lamb.device), [1]))
+            return t.reshape((-1,))
+        else:
+            log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
+            t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * (1. + self.cosine_s) / math.pi - self.cosine_s
+            t = t_fn(log_alpha)
+            return t
+
+
+def model_wrapper(
+    model,
+    noise_schedule,
+    model_type="noise",
+    model_kwargs={},
+    guidance_type="uncond",
+    condition=None,
+    unconditional_condition=None,
+    guidance_scale=1.,
+    classifier_fn=None,
+    classifier_kwargs={},
+):
+    """Create a wrapper function for the noise prediction model.
+
+    DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
+    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
+
+    We support four types of the diffusion model by setting `model_type`:
+
+        1. "noise": noise prediction model. (Trained by predicting noise).
+
+        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
+
+        3. "v": velocity prediction model. (Trained by predicting the velocity).
+            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
+
+            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
+                arXiv preprint arXiv:2202.00512 (2022).
+            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
+                arXiv preprint arXiv:2210.02303 (2022).
+    
+        4. "score": marginal score function. (Trained by denoising score matching).
+            Note that the score function and the noise prediction model follows a simple relationship:
+            ```
+                noise(x_t, t) = -sigma_t * score(x_t, t)
+            ```
+
+    We support three types of guided sampling by DPMs by setting `guidance_type`:
+        1. "uncond": unconditional sampling by DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+
+        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            `` 
+
+            The input `classifier_fn` has the following format:
+            ``
+                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
+            ``
+
+            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
+                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
+
+        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
+            `` 
+            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
+
+            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
+                arXiv preprint arXiv:2207.12598 (2022).
+        
+
+    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
+    or continuous-time labels (i.e. epsilon to T).
+
+    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
+    ``
+        def model_fn(x, t_continuous) -> noise:
+            t_input = get_model_input_time(t_continuous)
+            return noise_pred(model, x, t_input, **model_kwargs)         
+    ``
+    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
+
+    ===============================================================
+
+    Args:
+        model: A diffusion model with the corresponding format described above.
+        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+        model_type: A `str`. The parameterization type of the diffusion model.
+                    "noise" or "x_start" or "v" or "score".
+        model_kwargs: A `dict`. A dict for the other inputs of the model function.
+        guidance_type: A `str`. The type of the guidance for sampling.
+                    "uncond" or "classifier" or "classifier-free".
+        condition: A pytorch tensor. The condition for the guided sampling.
+                    Only used for "classifier" or "classifier-free" guidance type.
+        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
+                    Only used for "classifier-free" guidance type.
+        guidance_scale: A `float`. The scale for the guided sampling.
+        classifier_fn: A classifier function. Only used for the classifier guidance.
+        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
+    Returns:
+        A noise prediction model that accepts the noised data and the continuous time as the inputs.
+    """
+
+    def get_model_input_time(t_continuous):
+        """
+        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
+        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
+        For continuous-time DPMs, we just use `t_continuous`.
+        """
+        if noise_schedule.schedule == 'discrete':
+            return (t_continuous - 1. / noise_schedule.total_N) * 1000.
+        else:
+            return t_continuous
+
+    def noise_pred_fn(x, t_continuous, cond=None):
+        if t_continuous.reshape((-1,)).shape[0] == 1:
+            t_continuous = t_continuous.expand((x.shape[0]))
+        t_input = get_model_input_time(t_continuous)
+        if cond is None:
+            output = model(x, t_input, **model_kwargs)
+        else:
+            output = model(x, t_input, cond, **model_kwargs)
+        if model_type == "noise":
+            return output
+        elif model_type == "x_start":
+            alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
+            dims = x.dim()
+            return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims)
+        elif model_type == "v":
+            alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
+            dims = x.dim()
+            return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x
+        elif model_type == "score":
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            dims = x.dim()
+            return -expand_dims(sigma_t, dims) * output
+
+    def cond_grad_fn(x, t_input):
+        """
+        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
+        """
+        with torch.enable_grad():
+            x_in = x.detach().requires_grad_(True)
+            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
+            return torch.autograd.grad(log_prob.sum(), x_in)[0]
+
+    def model_fn(x, t_continuous):
+        """
+        The noise predicition model function that is used for DPM-Solver.
+        """
+        if t_continuous.reshape((-1,)).shape[0] == 1:
+            t_continuous = t_continuous.expand((x.shape[0]))
+        if guidance_type == "uncond":
+            return noise_pred_fn(x, t_continuous)
+        elif guidance_type == "classifier":
+            assert classifier_fn is not None
+            t_input = get_model_input_time(t_continuous)
+            cond_grad = cond_grad_fn(x, t_input)
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            noise = noise_pred_fn(x, t_continuous)
+            return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad
+        elif guidance_type == "classifier-free":
+            if guidance_scale == 1. or unconditional_condition is None:
+                return noise_pred_fn(x, t_continuous, cond=condition)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t_continuous] * 2)
+                c_in = torch.cat([unconditional_condition, condition])
+                noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
+                return noise_uncond + guidance_scale * (noise - noise_uncond)
+
+    assert model_type in ["noise", "x_start", "v"]
+    assert guidance_type in ["uncond", "classifier", "classifier-free"]
+    return model_fn
+
+
+class DPM_Solver:
+    def __init__(self, model_fn, noise_schedule, predict_x0=False, thresholding=False, max_val=1.):
+        """Construct a DPM-Solver. 
+
+        We support both the noise prediction model ("predicting epsilon") and the data prediction model ("predicting x0").
+        If `predict_x0` is False, we use the solver for the noise prediction model (DPM-Solver).
+        If `predict_x0` is True, we use the solver for the data prediction model (DPM-Solver++).
+            In such case, we further support the "dynamic thresholding" in [1] when `thresholding` is True.
+            The "dynamic thresholding" can greatly improve the sample quality for pixel-space DPMs with large guidance scales.
+
+        Args:
+            model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
+                ``
+                def model_fn(x, t_continuous):
+                    return noise
+                ``
+            noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+            predict_x0: A `bool`. If true, use the data prediction model; else, use the noise prediction model.
+            thresholding: A `bool`. Valid when `predict_x0` is True. Whether to use the "dynamic thresholding" in [1].
+            max_val: A `float`. Valid when both `predict_x0` and `thresholding` are True. The max value for thresholding.
+        
+        [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
+        """
+        self.model = model_fn
+        self.noise_schedule = noise_schedule
+        self.predict_x0 = predict_x0
+        self.thresholding = thresholding
+        self.max_val = max_val
+
+    def noise_prediction_fn(self, x, t):
+        """
+        Return the noise prediction model.
+        """
+        return self.model(x, t)
+
+    def data_prediction_fn(self, x, t):
+        """
+        Return the data prediction model (with thresholding).
+        """
+        noise = self.noise_prediction_fn(x, t)
+        dims = x.dim()
+        alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t)
+        x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims)
+        if self.thresholding:
+            p = 0.995   # A hyperparameter in the paper of "Imagen" [1].
+            s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
+            s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims)
+            x0 = torch.clamp(x0, -s, s) / s
+        return x0
+
+    def model_fn(self, x, t):
+        """
+        Convert the model to the noise prediction model or the data prediction model. 
+        """
+        if self.predict_x0:
+            return self.data_prediction_fn(x, t)
+        else:
+            return self.noise_prediction_fn(x, t)
+
+    def get_time_steps(self, skip_type, t_T, t_0, N, device):
+        """Compute the intermediate time steps for sampling.
+
+        Args:
+            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
+                - 'logSNR': uniform logSNR for the time steps.
+                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
+                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            N: A `int`. The total number of the spacing of the time steps.
+            device: A torch device.
+        Returns:
+            A pytorch tensor of the time steps, with the shape (N + 1,).
+        """
+        if skip_type == 'logSNR':
+            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
+            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
+            logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device)
+            return self.noise_schedule.inverse_lambda(logSNR_steps)
+        elif skip_type == 'time_uniform':
+            return torch.linspace(t_T, t_0, N + 1).to(device)
+        elif skip_type == 'time_quadratic':
+            t_order = 2
+            t = torch.linspace(t_T**(1. / t_order), t_0**(1. / t_order), N + 1).pow(t_order).to(device)
+            return t
+        else:
+            raise ValueError("Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type))
+
+    def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device):
+        """
+        Get the order of each step for sampling by the singlestep DPM-Solver.
+
+        We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
+        Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
+            - If order == 1:
+                We take `steps` of DPM-Solver-1 (i.e. DDIM).
+            - If order == 2:
+                - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
+                - If steps % 2 == 0, we use K steps of DPM-Solver-2.
+                - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
+            - If order == 3:
+                - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
+                - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
+                - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
+                - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
+
+        ============================================
+        Args:
+            order: A `int`. The max order for the solver (2 or 3).
+            steps: A `int`. The total number of function evaluations (NFE).
+            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
+                - 'logSNR': uniform logSNR for the time steps.
+                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
+                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            device: A torch device.
+        Returns:
+            orders: A list of the solver order of each step.
+        """
+        if order == 3:
+            K = steps // 3 + 1
+            if steps % 3 == 0:
+                orders = [3,] * (K - 2) + [2, 1]
+            elif steps % 3 == 1:
+                orders = [3,] * (K - 1) + [1]
+            else:
+                orders = [3,] * (K - 1) + [2]
+        elif order == 2:
+            if steps % 2 == 0:
+                K = steps // 2
+                orders = [2,] * K
+            else:
+                K = steps // 2 + 1
+                orders = [2,] * (K - 1) + [1]
+        elif order == 1:
+            K = 1
+            orders = [1,] * steps
+        else:
+            raise ValueError("'order' must be '1' or '2' or '3'.")
+        if skip_type == 'logSNR':
+            # To reproduce the results in DPM-Solver paper
+            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
+        else:
+            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[torch.cumsum(torch.tensor([0,] + orders)).to(device)]
+        return timesteps_outer, orders
+
+    def denoise_to_zero_fn(self, x, s):
+        """
+        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization. 
+        """
+        return self.data_prediction_fn(x, s)
+
+    def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
+        """
+        DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
+            t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        ns = self.noise_schedule
+        dims = x.dim()
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t)
+        sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        if self.predict_x0:
+            phi_1 = torch.expm1(-h)
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_t = (
+                expand_dims(sigma_t / sigma_s, dims) * x
+                - expand_dims(alpha_t * phi_1, dims) * model_s
+            )
+            if return_intermediate:
+                return x_t, {'model_s': model_s}
+            else:
+                return x_t
+        else:
+            phi_1 = torch.expm1(h)
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_t = (
+                expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
+                - expand_dims(sigma_t * phi_1, dims) * model_s
+            )
+            if return_intermediate:
+                return x_t, {'model_s': model_s}
+            else:
+                return x_t
+
+    def singlestep_dpm_solver_second_update(self, x, s, t, r1=0.5, model_s=None, return_intermediate=False, solver_type='dpm_solver'):
+        """
+        Singlestep solver DPM-Solver-2 from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
+            t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
+            r1: A `float`. The hyperparameter of the second-order solver.
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
+            solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ['dpm_solver', 'taylor']:
+            raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
+        if r1 is None:
+            r1 = 0.5
+        ns = self.noise_schedule
+        dims = x.dim()
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        lambda_s1 = lambda_s + r1 * h
+        s1 = ns.inverse_lambda(lambda_s1)
+        log_alpha_s, log_alpha_s1, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(t)
+        sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t)
+        alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
+
+        if self.predict_x0:
+            phi_11 = torch.expm1(-r1 * h)
+            phi_1 = torch.expm1(-h)
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_s1 = (
+                expand_dims(sigma_s1 / sigma_s, dims) * x
+                - expand_dims(alpha_s1 * phi_11, dims) * model_s
+            )
+            model_s1 = self.model_fn(x_s1, s1)
+            if solver_type == 'dpm_solver':
+                x_t = (
+                    expand_dims(sigma_t / sigma_s, dims) * x
+                    - expand_dims(alpha_t * phi_1, dims) * model_s
+                    - (0.5 / r1) * expand_dims(alpha_t * phi_1, dims) * (model_s1 - model_s)
+                )
+            elif solver_type == 'taylor':
+                x_t = (
+                    expand_dims(sigma_t / sigma_s, dims) * x
+                    - expand_dims(alpha_t * phi_1, dims) * model_s
+                    + (1. / r1) * expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * (model_s1 - model_s)
+                )
+        else:
+            phi_11 = torch.expm1(r1 * h)
+            phi_1 = torch.expm1(h)
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_s1 = (
+                expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
+                - expand_dims(sigma_s1 * phi_11, dims) * model_s
+            )
+            model_s1 = self.model_fn(x_s1, s1)
+            if solver_type == 'dpm_solver':
+                x_t = (
+                    expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
+                    - expand_dims(sigma_t * phi_1, dims) * model_s
+                    - (0.5 / r1) * expand_dims(sigma_t * phi_1, dims) * (model_s1 - model_s)
+                )
+            elif solver_type == 'taylor':
+                x_t = (
+                    expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
+                    - expand_dims(sigma_t * phi_1, dims) * model_s
+                    - (1. / r1) * expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * (model_s1 - model_s)
+                )
+        if return_intermediate:
+            return x_t, {'model_s': model_s, 'model_s1': model_s1}
+        else:
+            return x_t
+
+    def singlestep_dpm_solver_third_update(self, x, s, t, r1=1./3., r2=2./3., model_s=None, model_s1=None, return_intermediate=False, solver_type='dpm_solver'):
+        """
+        Singlestep solver DPM-Solver-3 from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
+            t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
+            r1: A `float`. The hyperparameter of the third-order solver.
+            r2: A `float`. The hyperparameter of the third-order solver.
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
+                If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
+            solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ['dpm_solver', 'taylor']:
+            raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
+        if r1 is None:
+            r1 = 1. / 3.
+        if r2 is None:
+            r2 = 2. / 3.
+        ns = self.noise_schedule
+        dims = x.dim()
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        lambda_s1 = lambda_s + r1 * h
+        lambda_s2 = lambda_s + r2 * h
+        s1 = ns.inverse_lambda(lambda_s1)
+        s2 = ns.inverse_lambda(lambda_s2)
+        log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(s2), ns.marginal_log_mean_coeff(t)
+        sigma_s, sigma_s1, sigma_s2, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(s2), ns.marginal_std(t)
+        alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t)
+
+        if self.predict_x0:
+            phi_11 = torch.expm1(-r1 * h)
+            phi_12 = torch.expm1(-r2 * h)
+            phi_1 = torch.expm1(-h)
+            phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.
+            phi_2 = phi_1 / h + 1.
+            phi_3 = phi_2 / h - 0.5
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            if model_s1 is None:
+                x_s1 = (
+                    expand_dims(sigma_s1 / sigma_s, dims) * x
+                    - expand_dims(alpha_s1 * phi_11, dims) * model_s
+                )
+                model_s1 = self.model_fn(x_s1, s1)
+            x_s2 = (
+                expand_dims(sigma_s2 / sigma_s, dims) * x
+                - expand_dims(alpha_s2 * phi_12, dims) * model_s
+                + r2 / r1 * expand_dims(alpha_s2 * phi_22, dims) * (model_s1 - model_s)
+            )
+            model_s2 = self.model_fn(x_s2, s2)
+            if solver_type == 'dpm_solver':
+                x_t = (
+                    expand_dims(sigma_t / sigma_s, dims) * x
+                    - expand_dims(alpha_t * phi_1, dims) * model_s
+                    + (1. / r2) * expand_dims(alpha_t * phi_2, dims) * (model_s2 - model_s)
+                )
+            elif solver_type == 'taylor':
+                D1_0 = (1. / r1) * (model_s1 - model_s)
+                D1_1 = (1. / r2) * (model_s2 - model_s)
+                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
+                D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
+                x_t = (
+                    expand_dims(sigma_t / sigma_s, dims) * x
+                    - expand_dims(alpha_t * phi_1, dims) * model_s
+                    + expand_dims(alpha_t * phi_2, dims) * D1
+                    - expand_dims(alpha_t * phi_3, dims) * D2
+                )
+        else:
+            phi_11 = torch.expm1(r1 * h)
+            phi_12 = torch.expm1(r2 * h)
+            phi_1 = torch.expm1(h)
+            phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.
+            phi_2 = phi_1 / h - 1.
+            phi_3 = phi_2 / h - 0.5
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            if model_s1 is None:
+                x_s1 = (
+                    expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
+                    - expand_dims(sigma_s1 * phi_11, dims) * model_s
+                )
+                model_s1 = self.model_fn(x_s1, s1)
+            x_s2 = (
+                expand_dims(torch.exp(log_alpha_s2 - log_alpha_s), dims) * x
+                - expand_dims(sigma_s2 * phi_12, dims) * model_s
+                - r2 / r1 * expand_dims(sigma_s2 * phi_22, dims) * (model_s1 - model_s)
+            )
+            model_s2 = self.model_fn(x_s2, s2)
+            if solver_type == 'dpm_solver':
+                x_t = (
+                    expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
+                    - expand_dims(sigma_t * phi_1, dims) * model_s
+                    - (1. / r2) * expand_dims(sigma_t * phi_2, dims) * (model_s2 - model_s)
+                )
+            elif solver_type == 'taylor':
+                D1_0 = (1. / r1) * (model_s1 - model_s)
+                D1_1 = (1. / r2) * (model_s2 - model_s)
+                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
+                D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
+                x_t = (
+                    expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
+                    - expand_dims(sigma_t * phi_1, dims) * model_s
+                    - expand_dims(sigma_t * phi_2, dims) * D1
+                    - expand_dims(sigma_t * phi_3, dims) * D2
+                )
+
+        if return_intermediate:
+            return x_t, {'model_s': model_s, 'model_s1': model_s1, 'model_s2': model_s2}
+        else:
+            return x_t
+
+    def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpm_solver"):
+        """
+        Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
+            t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
+            solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ['dpm_solver', 'taylor']:
+            raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
+        ns = self.noise_schedule
+        dims = x.dim()
+        model_prev_1, model_prev_0 = model_prev_list
+        t_prev_1, t_prev_0 = t_prev_list
+        lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t)
+        log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
+        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        h_0 = lambda_prev_0 - lambda_prev_1
+        h = lambda_t - lambda_prev_0
+        r0 = h_0 / h
+        D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
+        if self.predict_x0:
+            if solver_type == 'dpm_solver':
+                x_t = (
+                    expand_dims(sigma_t / sigma_prev_0, dims) * x
+                    - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
+                    - 0.5 * expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * D1_0
+                )
+            elif solver_type == 'taylor':
+                x_t = (
+                    expand_dims(sigma_t / sigma_prev_0, dims) * x
+                    - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
+                    + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1_0
+                )
+        else:
+            if solver_type == 'dpm_solver':
+                x_t = (
+                    expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
+                    - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
+                    - 0.5 * expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * D1_0
+                )
+            elif solver_type == 'taylor':
+                x_t = (
+                    expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
+                    - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
+                    - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1_0
+                )
+        return x_t
+
+    def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type='dpm_solver'):
+        """
+        Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
+            t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
+            solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        ns = self.noise_schedule
+        dims = x.dim()
+        model_prev_2, model_prev_1, model_prev_0 = model_prev_list
+        t_prev_2, t_prev_1, t_prev_0 = t_prev_list
+        lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_2), ns.marginal_lambda(t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t)
+        log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
+        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        h_1 = lambda_prev_1 - lambda_prev_2
+        h_0 = lambda_prev_0 - lambda_prev_1
+        h = lambda_t - lambda_prev_0
+        r0, r1 = h_0 / h, h_1 / h
+        D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
+        D1_1 = expand_dims(1. / r1, dims) * (model_prev_1 - model_prev_2)
+        D1 = D1_0 + expand_dims(r0 / (r0 + r1), dims) * (D1_0 - D1_1)
+        D2 = expand_dims(1. / (r0 + r1), dims) * (D1_0 - D1_1)
+        if self.predict_x0:
+            x_t = (
+                expand_dims(sigma_t / sigma_prev_0, dims) * x
+                - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
+                + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1
+                - expand_dims(alpha_t * ((torch.exp(-h) - 1. + h) / h**2 - 0.5), dims) * D2
+            )
+        else:
+            x_t = (
+                expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
+                - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
+                - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1
+                - expand_dims(sigma_t * ((torch.exp(h) - 1. - h) / h**2 - 0.5), dims) * D2
+            )
+        return x_t
+
+    def singlestep_dpm_solver_update(self, x, s, t, order, return_intermediate=False, solver_type='dpm_solver', r1=None, r2=None):
+        """
+        Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
+            t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
+            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
+            solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
+            r1: A `float`. The hyperparameter of the second-order or third-order solver.
+            r2: A `float`. The hyperparameter of the third-order solver.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if order == 1:
+            return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate)
+        elif order == 2:
+            return self.singlestep_dpm_solver_second_update(x, s, t, return_intermediate=return_intermediate, solver_type=solver_type, r1=r1)
+        elif order == 3:
+            return self.singlestep_dpm_solver_third_update(x, s, t, return_intermediate=return_intermediate, solver_type=solver_type, r1=r1, r2=r2)
+        else:
+            raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
+
+    def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type='dpm_solver'):
+        """
+        Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
+            t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
+            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
+            solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if order == 1:
+            return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1])
+        elif order == 2:
+            return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
+        elif order == 3:
+            return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
+        else:
+            raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
+
+    def dpm_solver_adaptive(self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5, solver_type='dpm_solver'):
+        """
+        The adaptive step size solver based on singlestep DPM-Solver.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `t_T`.
+            order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            h_init: A `float`. The initial step size (for logSNR).
+            atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
+            rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
+            theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
+            t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the 
+                current time and `t_0` is less than `t_err`. The default setting is 1e-5.
+            solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
+        Returns:
+            x_0: A pytorch tensor. The approximated solution at time `t_0`.
+
+        [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
+        """
+        ns = self.noise_schedule
+        s = t_T * torch.ones((x.shape[0],)).to(x)
+        lambda_s = ns.marginal_lambda(s)
+        lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
+        h = h_init * torch.ones_like(s).to(x)
+        x_prev = x
+        nfe = 0
+        if order == 2:
+            r1 = 0.5
+            lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True)
+            higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, solver_type=solver_type, **kwargs)
+        elif order == 3:
+            r1, r2 = 1. / 3., 2. / 3.
+            lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, return_intermediate=True, solver_type=solver_type)
+            higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(x, s, t, r1=r1, r2=r2, solver_type=solver_type, **kwargs)
+        else:
+            raise ValueError("For adaptive step size solver, order must be 2 or 3, got {}".format(order))
+        while torch.abs((s - t_0)).mean() > t_err:
+            t = ns.inverse_lambda(lambda_s + h)
+            x_lower, lower_noise_kwargs = lower_update(x, s, t)
+            x_higher = higher_update(x, s, t, **lower_noise_kwargs)
+            delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)))
+            norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True))
+            E = norm_fn((x_higher - x_lower) / delta).max()
+            if torch.all(E <= 1.):
+                x = x_higher
+                s = t
+                x_prev = x_lower
+                lambda_s = ns.marginal_lambda(s)
+            h = torch.min(theta * h * torch.float_power(E, -1. / order).float(), lambda_0 - lambda_s)
+            nfe += order
+        print('adaptive solver nfe', nfe)
+        return x
+
+    def sample(self, x, steps=20, t_start=None, t_end=None, order=3, skip_type='time_uniform',
+        method='singlestep', lower_order_final=True, denoise_to_zero=False, solver_type='dpm_solver',
+        atol=0.0078, rtol=0.05,
+    ):
+        """
+        Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
+
+        =====================================================
+
+        We support the following algorithms for both noise prediction model and data prediction model:
+            - 'singlestep':
+                Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver. 
+                We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
+                The total number of function evaluations (NFE) == `steps`.
+                Given a fixed NFE == `steps`, the sampling procedure is:
+                    - If `order` == 1:
+                        - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
+                    - If `order` == 2:
+                        - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
+                        - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
+                        - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
+                    - If `order` == 3:
+                        - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
+                        - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
+                        - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
+                        - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
+            - 'multistep':
+                Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
+                We initialize the first `order` values by lower order multistep solvers.
+                Given a fixed NFE == `steps`, the sampling procedure is:
+                    Denote K = steps.
+                    - If `order` == 1:
+                        - We use K steps of DPM-Solver-1 (i.e. DDIM).
+                    - If `order` == 2:
+                        - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
+                    - If `order` == 3:
+                        - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
+            - 'singlestep_fixed':
+                Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
+                We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
+            - 'adaptive':
+                Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
+                We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
+                You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
+                (NFE) and the sample quality.
+                    - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
+                    - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
+
+        =====================================================
+
+        Some advices for choosing the algorithm:
+            - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
+                Use singlestep DPM-Solver ("DPM-Solver-fast" in the paper) with `order = 3`.
+                e.g.
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=False)
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
+                            skip_type='time_uniform', method='singlestep')
+            - For **guided sampling with large guidance scale** by DPMs:
+                Use multistep DPM-Solver with `predict_x0 = True` and `order = 2`.
+                e.g.
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=True)
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
+                            skip_type='time_uniform', method='multistep')
+
+        We support three types of `skip_type`:
+            - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
+            - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
+            - 'time_quadratic': quadratic time for the time steps.
+
+        =====================================================
+        Args:
+            x: A pytorch tensor. The initial value at time `t_start`
+                e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
+            steps: A `int`. The total number of function evaluations (NFE).
+            t_start: A `float`. The starting time of the sampling.
+                If `T` is None, we use self.noise_schedule.T (default is 1.0).
+            t_end: A `float`. The ending time of the sampling.
+                If `t_end` is None, we use 1. / self.noise_schedule.total_N.
+                e.g. if total_N == 1000, we have `t_end` == 1e-3.
+                For discrete-time DPMs:
+                    - We recommend `t_end` == 1. / self.noise_schedule.total_N.
+                For continuous-time DPMs:
+                    - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
+            order: A `int`. The order of DPM-Solver.
+            skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
+            method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
+            denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step.
+                Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1).
+
+                This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and
+                score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID
+                for diffusion models sampling by diffusion SDEs for low-resolutional images
+                (such as CIFAR-10). However, we observed that such trick does not matter for
+                high-resolutional images. As it needs an additional NFE, we do not recommend
+                it for high-resolutional images.
+            lower_order_final: A `bool`. Whether to use lower order solvers at the final steps.
+                Only valid for `method=multistep` and `steps < 15`. We empirically find that
+                this trick is a key to stabilizing the sampling by DPM-Solver with very few steps
+                (especially for steps <= 10). So we recommend to set it to be `True`.
+            solver_type: A `str`. The taylor expansion type for the solver. `dpm_solver` or `taylor`. We recommend `dpm_solver`.
+            atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
+            rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
+        Returns:
+            x_end: A pytorch tensor. The approximated solution at time `t_end`.
+
+        """
+        t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end
+        t_T = self.noise_schedule.T if t_start is None else t_start
+        device = x.device
+        if method == 'adaptive':
+            with torch.no_grad():
+                x = self.dpm_solver_adaptive(x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol, solver_type=solver_type)
+        elif method == 'multistep':
+            assert steps >= order
+            timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device)
+            assert timesteps.shape[0] - 1 == steps
+            with torch.no_grad():
+                vec_t = timesteps[0].expand((x.shape[0]))
+                model_prev_list = [self.model_fn(x, vec_t)]
+                t_prev_list = [vec_t]
+                # Init the first `order` values by lower order multistep DPM-Solver.
+                for init_order in range(1, order):
+                    vec_t = timesteps[init_order].expand(x.shape[0])
+                    x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, init_order, solver_type=solver_type)
+                    model_prev_list.append(self.model_fn(x, vec_t))
+                    t_prev_list.append(vec_t)
+                # Compute the remaining values by `order`-th order multistep DPM-Solver.
+                for step in range(order, steps + 1):
+                    vec_t = timesteps[step].expand(x.shape[0])
+                    if lower_order_final and steps < 15:
+                        step_order = min(order, steps + 1 - step)
+                    else:
+                        step_order = order
+                    x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, step_order, solver_type=solver_type)
+                    for i in range(order - 1):
+                        t_prev_list[i] = t_prev_list[i + 1]
+                        model_prev_list[i] = model_prev_list[i + 1]
+                    t_prev_list[-1] = vec_t
+                    # We do not need to evaluate the final model value.
+                    if step < steps:
+                        model_prev_list[-1] = self.model_fn(x, vec_t)
+        elif method in ['singlestep', 'singlestep_fixed']:
+            if method == 'singlestep':
+                timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(steps=steps, order=order, skip_type=skip_type, t_T=t_T, t_0=t_0, device=device)
+            elif method == 'singlestep_fixed':
+                K = steps // order
+                orders = [order,] * K
+                timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device)
+            for i, order in enumerate(orders):
+                t_T_inner, t_0_inner = timesteps_outer[i], timesteps_outer[i + 1]
+                timesteps_inner = self.get_time_steps(skip_type=skip_type, t_T=t_T_inner.item(), t_0=t_0_inner.item(), N=order, device=device)
+                lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
+                vec_s, vec_t = t_T_inner.tile(x.shape[0]), t_0_inner.tile(x.shape[0])
+                h = lambda_inner[-1] - lambda_inner[0]
+                r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
+                r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
+                x = self.singlestep_dpm_solver_update(x, vec_s, vec_t, order, solver_type=solver_type, r1=r1, r2=r2)
+        if denoise_to_zero:
+            x = self.denoise_to_zero_fn(x, torch.ones((x.shape[0],)).to(device) * t_0)
+        return x
+
+
+
+#############################################################
+# other utility functions
+#############################################################
+
+def interpolate_fn(x, xp, yp):
+    """
+    A piecewise linear function y = f(x), using xp and yp as keypoints.
+    We implement f(x) in a differentiable way (i.e. applicable for autograd).
+    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
+
+    Args:
+        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
+        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
+        yp: PyTorch tensor with shape [C, K].
+    Returns:
+        The function values f(x), with shape [N, C].
+    """
+    N, K = x.shape[0], xp.shape[1]
+    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
+    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
+    x_idx = torch.argmin(x_indices, dim=2)
+    cand_start_idx = x_idx - 1
+    start_idx = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(1, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
+        ),
+    )
+    end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1)
+    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
+    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
+    start_idx2 = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(0, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
+        ),
+    )
+    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
+    start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2)
+    end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2)
+    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
+    return cand
+
+
+def expand_dims(v, dims):
+    """
+    Expand the tensor `v` to the dim `dims`.
+
+    Args:
+        `v`: a PyTorch tensor with shape [N].
+        `dim`: a `int`.
+    Returns:
+        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
+    """
+    return v[(...,) + (None,)*(dims - 1)]

stable-diffusion/ldm/models/diffusion/dpm_solver/sampler.py

@@ -0,0 +1,82 @@
+"""SAMPLING ONLY."""
+
+import torch
+
+from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver
+
+
+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="noise",
+            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

stable-diffusion/ldm/models/diffusion/plms.py

@@ -0,0 +1,236 @@
+"""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
+
+
+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, ...
+               **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,
+                                                    )
+        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,):
+        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)
+            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):
+        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)
+            # 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

stable-diffusion/ldm/modules/attention.py

@@ -0,0 +1,261 @@
+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 ldm.modules.diffusionmodules.util import checkpoint
+
+
+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 LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
+        k = k.softmax(dim=-1)  
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        return self.to_out(out)
+
+
+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
+
+        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
+        attn = sim.softmax(dim=-1)
+
+        out = einsum('b i j, b j d -> b i d', attn, v)
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True):
+        super().__init__()
+        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout)  # is a self-attention
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = CrossAttention(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)) + 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
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None):
+        super().__init__()
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = Normalize(in_channels)
+
+        self.proj_in = nn.Conv2d(in_channels,
+                                 inner_dim,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
+                for d in range(depth)]
+        )
+
+        self.proj_out = zero_module(nn.Conv2d(inner_dim,
+                                              in_channels,
+                                              kernel_size=1,
+                                              stride=1,
+                                              padding=0))
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c')
+        for block in self.transformer_blocks:
+            x = block(x, context=context)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
+        x = self.proj_out(x)
+        return x + x_in

stable-diffusion/ldm/modules/diffusionmodules/model.py

@@ -0,0 +1,835 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+
+from ldm.util import instantiate_from_config
+from ldm.modules.attention import LinearAttention
+
+
+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 LinAttnBlock(LinearAttention):
+    """to match AttnBlock usage"""
+    def __init__(self, in_channels):
+        super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
+
+
+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_
+
+
+def make_attn(in_channels, attn_type="vanilla"):
+    assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown'
+    print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        return AttnBlock(in_channels)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        return LinAttnBlock(in_channels)
+
+
+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
+
+class FirstStagePostProcessor(nn.Module):
+
+    def __init__(self, ch_mult:list, in_channels,
+                 pretrained_model:nn.Module=None,
+                 reshape=False,
+                 n_channels=None,
+                 dropout=0.,
+                 pretrained_config=None):
+        super().__init__()
+        if pretrained_config is None:
+            assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.pretrained_model = pretrained_model
+        else:
+            assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.instantiate_pretrained(pretrained_config)
+
+        self.do_reshape = reshape
+
+        if n_channels is None:
+            n_channels = self.pretrained_model.encoder.ch
+
+        self.proj_norm = Normalize(in_channels,num_groups=in_channels//2)
+        self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3,
+                            stride=1,padding=1)
+
+        blocks = []
+        downs = []
+        ch_in = n_channels
+        for m in ch_mult:
+            blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout))
+            ch_in = m * n_channels
+            downs.append(Downsample(ch_in, with_conv=False))
+
+        self.model = nn.ModuleList(blocks)
+        self.downsampler = nn.ModuleList(downs)
+
+
+    def instantiate_pretrained(self, config):
+        model = instantiate_from_config(config)
+        self.pretrained_model = model.eval()
+        # self.pretrained_model.train = False
+        for param in self.pretrained_model.parameters():
+            param.requires_grad = False
+
+
+    @torch.no_grad()
+    def encode_with_pretrained(self,x):
+        c = self.pretrained_model.encode(x)
+        if isinstance(c, DiagonalGaussianDistribution):
+            c = c.mode()
+        return  c
+
+    def forward(self,x):
+        z_fs = self.encode_with_pretrained(x)
+        z = self.proj_norm(z_fs)
+        z = self.proj(z)
+        z = nonlinearity(z)
+
+        for submodel, downmodel in zip(self.model,self.downsampler):
+            z = submodel(z,temb=None)
+            z = downmodel(z)
+
+        if self.do_reshape:
+            z = rearrange(z,'b c h w -> b (h w) c')
+        return z
+

stable-diffusion/ldm/modules/diffusionmodules/openaimodel.py

@@ -0,0 +1,961 @@
+from abc import abstractmethod
+from functools import partial
+import math
+from typing import Iterable
+
+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
+
+
+# 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,
+    ):
+        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
+        self.num_res_blocks = num_res_blocks
+        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:
+            self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+
+        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 _ in range(num_res_blocks):
+                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
+                    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
+                        )
+                    )
+                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(
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        ),
+            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(num_res_blocks + 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
+                    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
+                        )
+                    )
+                if level and i == num_res_blocks:
+                    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 == (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)
+
+
+class EncoderUNetModel(nn.Module):
+    """
+    The half UNet model with attention and timestep embedding.
+    For usage, see UNet.
+    """
+
+    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,
+        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,
+        pool="adaptive",
+        *args,
+        **kwargs
+    ):
+        super().__init__()
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        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
+
+        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),
+        )
+
+        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 _ in range(num_res_blocks):
+                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:
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=num_head_channels,
+                            use_new_attention_order=use_new_attention_order,
+                        )
+                    )
+                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
+
+        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=num_head_channels,
+                use_new_attention_order=use_new_attention_order,
+            ),
+            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.pool = pool
+        if pool == "adaptive":
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                nn.AdaptiveAvgPool2d((1, 1)),
+                zero_module(conv_nd(dims, ch, out_channels, 1)),
+                nn.Flatten(),
+            )
+        elif pool == "attention":
+            assert num_head_channels != -1
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                AttentionPool2d(
+                    (image_size // ds), ch, num_head_channels, out_channels
+                ),
+            )
+        elif pool == "spatial":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                nn.ReLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        elif pool == "spatial_v2":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                normalization(2048),
+                nn.SiLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        else:
+            raise NotImplementedError(f"Unexpected {pool} pooling")
+
+    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)
+
+    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)
+
+    def forward(self, x, timesteps):
+        """
+        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.
+        :return: an [N x K] Tensor of outputs.
+        """
+        emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
+        results = []
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb)
+            if self.pool.startswith("spatial"):
+                results.append(h.type(x.dtype).mean(dim=(2, 3)))
+        h = self.middle_block(h, emb)
+        if self.pool.startswith("spatial"):
+            results.append(h.type(x.dtype).mean(dim=(2, 3)))
+            h = th.cat(results, axis=-1)
+            return self.out(h)
+        else:
+            h = h.type(x.dtype)
+            return self.out(h)
+

stable-diffusion/ldm/modules/diffusionmodules/util.py

@@ -0,0 +1,267 @@
+# 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:])
+
+        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():
+            # 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()

stable-diffusion/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)
+    )

stable-diffusion/ldm/modules/ema.py

@@ -0,0 +1,76 @@
+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 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)

stable-diffusion/ldm/modules/encoders/modules.py

@@ -0,0 +1,234 @@
+import torch
+import torch.nn as nn
+from functools import partial
+import clip
+from einops import rearrange, repeat
+from transformers import CLIPTokenizer, CLIPTextModel
+import kornia
+
+from ldm.modules.x_transformer import Encoder, TransformerWrapper  # TODO: can we directly rely on lucidrains code and simply add this as a reuirement? --> test
+
+
+class AbstractEncoder(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key='class'):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+
+    def forward(self, batch, key=None):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]
+        c = self.embedding(c)
+        return c
+
+
+class TransformerEmbedder(AbstractEncoder):
+    """Some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size, max_seq_len=77, device="cuda"):
+        super().__init__()
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer))
+
+    def forward(self, tokens):
+        tokens = tokens.to(self.device)  # meh
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, x):
+        return self(x)
+
+
+class BERTTokenizer(AbstractEncoder):
+    """ Uses a pretrained BERT tokenizer by huggingface. Vocab size: 30522 (?)"""
+    def __init__(self, device="cuda", vq_interface=True, max_length=77):
+        super().__init__()
+        from transformers import BertTokenizerFast  # TODO: add to reuquirements
+        self.tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased")
+        self.device = device
+        self.vq_interface = vq_interface
+        self.max_length = max_length
+
+    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)
+        return tokens
+
+    @torch.no_grad()
+    def encode(self, text):
+        tokens = self(text)
+        if not self.vq_interface:
+            return tokens
+        return None, None, [None, None, tokens]
+
+    def decode(self, text):
+        return text
+
+
+class BERTEmbedder(AbstractEncoder):
+    """Uses the BERT tokenizr model and add some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size=30522, max_seq_len=77,
+                 device="cuda",use_tokenizer=True, embedding_dropout=0.0):
+        super().__init__()
+        self.use_tknz_fn = use_tokenizer
+        if self.use_tknz_fn:
+            self.tknz_fn = BERTTokenizer(vq_interface=False, max_length=max_seq_len)
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer),
+                                              emb_dropout=embedding_dropout)
+
+    def forward(self, text):
+        if self.use_tknz_fn:
+            tokens = self.tknz_fn(text)#.to(self.device)
+        else:
+            tokens = text
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, text):
+        # output of length 77
+        return self(text)
+
+
+class SpatialRescaler(nn.Module):
+    def __init__(self,
+                 n_stages=1,
+                 method='bilinear',
+                 multiplier=0.5,
+                 in_channels=3,
+                 out_channels=None,
+                 bias=False):
+        super().__init__()
+        self.n_stages = n_stages
+        assert self.n_stages >= 0
+        assert method in ['nearest','linear','bilinear','trilinear','bicubic','area']
+        self.multiplier = multiplier
+        self.interpolator = partial(torch.nn.functional.interpolate, mode=method)
+        self.remap_output = out_channels is not None
+        if self.remap_output:
+            print(f'Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing.')
+            self.channel_mapper = nn.Conv2d(in_channels,out_channels,1,bias=bias)
+
+    def forward(self,x):
+        for stage in range(self.n_stages):
+            x = self.interpolator(x, scale_factor=self.multiplier)
+
+
+        if self.remap_output:
+            x = self.channel_mapper(x)
+        return x
+
+    def encode(self, x):
+        return self(x)
+
+class FrozenCLIPEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from Hugging Face)"""
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77):
+        super().__init__()
+        self.tokenizer = CLIPTokenizer.from_pretrained(version)
+        self.transformer = CLIPTextModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length
+        self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        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 FrozenCLIPTextEmbedder(nn.Module):
+    """
+    Uses the CLIP transformer encoder for text.
+    """
+    def __init__(self, version='ViT-L/14', device="cuda", max_length=77, n_repeat=1, normalize=True):
+        super().__init__()
+        self.model, _ = clip.load(version, jit=False, device="cpu")
+        self.device = device
+        self.max_length = max_length
+        self.n_repeat = n_repeat
+        self.normalize = normalize
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        tokens = clip.tokenize(text).to(self.device)
+        z = self.model.encode_text(tokens)
+        if self.normalize:
+            z = z / torch.linalg.norm(z, dim=1, keepdim=True)
+        return z
+
+    def encode(self, text):
+        z = self(text)
+        if z.ndim==2:
+            z = z[:, None, :]
+        z = repeat(z, 'b 1 d -> b k d', k=self.n_repeat)
+        return z
+
+
+class FrozenClipImageEmbedder(nn.Module):
+    """
+        Uses the CLIP image encoder.
+        """
+    def __init__(
+            self,
+            model,
+            jit=False,
+            device='cuda' if torch.cuda.is_available() else 'cpu',
+            antialias=False,
+        ):
+        super().__init__()
+        self.model, _ = clip.load(name=model, device=device, jit=jit)
+
+        self.antialias = antialias
+
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+
+    def preprocess(self, x):
+        # normalize to [0,1]
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic',align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x):
+        # x is assumed to be in range [-1,1]
+        return self.model.encode_image(self.preprocess(x))
+
+
+if __name__ == "__main__":
+    from ldm.util import count_params
+    model = FrozenCLIPEmbedder()
+    count_params(model, verbose=True)

stable-diffusion/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

stable-diffusion/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')
+
+

stable-diffusion/ldm/modules/image_degradation/bsrgan_light.py

@@ -0,0 +1,650 @@
+# -*- 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.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
+
+    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.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(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.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=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):
+    """
+    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.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=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)
+    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')

stable-diffusion/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)

stable-diffusion/ldm/modules/losses/__init__.py

@@ -0,0 +1 @@
+from ldm.modules.losses.contperceptual import LPIPSWithDiscriminator

stable-diffusion/ldm/modules/losses/contperceptual.py

@@ -0,0 +1,111 @@
+import torch
+import torch.nn as nn
+
+from taming.modules.losses.vqperceptual import *  # TODO: taming dependency yes/no?
+
+
+class LPIPSWithDiscriminator(nn.Module):
+    def __init__(self, disc_start, logvar_init=0.0, kl_weight=1.0, pixelloss_weight=1.0,
+                 disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0,
+                 perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
+                 disc_loss="hinge"):
+
+        super().__init__()
+        assert disc_loss in ["hinge", "vanilla"]
+        self.kl_weight = kl_weight
+        self.pixel_weight = pixelloss_weight
+        self.perceptual_loss = LPIPS().eval()
+        self.perceptual_weight = perceptual_weight
+        # output log variance
+        self.logvar = nn.Parameter(torch.ones(size=()) * logvar_init)
+
+        self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
+                                                 n_layers=disc_num_layers,
+                                                 use_actnorm=use_actnorm
+                                                 ).apply(weights_init)
+        self.discriminator_iter_start = disc_start
+        self.disc_loss = hinge_d_loss if disc_loss == "hinge" else vanilla_d_loss
+        self.disc_factor = disc_factor
+        self.discriminator_weight = disc_weight
+        self.disc_conditional = disc_conditional
+
+    def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
+        if last_layer is not None:
+            nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
+            g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
+        else:
+            nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
+            g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]
+
+        d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
+        d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
+        d_weight = d_weight * self.discriminator_weight
+        return d_weight
+
+    def forward(self, inputs, reconstructions, posteriors, optimizer_idx,
+                global_step, last_layer=None, cond=None, split="train",
+                weights=None):
+        rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
+        if self.perceptual_weight > 0:
+            p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous())
+            rec_loss = rec_loss + self.perceptual_weight * p_loss
+
+        nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar
+        weighted_nll_loss = nll_loss
+        if weights is not None:
+            weighted_nll_loss = weights*nll_loss
+        weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0]
+        nll_loss = torch.sum(nll_loss) / nll_loss.shape[0]
+        kl_loss = posteriors.kl()
+        kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
+
+        # now the GAN part
+        if optimizer_idx == 0:
+            # generator update
+            if cond is None:
+                assert not self.disc_conditional
+                logits_fake = self.discriminator(reconstructions.contiguous())
+            else:
+                assert self.disc_conditional
+                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1))
+            g_loss = -torch.mean(logits_fake)
+
+            if self.disc_factor > 0.0:
+                try:
+                    d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
+                except RuntimeError:
+                    assert not self.training
+                    d_weight = torch.tensor(0.0)
+            else:
+                d_weight = torch.tensor(0.0)
+
+            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
+            loss = weighted_nll_loss + self.kl_weight * kl_loss + d_weight * disc_factor * g_loss
+
+            log = {"{}/total_loss".format(split): loss.clone().detach().mean(), "{}/logvar".format(split): self.logvar.detach(),
+                   "{}/kl_loss".format(split): kl_loss.detach().mean(), "{}/nll_loss".format(split): nll_loss.detach().mean(),
+                   "{}/rec_loss".format(split): rec_loss.detach().mean(),
+                   "{}/d_weight".format(split): d_weight.detach(),
+                   "{}/disc_factor".format(split): torch.tensor(disc_factor),
+                   "{}/g_loss".format(split): g_loss.detach().mean(),
+                   }
+            return loss, log
+
+        if optimizer_idx == 1:
+            # second pass for discriminator update
+            if cond is None:
+                logits_real = self.discriminator(inputs.contiguous().detach())
+                logits_fake = self.discriminator(reconstructions.contiguous().detach())
+            else:
+                logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1))
+                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1))
+
+            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
+            d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
+
+            log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(),
+                   "{}/logits_real".format(split): logits_real.detach().mean(),
+                   "{}/logits_fake".format(split): logits_fake.detach().mean()
+                   }
+            return d_loss, log
+

stable-diffusion/ldm/modules/losses/vqperceptual.py

@@ -0,0 +1,167 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+from einops import repeat
+
+from taming.modules.discriminator.model import NLayerDiscriminator, weights_init
+from taming.modules.losses.lpips import LPIPS
+from taming.modules.losses.vqperceptual import hinge_d_loss, vanilla_d_loss
+
+
+def hinge_d_loss_with_exemplar_weights(logits_real, logits_fake, weights):
+    assert weights.shape[0] == logits_real.shape[0] == logits_fake.shape[0]
+    loss_real = torch.mean(F.relu(1. - logits_real), dim=[1,2,3])
+    loss_fake = torch.mean(F.relu(1. + logits_fake), dim=[1,2,3])
+    loss_real = (weights * loss_real).sum() / weights.sum()
+    loss_fake = (weights * loss_fake).sum() / weights.sum()
+    d_loss = 0.5 * (loss_real + loss_fake)
+    return d_loss
+
+def adopt_weight(weight, global_step, threshold=0, value=0.):
+    if global_step < threshold:
+        weight = value
+    return weight
+
+
+def measure_perplexity(predicted_indices, n_embed):
+    # src: https://github.com/karpathy/deep-vector-quantization/blob/main/model.py
+    # eval cluster perplexity. when perplexity == num_embeddings then all clusters are used exactly equally
+    encodings = F.one_hot(predicted_indices, n_embed).float().reshape(-1, n_embed)
+    avg_probs = encodings.mean(0)
+    perplexity = (-(avg_probs * torch.log(avg_probs + 1e-10)).sum()).exp()
+    cluster_use = torch.sum(avg_probs > 0)
+    return perplexity, cluster_use
+
+def l1(x, y):
+    return torch.abs(x-y)
+
+
+def l2(x, y):
+    return torch.pow((x-y), 2)
+
+
+class VQLPIPSWithDiscriminator(nn.Module):
+    def __init__(self, disc_start, codebook_weight=1.0, pixelloss_weight=1.0,
+                 disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0,
+                 perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
+                 disc_ndf=64, disc_loss="hinge", n_classes=None, perceptual_loss="lpips",
+                 pixel_loss="l1"):
+        super().__init__()
+        assert disc_loss in ["hinge", "vanilla"]
+        assert perceptual_loss in ["lpips", "clips", "dists"]
+        assert pixel_loss in ["l1", "l2"]
+        self.codebook_weight = codebook_weight
+        self.pixel_weight = pixelloss_weight
+        if perceptual_loss == "lpips":
+            print(f"{self.__class__.__name__}: Running with LPIPS.")
+            self.perceptual_loss = LPIPS().eval()
+        else:
+            raise ValueError(f"Unknown perceptual loss: >> {perceptual_loss} <<")
+        self.perceptual_weight = perceptual_weight
+
+        if pixel_loss == "l1":
+            self.pixel_loss = l1
+        else:
+            self.pixel_loss = l2
+
+        self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
+                                                 n_layers=disc_num_layers,
+                                                 use_actnorm=use_actnorm,
+                                                 ndf=disc_ndf
+                                                 ).apply(weights_init)
+        self.discriminator_iter_start = disc_start
+        if disc_loss == "hinge":
+            self.disc_loss = hinge_d_loss
+        elif disc_loss == "vanilla":
+            self.disc_loss = vanilla_d_loss
+        else:
+            raise ValueError(f"Unknown GAN loss '{disc_loss}'.")
+        print(f"VQLPIPSWithDiscriminator running with {disc_loss} loss.")
+        self.disc_factor = disc_factor
+        self.discriminator_weight = disc_weight
+        self.disc_conditional = disc_conditional
+        self.n_classes = n_classes
+
+    def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
+        if last_layer is not None:
+            nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
+            g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
+        else:
+            nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
+            g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]
+
+        d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
+        d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
+        d_weight = d_weight * self.discriminator_weight
+        return d_weight
+
+    def forward(self, codebook_loss, inputs, reconstructions, optimizer_idx,
+                global_step, last_layer=None, cond=None, split="train", predicted_indices=None):
+        if not exists(codebook_loss):
+            codebook_loss = torch.tensor([0.]).to(inputs.device)
+        #rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
+        rec_loss = self.pixel_loss(inputs.contiguous(), reconstructions.contiguous())
+        if self.perceptual_weight > 0:
+            p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous())
+            rec_loss = rec_loss + self.perceptual_weight * p_loss
+        else:
+            p_loss = torch.tensor([0.0])
+
+        nll_loss = rec_loss
+        #nll_loss = torch.sum(nll_loss) / nll_loss.shape[0]
+        nll_loss = torch.mean(nll_loss)
+
+        # now the GAN part
+        if optimizer_idx == 0:
+            # generator update
+            if cond is None:
+                assert not self.disc_conditional
+                logits_fake = self.discriminator(reconstructions.contiguous())
+            else:
+                assert self.disc_conditional
+                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1))
+            g_loss = -torch.mean(logits_fake)
+
+            try:
+                d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
+            except RuntimeError:
+                assert not self.training
+                d_weight = torch.tensor(0.0)
+
+            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
+            loss = nll_loss + d_weight * disc_factor * g_loss + self.codebook_weight * codebook_loss.mean()
+
+            log = {"{}/total_loss".format(split): loss.clone().detach().mean(),
+                   "{}/quant_loss".format(split): codebook_loss.detach().mean(),
+                   "{}/nll_loss".format(split): nll_loss.detach().mean(),
+                   "{}/rec_loss".format(split): rec_loss.detach().mean(),
+                   "{}/p_loss".format(split): p_loss.detach().mean(),
+                   "{}/d_weight".format(split): d_weight.detach(),
+                   "{}/disc_factor".format(split): torch.tensor(disc_factor),
+                   "{}/g_loss".format(split): g_loss.detach().mean(),
+                   }
+            if predicted_indices is not None:
+                assert self.n_classes is not None
+                with torch.no_grad():
+                    perplexity, cluster_usage = measure_perplexity(predicted_indices, self.n_classes)
+                log[f"{split}/perplexity"] = perplexity
+                log[f"{split}/cluster_usage"] = cluster_usage
+            return loss, log
+
+        if optimizer_idx == 1:
+            # second pass for discriminator update
+            if cond is None:
+                logits_real = self.discriminator(inputs.contiguous().detach())
+                logits_fake = self.discriminator(reconstructions.contiguous().detach())
+            else:
+                logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1))
+                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1))
+
+            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
+            d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
+
+            log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(),
+                   "{}/logits_real".format(split): logits_real.detach().mean(),
+                   "{}/logits_fake".format(split): logits_fake.detach().mean()
+                   }
+            return d_loss, log