initial commit

bahjat-kawar · Nov 2, 2022 · e3704ce · e3704ce
commit e3704ce
Show file tree

Hide file tree

Showing 77 changed files with 5,098 additions and 0 deletions.
diff --git a/LICENSE b/LICENSE
@@ -0,0 +1,9 @@
+MIT License
+
+Copyright [2022] [Bahjat Kawar, Jiaming Song, Stefano Ermon, Michael Elad]
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
diff --git a/README.md b/README.md
@@ -0,0 +1,95 @@
+# JPEG Artifact Correction using Denoising Diffusion Restoration Models
+
+[arXiv](https://arxiv.org/abs/2209.11888) | [PDF](https://arxiv.org/pdf/2209.11888.pdf)
+
+[Bahjat Kawar](https://bahjat-kawar.github.io/)<sup>\* 1</sup>, [Jiaming Song](http://tsong.me)<sup>\* 2</sup>, [Stefano Ermon](http://cs.stanford.edu/~ermon)<sup>3</sup>, [Michael Elad](https://elad.cs.technion.ac.il/)<sup>1</sup><br />
+<sup>1</sup> Technion, <sup>2</sup>NVIDIA, <sup>3</sup>Stanford University, <sup>\*</sup>Equal contribution.
+
+![](assets/ddrm-jpeg-demo.png)
+
+
+We extend DDRM (Denoising Diffusion Restoration Models) for the problems of JPEG artifact correction and image dequantization.
+
+## Running the Experiments
+The code has been tested on PyTorch 1.8 and PyTorch 1.10. Please refer to `environment.yml` for a list of conda/mamba environments that can be used to run the code. The codebase is based heavily on the [original DDRM codebase](https://github.com/bahjat-kawar/ddrm).
+
+### Pretrained models
+We use pretrained models from [https://github.com/openai/guided-diffusion](https://github.com/openai/guided-diffusion), [https://github.com/pesser/pytorch_diffusion](https://github.com/pesser/pytorch_diffusion) and [https://github.com/ermongroup/SDEdit](https://github.com/ermongroup/SDEdit)
+
+We use 1,000 images from the ImageNet validation set for comparison with other methods. The list of images is taken from [https://github.com/XingangPan/deep-generative-prior/](https://github.com/XingangPan/deep-generative-prior/)
+
+The models and datasets are placed in the `exp/` folder as follows:
+```bash
+<exp> # a folder named by the argument `--exp` given to main.py
+├── datasets # all dataset files
+│   ├── celeba # all CelebA files
+│   ├── imagenet # all ImageNet files
+│   ├── ood # out of distribution ImageNet images
+│   ├── ood_bedroom # out of distribution bedroom images
+│   ├── ood_cat # out of distribution cat images
+│   └── ood_celeba # out of distribution CelebA images
+├── logs # contains checkpoints and samples produced during training
+│   ├── celeba
+│   │   └── celeba_hq.ckpt # the checkpoint file for CelebA-HQ
+│   ├── diffusion_models_converted
+│   │   └── ema_diffusion_lsun_<category>_model
+│   │       └── model-x.ckpt # the checkpoint file saved at the x-th training iteration
+│   ├── imagenet # ImageNet checkpoint files
+│   │   ├── 256x256_classifier.pt
+│   │   ├── 256x256_diffusion.pt
+│   │   ├── 256x256_diffusion_uncond.pt
+│   │   ├── 512x512_classifier.pt
+│   │   └── 512x512_diffusion.pt
+├── image_samples # contains generated samples
+└── imagenet_val_1k.txt # list of the 1k images used in ImageNet-1K.
+```
+
+### Sampling from the model
+
+The general command to sample from the model is as follows:
+```
+python main.py --ni --config {CONFIG}.yml --doc {DATASET} -i {IMAGE_FOLDER} --timesteps {STEPS} --init_timestep {INIT_T} --eta {ETA} --etaB {ETA_B} --deg {DEGRADATION} --num_avg_samples {NUM_AVG}
+```
+where the following are options
+- `ETA` is the eta hyperparameter in the paper. (default: `1`)
+- `ETA_B` is the eta_b hyperparameter in the paper. (default: `0.4`)
+- `STEPS` controls how many timesteps used in the process. (default: `20`)
+- `INIT_T` controls the timestep to start sampling from. (default: `300`)
+- `NUM_AVG` is the number of samples per input to average for the final result. (default: `1`)
+- `DEGREDATION` is the type of degredation used. (One of: `quant` for dequantization, or `jpegXX` for JPEG with quality factor `XX`, e.g. `jpeg80`)
+- `CONFIG` is the name of the config file (see `configs/` for a list), including hyperparameters such as batch size and network architectures.
+- `DATASET` is the name of the dataset used, to determine where the checkpoint file is found.
+- `IMAGE_FOLDER` is the name of the folder the resulting images will be placed in (default: `images`)
+
+For example, to use the default settings from the paper on the ImageNet 256x256 dataset, the problem of JPEG artifact correction for QF=80, and averaging 8 samples per input:
+```
+python main.py --ni --config imagenet_256.yml --doc imagenet -i imagenet --deg jpeg80 --num_avg_samples 8
+```
+The generated images are place in the `<exp>/image_samples/{IMAGE_FOLDER}` folder, where `orig_{id}.png`, `y0_{id}.png`, `{id}_-1.png` refer to the original, degraded, restored images respectively.
+
+The config files contain a setting controlling whether to test on samples from the trained dataset's distribution or not.
+
+## References and Acknowledgements
+```
+@inproceedings{kawar2022jpeg,
+      title={JPEG Artifact Correction using Denoising Diffusion Restoration Models}, 
+      author={Bahjat Kawar and Jiaming Song and Stefano Ermon and Michael Elad},
+      booktitle={Neural Information Processing Systems (NeurIPS) Workshop on Score-Based Methods},
+      year={2022}
+}
+```
+
+```
+@inproceedings{kawar2022denoising,
+      title={Denoising Diffusion Restoration Models},
+      author={Bahjat Kawar and Michael Elad and Stefano Ermon and Jiaming Song},
+      booktitle={Advances in Neural Information Processing Systems},
+      year={2022}
+}
+```
+
+This implementation is based on / inspired by [https://github.com/bahjat-kawar/ddrm](https://github.com/bahjat-kawar/ddrm) 
+
+## License
+
+The code is released under the MIT License.
diff --git a/assets/ddrm-jpeg-demo.png b/assets/ddrm-jpeg-demo.png
diff --git a/configs/bedroom.yml b/configs/bedroom.yml
@@ -0,0 +1,51 @@
+data:
+    dataset: "LSUN"
+    category: "bedroom"
+    image_size: 256
+    channels: 3
+    logit_transform: false
+    uniform_dequantization: false
+    gaussian_dequantization: false
+    random_flip: true
+    rescaled: true
+    num_workers: 32
+    out_of_dist: true
+
+model:
+    type: "simple"
+    in_channels: 3
+    out_ch: 3
+    ch: 128
+    ch_mult: [1, 1, 2, 2, 4, 4]
+    num_res_blocks: 2
+    attn_resolutions: [16, ]
+    dropout: 0.0
+    var_type: fixedsmall
+    ema_rate: 0.999
+    ema: True
+    resamp_with_conv: True
+
+diffusion:
+    beta_schedule: linear
+    beta_start: 0.0001
+    beta_end: 0.02
+    num_diffusion_timesteps: 1000
+
+training:
+    batch_size: 64
+    n_epochs: 10000
+    n_iters: 5000000
+    snapshot_freq: 5000
+    validation_freq: 2000
+
+sampling:
+    batch_size: 6
+    last_only: True
+
+optim:
+    weight_decay: 0.000
+    optimizer: "Adam"
+    lr: 0.00002
+    beta1: 0.9
+    amsgrad: false
+    eps: 0.00000001
diff --git a/configs/cat.yml b/configs/cat.yml
@@ -0,0 +1,51 @@
+data:
+    dataset: "LSUN"
+    category: "cat"
+    image_size: 256
+    channels: 3
+    logit_transform: false
+    uniform_dequantization: false
+    gaussian_dequantization: false
+    random_flip: true
+    rescaled: true
+    num_workers: 32
+    out_of_dist: false
+
+model:
+    type: "simple"
+    in_channels: 3
+    out_ch: 3
+    ch: 128
+    ch_mult: [1, 1, 2, 2, 4, 4]
+    num_res_blocks: 2
+    attn_resolutions: [16, ]
+    dropout: 0.0
+    var_type: fixedsmall
+    ema_rate: 0.999
+    ema: True
+    resamp_with_conv: True
+
+diffusion:
+    beta_schedule: linear
+    beta_start: 0.0001
+    beta_end: 0.02
+    num_diffusion_timesteps: 1000
+
+training:
+    batch_size: 64
+    n_epochs: 10000
+    n_iters: 5000000
+    snapshot_freq: 5000
+    validation_freq: 2000
+
+sampling:
+    batch_size: 32
+    last_only: True
+
+optim:
+    weight_decay: 0.000
+    optimizer: "Adam"
+    lr: 0.00002
+    beta1: 0.9
+    amsgrad: false
+    eps: 0.00000001
diff --git a/configs/celeba_hq.yml b/configs/celeba_hq.yml
@@ -0,0 +1,36 @@
+data:
+    dataset: "CelebA_HQ"
+    category: ""
+    image_size: 256
+    channels: 3
+    logit_transform: false
+    uniform_dequantization: false
+    gaussian_dequantization: false
+    random_flip: true
+    rescaled: true
+    num_workers: 32
+    out_of_dist: True
+
+model:
+    type: "simple"
+    in_channels: 3
+    out_ch: 3
+    ch: 128
+    ch_mult: [1, 1, 2, 2, 4, 4]
+    num_res_blocks: 2
+    attn_resolutions: [16, ]
+    dropout: 0.0
+    var_type: fixedsmall
+    ema_rate: 0.999
+    ema: True
+    resamp_with_conv: True
+
+diffusion:
+    beta_schedule: linear
+    beta_start: 0.0001
+    beta_end: 0.02
+    num_diffusion_timesteps: 1000
+
+sampling:
+    batch_size: 4
+    last_only: True
diff --git a/configs/church.yml b/configs/church.yml
@@ -0,0 +1,51 @@
+data:
+    dataset: "LSUN"
+    category: "church_outdoor"
+    image_size: 256
+    channels: 3
+    logit_transform: false
+    uniform_dequantization: false
+    gaussian_dequantization: false
+    random_flip: true
+    rescaled: true
+    num_workers: 32
+    out_of_dist: true
+
+model:
+    type: "simple"
+    in_channels: 3
+    out_ch: 3
+    ch: 128
+    ch_mult: [1, 1, 2, 2, 4, 4]
+    num_res_blocks: 2
+    attn_resolutions: [16, ]
+    dropout: 0.0
+    var_type: fixedsmall
+    ema_rate: 0.999
+    ema: True
+    resamp_with_conv: True
+
+diffusion:
+    beta_schedule: linear
+    beta_start: 0.0001
+    beta_end: 0.02
+    num_diffusion_timesteps: 1000
+
+training:
+    batch_size: 64
+    n_epochs: 10000
+    n_iters: 5000000
+    snapshot_freq: 5000
+    validation_freq: 2000
+
+sampling:
+    batch_size: 6
+    last_only: True
+
+optim:
+    weight_decay: 0.000
+    optimizer: "Adam"
+    lr: 0.00002
+    beta1: 0.9
+    amsgrad: false
+    eps: 0.00000001
diff --git a/configs/imagenet_256.yml b/configs/imagenet_256.yml
@@ -0,0 +1,43 @@
+data:
+    dataset: "ImageNet"
+    image_size: 256
+    channels: 3
+    logit_transform: false
+    uniform_dequantization: false
+    gaussian_dequantization: false
+    random_flip: true
+    rescaled: true
+    num_workers: 32
+    subset_1k: True
+    out_of_dist: False
+
+model:
+    type: "openai"
+    in_channels: 3
+    out_channels: 3
+    num_channels: 256
+    num_heads: 4
+    num_res_blocks: 2
+    attention_resolutions: "32,16,8"
+    dropout: 0.0
+    resamp_with_conv: True
+    learn_sigma: True
+    use_scale_shift_norm: true
+    use_fp16: true
+    resblock_updown: true
+    num_heads_upsample: -1
+    var_type: 'fixedsmall'
+    num_head_channels: 64
+    image_size: 256
+    class_cond: false
+    use_new_attention_order: false
+
+diffusion:
+    beta_schedule: linear
+    beta_start: 0.0001
+    beta_end: 0.02
+    num_diffusion_timesteps: 1000
+
+sampling:
+    batch_size: 8
+    last_only: True