├── .gitignore ├── Demo_hr ├── Real_09_bg.jpg └── Real_09_fg.png ├── LICENSE ├── PIH_train.py ├── README.md ├── __init__.py ├── dataset.py ├── demo.py ├── demo_data └── train │ ├── bg │ ├── 00057ce6-67c5-411e-8087-3799b638518a_after_mask.png │ └── 00057ce6-67c5-411e-8087-3799b638518a_before_mask.png │ ├── masks │ ├── 00057ce6-67c5-411e-8087-3799b638518a_after_mask.png │ └── 00057ce6-67c5-411e-8087-3799b638518a_before_mask.png │ └── real_images │ ├── 00057ce6-67c5-411e-8087-3799b638518a_after.png │ └── 00057ce6-67c5-411e-8087-3799b638518a_before.png ├── demo_light.py ├── environment.yml ├── github_images ├── Figure5.png ├── Figure_3.png ├── Figure_5_final.png ├── Figure_8_final.png ├── Figure_teaser.png ├── S1.png └── demo.gif ├── inference.py ├── inference_scripts ├── Inference.sh ├── Inference_Composite.sh ├── Inference_Composite_masking.sh ├── Inference_Composite_masking_3.sh ├── Inference_Composite_masking_3_depth.sh ├── Inference_Composite_masking_3_highres.sh ├── Inference_Composite_masking_3_noweb.sh ├── Inference_Composite_masking_highres.sh ├── Inference_Composite_masking_pixel.sh ├── Inference_Composite_unet.sh └── Inference_iHarmony.sh ├── model.py ├── pretrained └── pretrained.placeholder ├── results └── results.placeholder ├── scripts ├── installation.sh └── train_example.sh └── utils ├── efficientnet_v2.py ├── mobilenet_v3.py ├── modules.py ├── networks.py ├── resnet.py ├── resnet_ibn.py ├── unet ├── __init__.py ├── unet_model.py └── unet_parts.py └── unet_dis.py /.gitignore: -------------------------------------------------------------------------------- 1 | *.pyc 2 | PIH_ResNet/results/ 3 | .DS_Store 4 | __pycache__ 5 | .python-version 6 | -------------------------------------------------------------------------------- /Demo_hr/Real_09_bg.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/Demo_hr/Real_09_bg.jpg -------------------------------------------------------------------------------- /Demo_hr/Real_09_fg.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/Demo_hr/Real_09_fg.png -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright 2023 Adobe Research 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Parametric Image Harmonization (PIH) 2 | [Project Page](http://people.eecs.berkeley.edu/~kewang/sprih/) | [Paper](https://arxiv.org/abs/2303.00157) | [Bibtex](https://people.eecs.berkeley.edu/~kewang/sprih/sprih.txt) 3 | 4 | Semi-supervised Parametric Real-world Image Harmonization.\ 5 | _CVPR 2023_ \ 6 | [Ke Wang](people.eecs.berkeley.edu/~kewang), [Michaël Gharbi](http://mgharbi.com/), [He Zhang](https://sites.google.com/site/hezhangsprinter/), [Zhihao Xia](https://likesum.github.io/), [Eli Shechtman](https://research.adobe.com/person/eli-shechtman/) 7 | 8 | 9 | 10 | 11 | **A novel semi-supervised training strategy and the first harmonization method that learns complex local appearance harmonization from unpaired real composites.** 12 | 13 | The code was developed by [Ke Wang](people.eecs.berkeley.edu/~kewang) when Ke was a research scientist intern at Adobe research. 14 | 15 | Please contact Ke (kewang@berkeley.edu) or Michaël (mgharbi@adobe.com) if you have any question. 16 | 17 | **Results** 18 | 19 | 20 | Our results show better visual agreements with the ground truth compared to SOTA methods in terms of color harmonization 21 | (rows 1,2 and 4) and shading correction (row 3). 22 | 23 | 24 | 25 | RGB curves harmonize the global color/tone (center), while our shading map corrects the local shading in the harmonization output (right). 26 | 27 | 28 | --- 29 | 30 | ## Prerequisites 31 | 32 | - Linux 33 | - Python 3 34 | - NVIDIA GPU + CUDA CuDNN 35 | - [Conda](https://docs.conda.io/en/latest/) installed 36 | 37 | 38 | --- 39 | 40 | **Table of Contents:**
41 | 1. [Setup](#setup) - set up the enviroment
42 | 2. [Pretrained Models](#setup) - download pretrained models and resources
43 | 3. [Interactive Demo](#demo) - off-line interactive demo
44 | 4. [Inference](#inference) - inference on high-resolution images with pretrained model
45 | 5. [Dataset](#dataset) - prepare your own dataset for the training
46 | 6. [Training](#training) - pipeline for training PIH
47 | 7. [Citation](#citation) - bibtex citation
48 | 49 | 50 | 51 | --- 52 | 53 | ## Setup 54 | 55 | - Clone this repo: 56 | 57 | ```bash 58 | git clone git@github.com:adobe/PIH.git 59 | ``` 60 | 61 | - Install dependencies 62 | 63 | We create a `environment.yml` to install the dependencies, you need to have [Conda](https://docs.conda.io/en/latest/) installed. Run 64 | 65 | ``` 66 | conda env create -f environment.yml 67 | ``` 68 | (essentially install [PyTorch](https://pytorch.org/)) 69 | 70 | --- 71 | 72 | ## Pretrained models 73 | 74 | We provide our pre-trained model (93M parameters) on *Artist Retouched Dataset* from this [link](https://drive.google.com/file/d/1seW8qSnaBOQ4_S9bQ4ThVOdeJGYJ-f74/view?usp=sharing) and put it in the folder. 75 | 76 | ``` 77 | ./pretrained/ 78 | ``` 79 | 80 | --- 81 | 82 | ## Demo 83 | 84 | We provide an interactive demo host offline built with [PyGame](https://www.pygame.org/news) 85 | 86 | First, we install the dependencies: 87 | 88 | ``` 89 | python -m pip install -U pygame --user 90 | pip install pygame_gui 91 | pip install timm 92 | ``` 93 | 94 | Then, simpy run the following command to start the demo: 95 | ``` 96 | python demo.py 97 | ``` 98 | 99 | Here we provide a tutorial video for the demo. 100 | 101 | 102 | 103 | --- 104 | 105 | ## Inference 106 | 107 | We provide the inference code for evaluations: 108 | 109 | ``` 110 | python inference.py --bg --fg --checkpoints [--gpu] 111 | ``` 112 | 113 | notes: 114 | - arguments `--gpu` enable inference on GPU using cuda, default is by using CPU. 115 | - arguments `--checkpoints` specifies the dir for the checkpoint. 116 | 117 | 118 | Example: 119 | ``` 120 | python inference.py --bg Demo_hr/Real_09_bg.jpg --fg Demo_hr/Real_09_fg.png --checkpoints pretrained/ckpt_g39.pth --gpu 121 | ``` 122 | 123 | Check the `results/` folder for output images. 124 | 125 | --- 126 | 127 | ## Dataset 128 | 129 | 130 | 131 | We prepare a guidline of preparing *Artist Retouched Dataset*. 132 | 133 | For image with name ``, we organize the `data` directory like this: 134 | 135 | ``` 136 | data 137 | |--train 138 | |--bg 139 | |-- _before.png 140 | |-- _after.png 141 | |--masks 142 | |-- _before.png 143 | |-- _after.png 144 | |--real_images 145 | |-- _before.png 146 | |-- _after.png 147 | 148 | |--test 149 | |--bg 150 | |-- _before.png 151 | |-- _after.png 152 | |--masks 153 | |-- _before.png 154 | |-- _after.png 155 | |--real_images 156 | |-- _before.png 157 | |-- _after.png 158 | ``` 159 | 160 | notes: 161 | - bg (background): Inpainted background using foreground masks. Here we use [LAMA](https://github.com/advimman/lama) to perform inpainting. 162 | - masks: Foreground masks, should be consistent between `Before`, and `After`. 163 | - real_images: Ground truth real images. 164 | 165 | --- 166 | 167 | ## Training 168 | 169 | Our approach uses a dual-stream semi-supervised training to bridge the domain gap, alleviating the generalization issues that plague many state-of-the-art harmonization models 170 | 171 | 172 | 173 | 174 | We provide the script `train_example.sh` to perform training. 175 | 176 | Training notes: 177 | - modify `--dir_data` to the path of your custom dataset. 178 | - arguments `recon_weight` correspons to the weighting parameter to balance stream 1 and stream 2. 179 | 180 | Simply run: 181 | ``` 182 | bash scripts/train_example.sh 183 | ``` 184 | to start the training. 185 | 186 | --- 187 | 188 | ## Citation 189 | If you use this code for your research, please cite our paper. 190 | 191 | ``` 192 | @article{wang2023semi, 193 | title={Semi-supervised Parametric Real-world Image Harmonization}, 194 | author={Wang, Ke and Gharbi, Micha{\"e}l and Zhang, He and Xia, Zhihao and Shechtman, Eli}, 195 | booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, 196 | year = {2023} 197 | } 198 | 199 | -------------------------------------------------------------------------------- /__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. -------------------------------------------------------------------------------- /dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | 11 | 12 | from glob import glob 13 | import os 14 | import numpy as np 15 | import torch 16 | 17 | from PIL import Image 18 | 19 | import torchvision.transforms as T 20 | import torchvision.transforms.functional as F 21 | 22 | 23 | from torch.utils.data import Dataset 24 | import random 25 | import sys 26 | 27 | 28 | class PIHData(Dataset): 29 | def __init__(self, data_directory, device=torch.device("cpu")): 30 | """ 31 | 32 | Parameters 33 | ---------- 34 | data_directory : str 35 | The directory containing the training image data. 36 | max_offset : tuple 37 | The maximum offset to crop an image to. 38 | magnitude : bool 39 | If True, train using magnitude image as input. Otherwise, use real and imaginary image in separate channels. 40 | device : torch.device 41 | The device to load the data to. 42 | complex : bool 43 | If True, return images as complex data. Otherwise check for magnitude return or for real and imaginary 44 | channels. This is needed when training, since post processing is done in the model (adds phase augmentation 45 | and converts to magnitude or channels). Magnitude and channels are implemented for evaluation. 46 | """ 47 | 48 | self.image_paths = glob(f"{data_directory}/*_gt.jpg") 49 | print( 50 | f"Using data from: {data_directory}\nFound {len(self.image_paths)} image paths." 51 | ) 52 | self.device = device 53 | self.transforms = T.Compose([T.ToTensor()]) 54 | self.transforms_mask = T.Compose([T.Grayscale(), T.ToTensor()]) 55 | 56 | def __len__(self): 57 | return len(self.image_paths) 58 | 59 | def __getitem__(self, index): 60 | """Get image at the specified index. 61 | 62 | Parameters 63 | ---------- 64 | index : int 65 | The image index. 66 | 67 | Returns 68 | ------- 69 | patch: torch.Tensor 70 | 71 | """ 72 | 73 | image_path = self.image_paths[index] 74 | ground_truth = Image.open(image_path) 75 | input_image = Image.open(image_path[: image_path.rindex("_")] + ".jpg") 76 | input_mask = Image.open(image_path[: image_path.rindex("_")] + "_mask.jpg") 77 | 78 | # original_image = np.load(self.image_paths[index])[None].astype(np.complex64) 79 | 80 | return ( 81 | self.transforms(input_image), 82 | self.transforms_mask(input_mask), 83 | self.transforms(ground_truth), 84 | image_path, 85 | ) 86 | 87 | 88 | class PIHDataRandom(Dataset): 89 | def __init__(self, data_directory, device=torch.device("cpu")): 90 | """ 91 | 92 | Parameters 93 | ---------- 94 | data_directory : str 95 | The directory containing the training image data. 96 | max_offset : tuple 97 | The maximum offset to crop an image to. 98 | magnitude : bool 99 | If True, train using magnitude image as input. Otherwise, use real and imaginary image in separate channels. 100 | device : torch.device 101 | The device to load the data to. 102 | complex : bool 103 | If True, return images as complex data. Otherwise check for magnitude return or for real and imaginary 104 | channels. This is needed when training, since post processing is done in the model (adds phase augmentation 105 | and converts to magnitude or channels). Magnitude and channels are implemented for evaluation. 106 | """ 107 | 108 | self.image_paths = glob(f"{data_directory}/*_gt.jpg") 109 | print( 110 | f"Using data from: {data_directory}\nFound {len(self.image_paths)} image paths." 111 | ) 112 | self.device = device 113 | self.transforms = T.Compose([T.ToTensor()]) 114 | self.transforms_mask = T.Compose([T.Grayscale(), T.ToTensor()]) 115 | 116 | def __len__(self): 117 | return len(self.image_paths) 118 | 119 | def __getitem__(self, index): 120 | """Get image at the specified index. 121 | 122 | Parameters 123 | ---------- 124 | index : int 125 | The image index. 126 | 127 | Returns 128 | ------- 129 | patch: torch.Tensor 130 | 131 | """ 132 | 133 | image_path = self.image_paths[index] 134 | 135 | ground_truth = self.transforms(Image.open(image_path)) 136 | mask_torch = self.transforms( 137 | Image.open(image_path[: image_path.rindex("_")] + "_mask.jpg") 138 | ) 139 | 140 | imag_torch = T.functional.adjust_contrast( 141 | ground_truth, (np.random.rand() * 0.4 + 0.8) 142 | ) 143 | 144 | imag_torch = T.functional.adjust_brightness( 145 | imag_torch, (np.random.rand() * 0.4 + 0.8) 146 | ) 147 | 148 | imag_torch = T.functional.adjust_saturation( 149 | imag_torch, (np.random.rand() * 0.4 + 0.8) 150 | ) 151 | 152 | # Read functions for color transform: Cross - chaneel - YCC 153 | imag_torch[0, ...] = ( 154 | imag_torch[0, ...] * (np.random.rand() * 0.3 + 0.70) 155 | + imag_torch[0, ...] * imag_torch[0, ...] * (np.random.rand() - 0.5) * 0.1 156 | + imag_torch[0, ...] 157 | * imag_torch[0, ...] 158 | * imag_torch[0, ...] 159 | * (np.random.rand() - 0.5) 160 | * 0.05 161 | ) 162 | 163 | imag_torch[1, ...] = ( 164 | imag_torch[1, ...] * (np.random.rand() * 0.3 + 0.70) 165 | + +imag_torch[1, ...] * imag_torch[1, ...] * (np.random.rand() - 0.5) * 0.1 166 | + imag_torch[1, ...] 167 | * imag_torch[1, ...] 168 | * imag_torch[1, ...] 169 | * (np.random.rand() - 0.5) 170 | * 0.05 171 | ) 172 | 173 | imag_torch[2, ...] = ( 174 | imag_torch[2, ...] * (np.random.rand() * 0.3 + 0.70) 175 | + imag_torch[2, ...] * imag_torch[2, ...] * (np.random.rand() - 0.5) * 0.1 176 | + imag_torch[2, ...] 177 | * imag_torch[2, ...] 178 | * imag_torch[2, ...] 179 | * (np.random.rand() - 0.5) 180 | * 0.05 181 | ) 182 | 183 | imag_composite = ground_truth * (1 - mask_torch) + imag_torch * mask_torch 184 | # original_image = np.load(self.image_paths[index])[None].astype(np.complex64) 185 | 186 | return ( 187 | imag_composite, 188 | mask_torch, 189 | ground_truth, 190 | image_path, 191 | ) 192 | 193 | 194 | class PIHDataNGT(Dataset): 195 | def __init__(self, data_directory, device=torch.device("cpu")): 196 | """ 197 | 198 | Parameters 199 | ---------- 200 | data_directory : str 201 | The directory containing the training image data. 202 | max_offset : tuple 203 | The maximum offset to crop an image to. 204 | magnitude : bool 205 | If True, train using magnitude image as input. Otherwise, use real and imaginary image in separate channels. 206 | device : torch.device 207 | The device to load the data to. 208 | complex : bool 209 | If True, return images as complex data. Otherwise check for magnitude return or for real and imaginary 210 | channels. This is needed when training, since post processing is done in the model (adds phase augmentation 211 | and converts to magnitude or channels). Magnitude and channels are implemented for evaluation. 212 | """ 213 | 214 | self.image_paths = glob(f"{data_directory}/*_mask.jpg") 215 | print( 216 | f"Using data from: {data_directory}\nFound {len(self.image_paths)} image paths." 217 | ) 218 | self.device = device 219 | self.transforms = T.Compose([T.ToTensor()]) 220 | self.transforms_mask = T.Compose([T.Grayscale(), T.ToTensor()]) 221 | 222 | def __len__(self): 223 | return len(self.image_paths) 224 | 225 | def __getitem__(self, index): 226 | """Get image at the specified index. 227 | 228 | Parameters 229 | ---------- 230 | index : int 231 | The image index. 232 | 233 | Returns 234 | ------- 235 | patch: torch.Tensor 236 | 237 | """ 238 | 239 | image_path = self.image_paths[index] 240 | # ground_truth = Image.open(image_path) 241 | input_image = Image.open(image_path[: image_path.rindex("_")] + ".jpg") 242 | input_mask = Image.open(image_path[: image_path.rindex("_")] + "_mask.jpg") 243 | 244 | # original_image = np.load(self.image_paths[index])[None].astype(np.complex64) 245 | 246 | return ( 247 | self.transforms(input_image), 248 | self.transforms_mask(input_mask), 249 | self.transforms_mask(input_mask), 250 | image_path, 251 | ) 252 | 253 | 254 | class IhdDataset(Dataset): 255 | def __init__(self, opt): 256 | self.image_paths = [] 257 | self.isTrain = opt.train 258 | if opt.train == True: 259 | print("loading training file") 260 | self.trainfile = opt.datadir + "IHD_train.txt" 261 | with open(self.trainfile, "r") as f: 262 | for line in f.readlines(): 263 | self.image_paths.append( 264 | os.path.join(opt.datadir, "", line.rstrip()) 265 | ) 266 | else: 267 | print("loading test file") 268 | self.trainfile = opt.datadir + "IHD_test.txt" 269 | with open(self.trainfile, "r") as f: 270 | for line in f.readlines(): 271 | self.image_paths.append( 272 | os.path.join(opt.datadir, "", line.rstrip()) 273 | ) 274 | self.transforms = T.Compose([T.ToTensor()]) 275 | self.transforms_mask = T.Compose([T.Grayscale(), T.ToTensor()]) 276 | self.image_size = 512 277 | 278 | print( 279 | f"Using data from: {opt.datadir}\nFound {len(self.image_paths)} image paths." 280 | ) 281 | 282 | def __getitem__(self, index): 283 | 284 | path = self.image_paths[index] 285 | name_parts = path.split("_") 286 | mask_path = self.image_paths[index].replace("composite_images", "masks") 287 | mask_path = mask_path.replace(("_" + name_parts[-1]), ".png") 288 | target_path = self.image_paths[index].replace("composite_images", "real_images") 289 | target_path = target_path.replace( 290 | ("_" + name_parts[-2] + "_" + name_parts[-1]), ".jpg" 291 | ) 292 | 293 | comp = Image.open(path).convert("RGB") 294 | real = Image.open(target_path).convert("RGB") 295 | mask = Image.open(mask_path).convert("RGB") 296 | 297 | if np.random.rand() > 0.5 and self.isTrain: 298 | comp, mask, real = F.hflip(comp), F.hflip(mask), F.hflip(real) 299 | 300 | if not (comp.size[0] == self.image_size and comp.size[1] == self.image_size): 301 | # assert 0 302 | comp = F.resize(comp, [self.image_size, self.image_size]) 303 | mask = F.resize(mask, [self.image_size, self.image_size]) 304 | real = F.resize(real, [self.image_size, self.image_size]) 305 | 306 | comp = self.transforms(comp) 307 | mask = self.transforms_mask(mask) 308 | 309 | real = self.transforms(real) 310 | 311 | return (comp, mask, real, path) 312 | 313 | def __len__(self): 314 | """Return the total number of images.""" 315 | return len(self.image_paths) 316 | 317 | 318 | class DataCompositeGAN(Dataset): 319 | def __init__(self, data_directory, ratio=1, augment=False, colorjitter=True, lowres=False,return_raw=False, ratio_constrain=False): 320 | """ 321 | 322 | Parameters 323 | ---------- 324 | data_directory : str 325 | The directory containing the training image data. 326 | """ 327 | self.lowres = lowres 328 | self.image_paths = glob(f"{data_directory}/masks/*_mask.png") 329 | 330 | self.image_paths = self.image_paths[0 : int(len(self.image_paths) * ratio)] 331 | 332 | self.length = len(self.image_paths) 333 | print( 334 | f"Using data from: {data_directory}\nFound {len(self.image_paths)} image paths." 335 | ) 336 | self.transforms = T.Compose([T.ToTensor()]) 337 | self.transforms_mask = T.Compose([T.Grayscale(), T.ToTensor()]) 338 | self.colorjitter = colorjitter 339 | if self.colorjitter: 340 | self.transform_color = T.ColorJitter( 341 | brightness=[0.65, 1.35], contrast=0.2, saturation=0, hue=0 342 | ) ## 0.3 0.7 343 | self.augment = augment 344 | self.returnraw = return_raw 345 | self.ratio_constrain = ratio_constrain 346 | if ratio_constrain: 347 | print("Using Constrained Ratio") 348 | 349 | def __len__(self): 350 | return len(self.image_paths) 351 | 352 | def __getitem__(self, index): 353 | """Get image at the specified index. 354 | 355 | Parameters 356 | ---------- 357 | index : int 358 | The image index. 359 | 360 | Returns 361 | ------- 362 | patch: torch.Tensor 363 | 364 | Foreground 365 | """ 366 | 367 | path_bg = self.image_paths[index] # ForeGround 368 | 369 | path_fg = self.image_paths[np.random.randint(0, self.length)] 370 | 371 | ### fore-ground image loading 372 | 373 | path_fg_image = path_fg.replace("masks/", "real_images/") 374 | path_fg_image = path_fg_image.replace("_mask.png", ".png") 375 | 376 | path_fg_bg = path_fg.replace("masks/", "bg/") 377 | 378 | mask_fg = Image.open(path_fg) 379 | 380 | image_fg = Image.open(path_fg_image) 381 | 382 | image_fg_bg = Image.open(path_fg_bg) 383 | 384 | 385 | 386 | ### back-ground image loading 387 | 388 | path_bg_image = path_bg.replace("masks/", "real_images/") 389 | path_bg_image = path_bg_image.replace("_mask.png", ".png") 390 | 391 | path_bg_bg = path_bg.replace("masks/", "bg/") 392 | 393 | mask_bg = Image.open(path_bg) 394 | 395 | image_bg = Image.open(path_bg_image) 396 | 397 | 398 | 399 | image_bg_augment = image_bg 400 | 401 | if self.augment: 402 | if "before" in path_bg_image: 403 | path_bg_image_augment = path_bg_image.replace("before", "after") 404 | elif "after" in path_bg_image: 405 | path_bg_image_augment = path_bg_image.replace("after", "before") 406 | 407 | 408 | # image_bg_augment = Image.open(path_bg.replace("masks/",'composite/')) 409 | image_bg_augment = Image.open(path_bg_image_augment) 410 | 411 | 412 | image_bg_bg = Image.open(path_bg_bg) 413 | if self.lowres: 414 | mask_fg = mask_fg.resize((256,256)) 415 | image_fg = image_fg.resize((256,256)) 416 | image_fg_bg = image_fg_bg.resize((256,256)) 417 | mask_bg = mask_bg.resize((256,256)) 418 | image_bg = image_bg.resize((256,256)) 419 | image_bg_augment = image_bg_augment.resize((256,256)) 420 | image_bg_bg = image_bg_bg.resize((256,256)) 421 | 422 | 423 | 424 | 425 | mask_bg_bbox = mask_bg.getbbox() 426 | mask_fg_bbox = mask_fg.getbbox() 427 | 428 | 429 | 430 | ## Target 431 | x_1_1, y_1_1, x_1_2, y_1_2 = mask_bg_bbox 432 | center_1_x = (x_1_1 + x_1_2) / 2 433 | center_1_y = (y_1_1 + y_1_2) / 2 434 | 435 | ## 436 | x_2_1, y_2_1, x_2_2, y_2_2 = mask_fg_bbox 437 | ration_x = (x_1_2 - x_1_1) / (x_2_2 - x_2_1) if x_2_2 != x_2_1 else 1 438 | ration_y = (y_1_2 - y_1_1) / (y_2_2 - y_2_1) if y_2_2 != y_2_1 else 1 439 | 440 | ## Scaling 441 | 442 | if not self.ratio_constrain: 443 | 444 | mask_fg_aff = F.affine( 445 | mask_fg, angle=0, translate=[0, 0], scale=min(ration_y, ration_x), shear=0 446 | ) 447 | image_fg_aff = F.affine( 448 | image_fg, angle=0, translate=[0, 0], scale=min(ration_y, ration_x), shear=0 449 | ) 450 | else: 451 | 452 | length_box = max(y_1_2-y_1_1,x_1_2-x_1_1) 453 | 454 | if length_box < 100: 455 | ration_x = (100) / (x_2_2 - x_2_1) if x_2_2 != x_2_1 else 1 456 | ration_y = (100) / (y_2_2 - y_2_1) if y_2_2 != y_2_1 else 1 457 | 458 | mask_fg_aff = F.affine( 459 | mask_fg, angle=0, translate=[0, 0], scale=min(ration_y, ration_x), shear=0) 460 | image_fg_aff = F.affine( 461 | image_fg, angle=0, translate=[0, 0], scale=min(ration_y, ration_x), shear=0 462 | ) 463 | 464 | 465 | 466 | if mask_fg_aff.getbbox() == None: 467 | mask_fg_aff = F.affine(mask_fg, angle=0, translate=[0, 0], scale=1, shear=0) 468 | 469 | x_2_1_a, y_2_1_a, x_2_2_a, y_2_2_a = mask_fg_aff.getbbox() 470 | center_2_x_a = (x_2_1_a + x_2_2_a) / 2 471 | center_2_y_a = (y_2_1_a + y_2_2_a) / 2 472 | 473 | shift_fg_x = np.random.randint(-10, 10) 474 | shift_fg_y = np.random.randint(-10, 10) 475 | 476 | mask_fg_aff_all = F.affine( 477 | mask_fg_aff, 478 | angle=0, 479 | translate=[ 480 | center_1_x - center_2_x_a + shift_fg_x, 481 | center_1_y - center_2_y_a + shift_fg_y, 482 | ], 483 | scale=1, 484 | shear=0, 485 | ) 486 | image_fg_aff_all = F.affine( 487 | image_fg_aff, 488 | angle=0, 489 | translate=[ 490 | center_1_x - center_2_x_a + shift_fg_x, 491 | center_1_y - center_2_y_a + shift_fg_y, 492 | ], 493 | scale=1, 494 | shear=0, 495 | ) 496 | 497 | if self.colorjitter: 498 | if np.random.rand() < 1: 499 | # print("i love you one") 500 | image_fg_aff_all = self.transform_color(image_fg_aff_all) 501 | 502 | im_composite = Image.composite(image_fg_aff_all, image_bg_bg, mask_fg_aff_all) 503 | 504 | ## What we want to output? Background, im_composite, mask_fg_aff_all, real_image 505 | 506 | if self.returnraw: 507 | if self.colorjitter: 508 | if np.random.rand() < 1: 509 | # print("i love you two") 510 | 511 | image_bg_augment_f = self.transform_color(image_bg_augment) 512 | image_bg_augment = Image.composite(image_bg_augment_f, image_bg, mask_bg) 513 | else: 514 | image_bg_augment = Image.composite(image_bg_augment, image_bg, mask_bg) 515 | 516 | 517 | return ( 518 | self.transforms(image_bg_bg), 519 | self.transforms(im_composite), 520 | self.transforms_mask(mask_fg_aff_all), 521 | self.transforms(image_bg), 522 | self.transforms_mask(mask_bg), 523 | self.transforms(image_bg_augment), 524 | path_fg, 525 | path_bg, 526 | ) 527 | 528 | else: 529 | 530 | shift_bg_x = np.random.randint(-10, 10) 531 | shift_bg_y = np.random.randint(-10, 10) 532 | 533 | mask_bg_shift = F.affine( 534 | mask_bg, 535 | angle=0, 536 | translate=[ 537 | shift_bg_x, 538 | shift_bg_y, 539 | ], 540 | scale=1, 541 | shear=0, 542 | ) 543 | 544 | image_bg_shift = F.affine( 545 | image_bg, 546 | angle=0, 547 | translate=[ 548 | shift_bg_x, 549 | shift_bg_y, 550 | ], 551 | scale=1, 552 | shear=0, 553 | ) 554 | 555 | image_bg_augment_shift = F.affine( 556 | image_bg_augment, 557 | angle=0, 558 | translate=[ 559 | shift_bg_x, 560 | shift_bg_y, 561 | ], 562 | scale=1, 563 | shear=0, 564 | ) 565 | 566 | im_real = Image.composite(image_bg_shift, image_bg_bg, mask_bg_shift) 567 | 568 | if self.colorjitter: 569 | if np.random.rand() < 1: 570 | # print("i love you two") 571 | 572 | image_bg_augment_shift = self.transform_color(image_bg_augment_shift) 573 | 574 | im_real_augment = Image.composite( 575 | image_bg_augment_shift, image_bg_bg, mask_bg_shift 576 | ) 577 | 578 | # Dataset output orders: 1. Background (inpainted) 2. Image Composite 3. Mask 4. Real Image 579 | return ( 580 | self.transforms(image_bg_bg), 581 | self.transforms(im_composite), 582 | self.transforms_mask(mask_fg_aff_all), 583 | self.transforms(im_real), 584 | self.transforms_mask(mask_bg_shift), 585 | self.transforms(im_real_augment), 586 | path_fg, 587 | path_bg, 588 | ) 589 | 590 | 591 | class PIHData_Composite(Dataset): 592 | def __init__(self, data_directory,lowres,original=False): 593 | """ 594 | 595 | Parameters 596 | ---------- 597 | data_directory : str 598 | The directory containing the training image data. 599 | max_offset : tuple 600 | The maximum offset to crop an image to. 601 | magnitude : bool 602 | If True, train using magnitude image as input. Otherwise, use real and imaginary image in separate channels. 603 | device : torch.device 604 | The device to load the data to. 605 | complex : bool 606 | If True, return images as complex data. Otherwise check for magnitude return or for real and imaginary 607 | channels. This is needed when training, since post processing is done in the model (adds phase augmentation 608 | and converts to magnitude or channels). Magnitude and channels are implemented for evaluation. 609 | """ 610 | 611 | self.image_paths = glob(f"{data_directory}/*_bg.jpg") 612 | print( 613 | f"Using data from: {data_directory}\nFound {len(self.image_paths)} image paths." 614 | ) 615 | self.transforms = T.Compose([T.ToTensor()]) 616 | self.transforms_mask = T.Compose([T.Grayscale(), T.ToTensor()]) 617 | self.lowres = lowres 618 | self.original = original 619 | if lowres: 620 | self.res = 256 621 | else: 622 | self.res = 512 623 | 624 | def __len__(self): 625 | return len(self.image_paths) 626 | 627 | def __getitem__(self, index): 628 | """Get image at the specified index. 629 | 630 | Parameters 631 | ---------- 632 | index : int 633 | The image index. 634 | 635 | Returns 636 | ------- 637 | patch: torch.Tensor 638 | 639 | """ 640 | 641 | image_path = self.image_paths[index] 642 | 643 | if self.original: 644 | input_bg = Image.open(image_path) 645 | 646 | input_composite = Image.open(image_path.replace("bg", "composite")) 647 | input_mask = Image.open(image_path.replace("bg", "mask")) 648 | if os.path.exists(image_path.replace("bg", "real")): 649 | input_real = Image.open(image_path.replace("bg", "real")) 650 | else: 651 | input_real = Image.open(image_path.replace("bg", "gt")) 652 | 653 | else: 654 | input_bg = Image.open(image_path).resize((self.res, self.res)) 655 | 656 | input_composite = Image.open(image_path.replace("bg", "composite")).resize( 657 | (self.res, self.res) 658 | ) 659 | input_mask = Image.open(image_path.replace("bg", "mask")).resize((self.res, self.res)) 660 | if os.path.exists(image_path.replace("bg", "real")): 661 | input_real = Image.open(image_path.replace("bg", "real")).resize((self.res, self.res)) 662 | else: 663 | input_real = Image.open(image_path.replace("bg", "gt")).resize((self.res, self.res)) 664 | 665 | # original_image = np.load(self.image_paths[index])[None].astype(np.complex64) 666 | 667 | return ( 668 | self.transforms(input_bg), 669 | self.transforms(input_composite), 670 | self.transforms_mask(input_mask), 671 | self.transforms(input_real), 672 | image_path, 673 | ) 674 | 675 | 676 | class DataCompositeGAN_iharmony(Dataset): 677 | def __init__( 678 | self, data_directory, ratio=1, augment=False, colorjitter=True, return_raw=False,lowres=False 679 | ): 680 | """ 681 | 682 | Parameters 683 | ---------- 684 | data_directory : str 685 | The directory containing the training image data. 686 | """ 687 | 688 | self.image_paths = glob(f"{data_directory}/masks/*_mask.png") 689 | 690 | self.image_paths = self.image_paths[0 : int(len(self.image_paths) * ratio)] 691 | 692 | self.length = len(self.image_paths) 693 | print( 694 | f"Using data from: {data_directory}\nFound {len(self.image_paths)} image paths." 695 | ) 696 | self.transforms = T.Compose([T.ToTensor()]) 697 | self.transforms_mask = T.Compose([T.Grayscale(), T.ToTensor()]) 698 | self.colorjitter = colorjitter 699 | self.lowres = lowres 700 | if self.colorjitter: 701 | self.transform_color = T.ColorJitter( 702 | brightness=0.3, contrast=0.1, saturation=0.0, hue=0.0 703 | ) 704 | self.augment = augment 705 | self.return_raw = return_raw 706 | 707 | def __len__(self): 708 | return len(self.image_paths) 709 | 710 | def __getitem__(self, index): 711 | """Get image at the specified index. 712 | 713 | Parameters 714 | ---------- 715 | index : int 716 | The image index. 717 | 718 | Returns 719 | ------- 720 | patch: torch.Tensor 721 | 722 | Foreground 723 | """ 724 | 725 | path_bg = self.image_paths[index] # ForeGround 726 | 727 | path_fg = self.image_paths[np.random.randint(0, self.length)] 728 | 729 | ### fore-ground image loading 730 | 731 | path_fg_image = path_fg.replace("masks/", "real_images/") 732 | path_fg_image = path_fg_image.replace("_mask.png", ".jpg") 733 | 734 | path_fg_bg = path_fg.replace("masks/", "bg/") 735 | 736 | mask_fg = Image.open(path_fg) 737 | 738 | image_fg = Image.open(path_fg_image) 739 | 740 | image_fg_bg = Image.open(path_fg_bg) 741 | 742 | ### back-ground image loading 743 | 744 | path_bg_image = path_bg.replace("masks/", "real_images/") 745 | path_bg_image = path_bg_image.replace("_mask.png", ".jpg") 746 | 747 | path_bg_bg = path_bg.replace("masks/", "bg/") 748 | 749 | mask_bg = Image.open(path_bg) 750 | 751 | image_bg = Image.open(path_bg_image) 752 | 753 | image_bg_augment = image_bg 754 | 755 | if self.augment: 756 | path_bg_image_augment = path_bg_image.replace( 757 | "real_images", "composite" 758 | ).replace(".jpg", "_composite.jpg") 759 | # if "before" in path_bg_image: 760 | # path_bg_image_augment = path_bg_image.replace("before", "after") 761 | # elif "after" in path_bg_image: 762 | # path_bg_image_augment = path_bg_image.replace("after", "before") 763 | image_bg_augment = Image.open(path_bg_image_augment) 764 | 765 | image_bg_bg = Image.open(path_bg_bg) 766 | 767 | 768 | if self.lowres: 769 | mask_fg = mask_fg.resize((256,256)) 770 | image_fg = image_fg.resize((256,256)) 771 | image_fg_bg = image_fg_bg.resize((256,256)) 772 | mask_bg = mask_bg.resize((256,256)) 773 | image_bg = image_bg.resize((256,256)) 774 | image_bg_augment = image_bg_augment.resize((256,256)) 775 | image_bg_bg = image_bg_bg.resize((256,256)) 776 | 777 | mask_bg_bbox = mask_bg.getbbox() 778 | mask_fg_bbox = mask_fg.getbbox() 779 | 780 | ## Target 781 | x_1_1, y_1_1, x_1_2, y_1_2 = mask_bg_bbox 782 | center_1_x = (x_1_1 + x_1_2) / 2 783 | center_1_y = (y_1_1 + y_1_2) / 2 784 | 785 | ## 786 | x_2_1, y_2_1, x_2_2, y_2_2 = mask_fg_bbox 787 | ration_x = (x_1_2 - x_1_1) / (x_2_2 - x_2_1) if x_2_2 != x_2_1 else 1 788 | ration_y = (y_1_2 - y_1_1) / (y_2_2 - y_2_1) if y_2_2 != y_2_1 else 1 789 | 790 | ## Scaling 791 | mask_fg_aff = F.affine( 792 | mask_fg, angle=0, translate=[0, 0], scale=min(ration_y, ration_x), shear=0 793 | ) 794 | image_fg_aff = F.affine( 795 | image_fg, angle=0, translate=[0, 0], scale=min(ration_y, ration_x), shear=0 796 | ) 797 | if mask_fg_aff.getbbox() == None: 798 | mask_fg_aff = F.affine(mask_fg, angle=0, translate=[0, 0], scale=1, shear=0) 799 | 800 | x_2_1_a, y_2_1_a, x_2_2_a, y_2_2_a = mask_fg_aff.getbbox() 801 | center_2_x_a = (x_2_1_a + x_2_2_a) / 2 802 | center_2_y_a = (y_2_1_a + y_2_2_a) / 2 803 | 804 | shift_fg_x = np.random.randint(-10, 10) 805 | shift_fg_y = np.random.randint(-10, 10) 806 | 807 | mask_fg_aff_all = F.affine( 808 | mask_fg_aff, 809 | angle=0, 810 | translate=[ 811 | center_1_x - center_2_x_a + shift_fg_x, 812 | center_1_y - center_2_y_a + shift_fg_y, 813 | ], 814 | scale=1, 815 | shear=0, 816 | ) 817 | image_fg_aff_all = F.affine( 818 | image_fg_aff, 819 | angle=0, 820 | translate=[ 821 | center_1_x - center_2_x_a + shift_fg_x, 822 | center_1_y - center_2_y_a + shift_fg_y, 823 | ], 824 | scale=1, 825 | shear=0, 826 | ) 827 | 828 | if self.colorjitter: 829 | if np.random.rand() < 1: 830 | # print("i love you one") 831 | image_fg_aff_all = self.transform_color(image_fg_aff_all) 832 | 833 | im_composite = Image.composite(image_fg_aff_all, image_bg_bg, mask_fg_aff_all) 834 | 835 | ## What we want to output? Background, im_composite, mask_fg_aff_all, real_image 836 | 837 | if self.return_raw: 838 | 839 | if self.colorjitter: 840 | if np.random.rand() < 1: 841 | # print("i love you two") 842 | 843 | image_bg_augment_f = self.transform_color(image_bg_augment) 844 | image_bg_augment = Image.composite(image_bg_augment_f, image_bg_augment, mask_bg) 845 | # else: 846 | # image_bg_augment = Image.composite(image_bg_augment, image_bg, mask_bg) 847 | 848 | 849 | 850 | return ( 851 | self.transforms(image_bg_bg), 852 | self.transforms(im_composite), 853 | self.transforms_mask(mask_fg_aff_all), 854 | self.transforms(image_bg), 855 | self.transforms_mask(mask_bg), 856 | self.transforms(image_bg_augment), 857 | path_fg, 858 | path_bg, 859 | ) 860 | else: 861 | 862 | shift_bg_x = np.random.randint(-10, 10) 863 | shift_bg_y = np.random.randint(-10, 10) 864 | 865 | mask_bg_shift = F.affine( 866 | mask_bg, 867 | angle=0, 868 | translate=[ 869 | shift_bg_x, 870 | shift_bg_y, 871 | ], 872 | scale=1, 873 | shear=0, 874 | ) 875 | 876 | image_bg_shift = F.affine( 877 | image_bg, 878 | angle=0, 879 | translate=[ 880 | shift_bg_x, 881 | shift_bg_y, 882 | ], 883 | scale=1, 884 | shear=0, 885 | ) 886 | 887 | image_bg_augment_shift = F.affine( 888 | image_bg_augment, 889 | angle=0, 890 | translate=[ 891 | shift_bg_x, 892 | shift_bg_y, 893 | ], 894 | scale=1, 895 | shear=0, 896 | ) 897 | 898 | im_real = Image.composite(image_bg_shift, image_bg_bg, mask_bg_shift) 899 | 900 | if self.colorjitter: 901 | if np.random.rand() < 1: 902 | # print("i love you two") 903 | 904 | image_bg_augment_shift = self.transform_color( 905 | image_bg_augment_shift 906 | ) 907 | 908 | im_real_augment = Image.composite( 909 | image_bg_augment_shift, image_bg_bg, mask_bg_shift 910 | ) 911 | 912 | # Dataset output orders: 1. Background (inpainted) 2. Image Composite 3. Mask 4. Real Image 913 | return ( 914 | self.transforms(image_bg_bg), 915 | self.transforms(im_composite), 916 | self.transforms_mask(mask_fg_aff_all), 917 | self.transforms(im_real), 918 | self.transforms_mask(mask_bg_shift), 919 | self.transforms(im_real_augment), 920 | path_fg, 921 | path_bg, 922 | ) 923 | -------------------------------------------------------------------------------- /demo_data/train/bg/00057ce6-67c5-411e-8087-3799b638518a_after_mask.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/demo_data/train/bg/00057ce6-67c5-411e-8087-3799b638518a_after_mask.png -------------------------------------------------------------------------------- /demo_data/train/bg/00057ce6-67c5-411e-8087-3799b638518a_before_mask.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/demo_data/train/bg/00057ce6-67c5-411e-8087-3799b638518a_before_mask.png -------------------------------------------------------------------------------- /demo_data/train/masks/00057ce6-67c5-411e-8087-3799b638518a_after_mask.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/demo_data/train/masks/00057ce6-67c5-411e-8087-3799b638518a_after_mask.png -------------------------------------------------------------------------------- /demo_data/train/masks/00057ce6-67c5-411e-8087-3799b638518a_before_mask.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/demo_data/train/masks/00057ce6-67c5-411e-8087-3799b638518a_before_mask.png -------------------------------------------------------------------------------- /demo_data/train/real_images/00057ce6-67c5-411e-8087-3799b638518a_after.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/demo_data/train/real_images/00057ce6-67c5-411e-8087-3799b638518a_after.png -------------------------------------------------------------------------------- /demo_data/train/real_images/00057ce6-67c5-411e-8087-3799b638518a_before.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/demo_data/train/real_images/00057ce6-67c5-411e-8087-3799b638518a_before.png -------------------------------------------------------------------------------- /environment.yml: -------------------------------------------------------------------------------- 1 | name: pytorch_pih 2 | channels: 3 | - pytorch 4 | - conda-forge 5 | dependencies: 6 | - python=3.8 7 | - cudatoolkit=11.3 8 | - pytorch=1.11.0 9 | - torchvision=0.12.0 10 | - cudnn 11 | - numpy 12 | - pandas 13 | - jupyter 14 | - pip 15 | - tqdm 16 | - ipython 17 | - pillow 18 | - pip: 19 | - matplotlib 20 | - opencv-python 21 | 22 | 23 | -------------------------------------------------------------------------------- /github_images/Figure5.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/github_images/Figure5.png -------------------------------------------------------------------------------- /github_images/Figure_3.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/github_images/Figure_3.png -------------------------------------------------------------------------------- /github_images/Figure_5_final.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/github_images/Figure_5_final.png -------------------------------------------------------------------------------- /github_images/Figure_8_final.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/github_images/Figure_8_final.png -------------------------------------------------------------------------------- /github_images/Figure_teaser.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/github_images/Figure_teaser.png -------------------------------------------------------------------------------- /github_images/S1.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/github_images/S1.png -------------------------------------------------------------------------------- /github_images/demo.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/github_images/demo.gif -------------------------------------------------------------------------------- /inference.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | 11 | import matplotlib.pyplot as plt 12 | import torch 13 | import numpy as np 14 | import torchvision.transforms as T 15 | import torchvision.transforms.functional as F 16 | import PIL 17 | from PIL import Image 18 | import cv2 19 | from model import Model_Composite_PL, Model_Composite 20 | from optparse import OptionParser 21 | import os 22 | import time 23 | 24 | 25 | transforms_mask = T.Compose([T.Grayscale(), T.ToTensor()]) 26 | transform = T.Compose([T.ToTensor()]) 27 | resize = T.Resize((512, 512)) 28 | 29 | 30 | def get_concat_h(im1, im2): 31 | dst = Image.new("RGB", (im1.width + im2.width, im1.height)) 32 | dst.paste(im1, (0, 0)) 33 | dst.paste(im2, (im1.width, 0)) 34 | return dst 35 | 36 | 37 | 38 | 39 | def get_args(): 40 | parser = OptionParser() 41 | parser.add_option("--bg", help="Directory to the background image.") 42 | parser.add_option("--fg", help="Directory to the foreground image.") 43 | 44 | parser.add_option("--checkpoints", "--ld", help="Directory to checkpoints, default is model/ckpt_g39.pth") 45 | 46 | parser.add_option( 47 | "--gpu", 48 | action="store_true", 49 | help="If specified, will use GPU", 50 | ) 51 | 52 | parser.add_option( 53 | "--light", 54 | action="store_true", 55 | help="If specified, will use light model", 56 | ) 57 | 58 | (options, args) = parser.parse_args() 59 | return options 60 | 61 | 62 | 63 | 64 | 65 | class Evaluater: 66 | def __init__(self): 67 | 68 | self.args = get_args() 69 | 70 | self.name_cat = self.args.bg.split('/')[-1].split('.')[0]+'_'+self.args.fg.split('/')[-1].split('.')[0] 71 | 72 | 73 | self.fg = Image.open(self.args.fg) 74 | 75 | self.mask = self.fg.split()[-1] 76 | 77 | self.background = Image.open(self.args.bg).resize(self.fg.size) 78 | 79 | self.img_composite = Image.composite(self.fg, self.background, self.mask) 80 | 81 | 82 | 83 | 84 | 85 | if self.args.gpu: 86 | device = "cuda" 87 | else: 88 | device = "cpu" 89 | 90 | self.Model = Model_Composite_PL( 91 | dim=32, 92 | masking=True, 93 | brush=True, 94 | maskoffset=0.6, 95 | swap=True, 96 | Vit_bool=False, 97 | onlyupsample=True, 98 | aggupsample=True, 99 | light=self.args.light, 100 | Eff_bool=self.args.light, 101 | ).to(device) 102 | 103 | if self.args.checkpoints is not None: 104 | model_path = self.args.checkpoints 105 | else: 106 | model_path = os.getcwd() + '/pretrained/ckpt_g39.pth' 107 | 108 | checkpoint = torch.load(model_path, map_location=device) 109 | self.Model.load_state_dict(checkpoint["state_dict"]) 110 | 111 | self.Model.eval() 112 | self.bg_low= resize(self.background) 113 | self.composite_low= resize(self.img_composite) 114 | self.mask_low = resize(self.mask) 115 | 116 | # Load image 117 | 118 | 119 | self.torch_bg = transform(self.background).to(device) 120 | self.torch_composite = transform(self.img_composite).to(device) 121 | self.torch_mask = transforms_mask(self.mask).to(device) 122 | 123 | self.torch_bg_low = transform(self.bg_low).to(device) 124 | self.torch_composite_low = transform(self.composite_low).to(device) 125 | self.torch_mask_low = transforms_mask(self.mask_low).to(device) 126 | 127 | def evaluate(self): 128 | 129 | with torch.no_grad(): 130 | inter_composite, output_composite, par1, par2 = self.Model( 131 | self.torch_bg_low[None, ...], 132 | self.torch_composite_low[None, ...], 133 | self.torch_mask_low[None, ...], 134 | ) 135 | 136 | 137 | hr_intermediate = ( 138 | self.Model.PL3D(self.Model.pl_table, self.torch_composite[None,...]) * self.torch_mask 139 | + (1 - self.torch_mask) * self.torch_bg 140 | ) 141 | 142 | Gainmap_Resize = T.Resize(self.torch_bg.shape[-2:]) 143 | # print(Gain_map) 144 | 145 | output_results = ( 146 | hr_intermediate * Gainmap_Resize(self.Model.gainmap) * self.torch_mask 147 | + (1 - self.torch_mask) * self.torch_bg 148 | ) 149 | 150 | output_lr = T.ToPILImage()(output_results[0,...]) 151 | output_lr.save('results/%s_final.png'%(self.name_cat)) 152 | 153 | output_gm = T.ToPILImage()( (Gainmap_Resize(self.Model.gainmap) * self.torch_mask)[0,...]) 154 | 155 | output_gm.save('results/%s_gainmap.png'%(self.name_cat)) 156 | 157 | 158 | 159 | 160 | 161 | #### Save Fig 162 | 163 | curves = par2.cpu().detach().numpy() 164 | 165 | red_curve = curves[0, 0, 0, 0, :] 166 | green_curve = curves[0, 1, 0, :, 0] 167 | blue_curve = curves[0, 2, :, 0, 0] 168 | 169 | plt.figure() 170 | plt.plot(np.linspace(0, 1, 32), red_curve, "r") 171 | plt.plot(np.linspace(0, 1, 32), green_curve, "g") 172 | plt.plot(np.linspace(0, 1, 32), blue_curve, "b") 173 | plt.ylim(0, 1) 174 | plt.legend(["Reg", "Green", "Blue"]) 175 | plt.title("Learned Color Curves") 176 | 177 | plt.savefig("results/%s_color.jpg"%(self.name_cat)) 178 | 179 | 180 | 181 | plt.close() 182 | 183 | im_final = get_concat_h( self.img_composite,get_concat_h(self.mask,output_lr)) 184 | 185 | im_final.save('results/%s_results_summary.png'%(self.name_cat)) 186 | 187 | 188 | if __name__ == "__main__": 189 | torch.backends.cudnn.benchmark = False 190 | torch.backends.cudnn.deterministic = True 191 | evaluater = Evaluater() 192 | evaluater.evaluate() -------------------------------------------------------------------------------- /inference_scripts/Inference.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | 7 | CUDA_VISIBLE_DEVICES=$3 8 | 9 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 10 | 11 | CUDA_VISIBLE_DEVICES=$3 python PIH_test.py --datadir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing/ \ 12 | -g 0 \ 13 | --checkpoints $1 \ 14 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 15 | --bs 1 \ 16 | 17 | CUDA_VISIBLE_DEVICES=$3 python PIH_test.py --datadir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_different_large/ \ 18 | -g 0 \ 19 | --checkpoints $1 \ 20 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing_different/ \ 21 | --bs 1 \ 22 | 23 | 24 | CUDA_VISIBLE_DEVICES=$3 python PIH_test.py --datadir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_real_large/ \ 25 | -g 0 \ 26 | --checkpoints $1 \ 27 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing_real/ \ 28 | --ngt \ 29 | --bs 1 \ 30 | 31 | CUDA_VISIBLE_DEVICES=$3 python PIH_test.py --datadir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_real_aug/ \ 32 | -g 0 \ 33 | --checkpoints $1 \ 34 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing_real_aug/ \ 35 | --ngt \ 36 | --bs 1 \ 37 | 38 | mkdir /home/kewang/website_michael/results/$2/ 39 | 40 | cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 41 | 42 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 43 | /home/kewang/website_michael/results/$2/results_testing/ \ 44 | --order original intermediate results gt \ 45 | --name $2-results_testing 46 | 47 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing_different/ \ 48 | /home/kewang/website_michael/results/$2/results_testing_different/ \ 49 | --order original intermediate results gt \ 50 | --name $2-results_testing_different 51 | 52 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing_real/ \ 53 | /home/kewang/website_michael/results/$2/results_testing_real/ \ 54 | --order original intermediate results\ 55 | --name $2-real 56 | 57 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing_real_aug/ \ 58 | /home/kewang/website_michael/results/$2/results_testing_real_aug/ \ 59 | --order original intermediate results\ 60 | --name $2-real-aug 61 | 62 | # # Network hyperparameters 63 | # device=1 64 | # lr=1e-5 65 | # batch_size=16 66 | # date=202206070 67 | # name=iharmonysimplemodel 68 | 69 | # model_name=exp_${date}_batch_size_$((batch_size))_lr_${lr}_${name} 70 | 71 | # # Set folder names 72 | # dir_data=/mnt/localssd/Image_Harmonization_Dataset/ 73 | # dir_log=/home/kewang/sensei-fs-symlink/users/kewang/projects/PIH/PIH_ResNet/results/$model_name 74 | 75 | 76 | 77 | # CUDA_VISIBLE_DEVICES=$device python PIH_train.py --datadir $dir_data \ 78 | # -g 0 \ 79 | # --logdir $dir_log \ 80 | # --bs $batch_size \ 81 | # --lr $lr \ 82 | # --force_train_from_scratch \ 83 | # --ihd \ 84 | -------------------------------------------------------------------------------- /inference_scripts/Inference_Composite.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | 7 | CUDA_VISIBLE_DEVICES=$3 8 | 9 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 10 | 11 | CUDA_VISIBLE_DEVICES=$3 python PIH_test_compositeGAN.py --datadir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_LR_self_2000/ \ 12 | -g 0 \ 13 | --checkpoints $1 \ 14 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 15 | --bs 1 \ 16 | --composite \ 17 | --lut \ 18 | --lut-dim 16 \ 19 | --num-testing 2000 \ 20 | --nocurve \ 21 | --piecewiselinear \ 22 | 23 | mkdir /home/kewang/website_michael/results/$2/ 24 | 25 | cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 26 | 27 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 28 | /home/kewang/website_michael/results/$2/results_testing/ \ 29 | --order bg mask original results real curves\ 30 | --name $2-results_testing 31 | 32 | 33 | -------------------------------------------------------------------------------- /inference_scripts/Inference_Composite_masking.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | echo mask offset: $4 7 | 8 | echo $5 9 | 10 | if [ $5 == realhm ] 11 | then 12 | a=File_for_testing_composite_realhm 13 | elif [ $5 == realhr ] 14 | then 15 | a=File_for_testing_composite_realhr 16 | elif [ $5 == iharm ] 17 | then 18 | a=File_for_testing_composite_iharm 19 | elif [ $5 == LR ] 20 | then 21 | a=File_for_testing_composite_LR_2000 22 | 23 | elif [ $5 == LRself ] 24 | then 25 | a=File_for_testing_composite_LR_self_2000 26 | 27 | 28 | elif [ $5 == self ] 29 | then 30 | a=File_for_testing_self 31 | 32 | elif [ $5 == realself ] 33 | then 34 | a=File_for_testing_composite_realself 35 | 36 | elif [ $5 == adobe ] 37 | then 38 | a=File_for_testing_composite_adobereal 39 | 40 | else 41 | a=0 42 | fi 43 | 44 | echo Data stream: $a 45 | 46 | CUDA_VISIBLE_DEVICES=$3 47 | 48 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 49 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_realhm/ 50 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_LR_self_2000/ 51 | 52 | CUDA_VISIBLE_DEVICES=$3 python PIH_test_compositeGAN_masking.py --datadir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/$a \ 53 | -g 0 \ 54 | --checkpoints $1 \ 55 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 56 | --bs 1 \ 57 | --composite \ 58 | --num-testing 50000 \ 59 | --nocurve \ 60 | --piecewiselinear \ 61 | --masking \ 62 | --brush \ 63 | --maskoffset $4 \ 64 | --swap \ 65 | --onlyupsample \ 66 | --twoinputs \ 67 | --aggupsample \ 68 | --dim 64 \ 69 | # --effbool \ 70 | 71 | # --lowres \ 72 | # --effbool \ 73 | # --dim 64 \ 74 | # --lut \ 75 | # --lut-dim 16 \ 76 | 77 | 78 | # --vitbool \ 79 | 80 | # --pihnetbool \ 81 | # --pihnetbool \ 82 | # --lut \ 83 | 84 | # --lut \ 85 | # --onlyupsample \ 86 | # --maskconvkernel 3 \ 87 | 88 | # # --onlyupsample \ 89 | # # --maskconvkernel 3 \ 90 | # # --maskoffset 0 \ 91 | 92 | # # --onlyupsample \ 93 | # # --maskconvkernel 3 \ 94 | # # --maskoffset 0 \ 95 | # # --mask 96 | 97 | # # --onlyupsample \ 98 | # # --nosig \ 99 | 100 | 101 | mkdir /home/kewang/website_michael/results/$2/ 102 | 103 | cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 104 | 105 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 106 | /home/kewang/website_michael/results/$2/results_testing/ \ 107 | --order bg mask original intermediate results real curves\ 108 | --name $2-results_testing 109 | 110 | 111 | -------------------------------------------------------------------------------- /inference_scripts/Inference_Composite_masking_3.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | echo mask offset: $4 7 | 8 | echo $5 9 | 10 | if [ $5 == realhm ] 11 | then 12 | a=File_for_testing_composite_realhm 13 | elif [ $5 == realhr ] 14 | then 15 | a=File_for_testing_composite_realhr 16 | elif [ $5 == iharm ] 17 | then 18 | a=File_for_testing_composite_iharm 19 | elif [ $5 == LR ] 20 | then 21 | a=File_for_testing_composite_LR_2000 22 | 23 | elif [ $5 == LRself ] 24 | then 25 | a=File_for_testing_composite_LR_self_2000 26 | 27 | 28 | elif [ $5 == self ] 29 | then 30 | a=File_for_testing_self 31 | 32 | 33 | elif [ $5 == adobe ] 34 | then 35 | a=File_for_testing_composite_adobereal 36 | 37 | elif [ $5 == realself ] 38 | then 39 | a=File_for_testing_composite_realself 40 | 41 | 42 | else 43 | a=0 44 | fi 45 | 46 | echo Data stream: $a 47 | 48 | CUDA_VISIBLE_DEVICES=$3 49 | 50 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 51 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_realhm/ 52 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_LR_self_2000/ 53 | 54 | CUDA_VISIBLE_DEVICES=$3 python PIH_test_compositeGAN_masking.py --datadir /mnt/localssd/$a \ 55 | -g 0 \ 56 | --checkpoints $1 \ 57 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 58 | --bs 1 \ 59 | --composite \ 60 | --num-testing 500 \ 61 | --nocurve \ 62 | --piecewiselinear \ 63 | --masking \ 64 | --brush \ 65 | --maskoffset $4 \ 66 | --swap \ 67 | --onlyupsample \ 68 | --aggupsample \ 69 | --dim 32 \ 70 | # --lut \ 71 | # --lut-dim 16 \ 72 | # --lowres \ 73 | # --vitbool \ 74 | # --ibn \ 75 | # --effbool \ 76 | # --ibn \ 77 | # --vitbool \ 78 | # --ibn \ 79 | # --bgshadow \ 80 | # --twoinputs \ 81 | 82 | 83 | # --vitbool \ 84 | 85 | # --pihnetbool \ 86 | # --pihnetbool \ 87 | # --lut \ 88 | 89 | # --lut \ 90 | # --onlyupsample \ 91 | # --maskconvkernel 3 \ 92 | 93 | # # --onlyupsample \ 94 | # # --maskconvkernel 3 \ 95 | # # --maskoffset 0 \ 96 | 97 | # # --onlyupsample \ 98 | # # --maskconvkernel 3 \ 99 | # # --maskoffset 0 \ 100 | # # --mask 101 | 102 | # # --onlyupsample \ 103 | # # --nosig \ 104 | 105 | 106 | mkdir /home/kewang/website_michael/results/$2/ 107 | 108 | cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 109 | 110 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 111 | /home/kewang/website_michael/results/$2/results_testing/ \ 112 | --order bg mask original intermediate results real curves\ 113 | --name $2-results_testing 114 | 115 | 116 | -------------------------------------------------------------------------------- /inference_scripts/Inference_Composite_masking_3_depth.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | echo mask offset: $4 7 | 8 | echo $5 9 | 10 | echo dim $6 11 | if [ $5 == realhm ] 12 | then 13 | a=File_for_testing_composite_realhm 14 | elif [ $5 == realhr ] 15 | then 16 | a=File_for_testing_composite_realhr 17 | elif [ $5 == iharm ] 18 | then 19 | a=File_for_testing_composite_iharm 20 | elif [ $5 == LR ] 21 | then 22 | a=File_for_testing_composite_LR_2000 23 | 24 | elif [ $5 == LRself ] 25 | then 26 | a=File_for_testing_composite_LR_self_2000 27 | 28 | elif [ $5 == adobe ] 29 | then 30 | a=File_for_testing_composite_adobereal 31 | 32 | elif [ $5 == realself ] 33 | then 34 | a=File_for_testing_composite_realself 35 | 36 | 37 | 38 | 39 | elif [ $5 == self ] 40 | then 41 | a=File_for_testing_self 42 | else 43 | a=0 44 | fi 45 | 46 | echo Data stream: $a 47 | 48 | CUDA_VISIBLE_DEVICES=$3 49 | 50 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 51 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_realhm/ 52 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_LR_self_2000/ 53 | 54 | CUDA_VISIBLE_DEVICES=$3 python PIH_test_compositeGAN_masking.py --datadir /mnt/localssd/$a \ 55 | -g 0 \ 56 | --checkpoints $1 \ 57 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 58 | --bs 1 \ 59 | --composite \ 60 | --num-testing 500 \ 61 | --nocurve \ 62 | --piecewiselinear \ 63 | --masking \ 64 | --brush \ 65 | --maskoffset $4 \ 66 | --swap \ 67 | --onlyupsample \ 68 | --aggupsample \ 69 | --depthmap \ 70 | --dim $6 \ 71 | --bgshadow \ 72 | --dual \ 73 | # --ibn \ 74 | # --effbool \ 75 | # --twoinputs \ 76 | 77 | # --ibn \ 78 | # --depth \ 79 | # --twoinputs \ 80 | 81 | 82 | # --vitbool \ 83 | 84 | # --pihnetbool \ 85 | # --pihnetbool \ 86 | # --lut \ 87 | 88 | # --lut \ 89 | # --onlyupsample \ 90 | # --maskconvkernel 3 \ 91 | 92 | # # --onlyupsample \ 93 | # # --maskconvkernel 3 \ 94 | # # --maskoffset 0 \ 95 | 96 | # # --onlyupsample \ 97 | # # --maskconvkernel 3 \ 98 | # # --maskoffset 0 \ 99 | # # --mask 100 | 101 | # # --onlyupsample \ 102 | # # --nosig \ 103 | 104 | 105 | mkdir /home/kewang/website_michael/results/$2/ 106 | 107 | cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 108 | 109 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 110 | /home/kewang/website_michael/results/$2/results_testing/ \ 111 | --order bg mask original intermediate results real curves\ 112 | --name $2-results_testing 113 | 114 | 115 | -------------------------------------------------------------------------------- /inference_scripts/Inference_Composite_masking_3_highres.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | echo mask offset: $4 7 | 8 | echo $5 9 | 10 | if [ $5 == realhm ] 11 | then 12 | a=File_for_testing_composite_realhm 13 | 14 | 15 | elif [ $5 == realhmfull ] 16 | then 17 | a=File_for_testing_composite_realhmfullres 18 | 19 | elif [ $5 == realhr ] 20 | then 21 | a=File_for_testing_composite_realhr 22 | elif [ $5 == iharm ] 23 | then 24 | a=File_for_testing_composite_iharm 25 | elif [ $5 == LR ] 26 | then 27 | a=File_for_testing_composite_LR_2000 28 | 29 | elif [ $5 == LRself ] 30 | then 31 | a=File_for_testing_composite_LR_self_1000_HR_final 32 | 33 | 34 | 35 | elif [ $5 == self ] 36 | then 37 | a=File_for_testing_self 38 | 39 | 40 | elif [ $5 == adobe ] 41 | then 42 | a=File_for_testing_composite_adobereal 43 | 44 | elif [ $5 == realself ] 45 | then 46 | a=File_for_testing_composite_realself 47 | 48 | 49 | else 50 | a=0 51 | fi 52 | 53 | echo Data stream: $a 54 | 55 | CUDA_VISIBLE_DEVICES=$3 56 | 57 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 58 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_realhm/ 59 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_LR_self_2000/ 60 | 61 | CUDA_VISIBLE_DEVICES=$3 python PIH_test_compositeGAN_masking_highres.py --datadir /mnt/localssd/$a \ 62 | -g 0 \ 63 | --checkpoints $1 \ 64 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 65 | --bs 1 \ 66 | --composite \ 67 | --num-testing 500000 \ 68 | --nocurve \ 69 | --piecewiselinear \ 70 | --masking \ 71 | --brush \ 72 | --maskoffset $4 \ 73 | --swap \ 74 | --onlyupsample \ 75 | --aggupsample \ 76 | --dim 32 \ 77 | --original \ 78 | --lrdata \ 79 | 80 | # --lut \ 81 | # --lut-dim 16 \ 82 | # --lowres \ 83 | # --vitbool \ 84 | # --ibn \ 85 | # --effbool \ 86 | # --ibn \ 87 | # --vitbool \ 88 | # --ibn \ 89 | # --bgshadow \ 90 | # --twoinputs \ 91 | 92 | 93 | # --vitbool \ 94 | 95 | # --pihnetbool \ 96 | # --pihnetbool \ 97 | # --lut \ 98 | 99 | # --lut \ 100 | # --onlyupsample \ 101 | # --maskconvkernel 3 \ 102 | 103 | # # --onlyupsample \ 104 | # # --maskconvkernel 3 \ 105 | # # --maskoffset 0 \ 106 | 107 | # # --onlyupsample \ 108 | # # --maskconvkernel 3 \ 109 | # # --maskoffset 0 \ 110 | # # --mask 111 | 112 | # # --onlyupsample \ 113 | # # --nosig \ 114 | 115 | 116 | # mkdir /home/kewang/website_michael/results/$2/ 117 | 118 | # cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 119 | 120 | # python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 121 | # /home/kewang/website_michael/results/$2/results_testing/ \ 122 | # --order bg mask original intermediate results real curves\ 123 | # --name $2-results_testing 124 | 125 | 126 | -------------------------------------------------------------------------------- /inference_scripts/Inference_Composite_masking_3_noweb.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | echo mask offset: $4 7 | 8 | echo $5 9 | 10 | if [ $5 == realhm ] 11 | then 12 | a=File_for_testing_composite_realhm 13 | elif [ $5 == realhr ] 14 | then 15 | a=File_for_testing_composite_realhr 16 | elif [ $5 == iharm ] 17 | then 18 | a=File_for_testing_composite_iharm 19 | elif [ $5 == LR ] 20 | then 21 | a=File_for_testing_composite_LR_2000 22 | 23 | elif [ $5 == LRself ] 24 | then 25 | a=File_for_testing_composite_LR_self_2000 26 | 27 | 28 | elif [ $5 == self ] 29 | then 30 | a=File_for_testing_self 31 | 32 | 33 | elif [ $5 == adobe ] 34 | then 35 | a=File_for_testing_composite_adobereal 36 | 37 | elif [ $5 == realself ] 38 | then 39 | a=File_for_testing_composite_realself 40 | 41 | 42 | else 43 | a=0 44 | fi 45 | 46 | echo Data stream: $a 47 | 48 | CUDA_VISIBLE_DEVICES=$3 49 | 50 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 51 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_realhm/ 52 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_LR_self_2000/ 53 | 54 | CUDA_VISIBLE_DEVICES=$3 python PIH_test_compositeGAN_masking_tiny.py --datadir /mnt/localssd/$a \ 55 | -g 0 \ 56 | --checkpoints $1 \ 57 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 58 | --bs 1 \ 59 | --composite \ 60 | --num-testing 500 \ 61 | --nocurve \ 62 | --piecewiselinear \ 63 | --masking \ 64 | --brush \ 65 | --maskoffset $4 \ 66 | --swap \ 67 | --onlyupsample \ 68 | --aggupsample \ 69 | --dim 32 \ 70 | # --lut \ 71 | # --lut-dim 16 \ 72 | # --lowres \ 73 | # --vitbool \ 74 | # --ibn \ 75 | # --effbool \ 76 | # --ibn \ 77 | # --vitbool \ 78 | # --ibn \ 79 | # --bgshadow \ 80 | # --twoinputs \ 81 | 82 | 83 | # --vitbool \ 84 | 85 | # --pihnetbool \ 86 | # --pihnetbool \ 87 | # --lut \ 88 | 89 | # --lut \ 90 | # --onlyupsample \ 91 | # --maskconvkernel 3 \ 92 | 93 | # # --onlyupsample \ 94 | # # --maskconvkernel 3 \ 95 | # # --maskoffset 0 \ 96 | 97 | # # --onlyupsample \ 98 | # # --maskconvkernel 3 \ 99 | # # --maskoffset 0 \ 100 | # # --mask 101 | 102 | # # --onlyupsample \ 103 | # # --nosig \ 104 | 105 | 106 | 107 | 108 | -------------------------------------------------------------------------------- /inference_scripts/Inference_Composite_masking_highres.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | echo mask offset: $4 7 | 8 | echo $5 9 | 10 | if [ $5 == realhm ] 11 | then 12 | a=File_for_testing_composite_realhm 13 | elif [ $5 == realhr ] 14 | then 15 | a=File_for_testing_composite_realhr 16 | elif [ $5 == iharm ] 17 | then 18 | a=File_for_testing_composite_iharm 19 | elif [ $5 == LR ] 20 | then 21 | a=File_for_testing_composite_LR_2000 22 | 23 | elif [ $5 == LRself ] 24 | then 25 | a=File_for_testing_composite_LR_self_2000 26 | 27 | 28 | elif [ $5 == self ] 29 | then 30 | a=File_for_testing_self 31 | 32 | elif [ $5 == realself ] 33 | then 34 | a=File_for_testing_composite_realself 35 | 36 | elif [ $5 == adobe ] 37 | then 38 | a=File_for_testing_composite_adobereal 39 | 40 | elif [ $5 == adobe5k ] 41 | then 42 | a=File_for_testing_HAdobe5k 43 | 44 | 45 | 46 | else 47 | a=0 48 | fi 49 | 50 | echo Data stream: $a 51 | 52 | CUDA_VISIBLE_DEVICES=$3 53 | 54 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 55 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_realhm/ 56 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_LR_self_2000/ 57 | 58 | CUDA_VISIBLE_DEVICES=$3 python PIH_test_compositeGAN_masking_highres.py --datadir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/$a \ 59 | -g 0 \ 60 | --checkpoints $1 \ 61 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 62 | --bs 1 \ 63 | --composite \ 64 | --num-testing 50000 \ 65 | --nocurve \ 66 | --piecewiselinear \ 67 | --masking \ 68 | --brush \ 69 | --maskoffset $4 \ 70 | --swap \ 71 | --onlyupsample \ 72 | --twoinputs \ 73 | --aggupsample \ 74 | --dim 64 \ 75 | --original \ 76 | # --effbool \ 77 | 78 | # --lowres \ 79 | # --effbool \ 80 | # --dim 64 \ 81 | # --lut \ 82 | # --lut-dim 16 \ 83 | 84 | 85 | # --vitbool \ 86 | 87 | # --pihnetbool \ 88 | # --pihnetbool \ 89 | # --lut \ 90 | 91 | # --lut \ 92 | # --onlyupsample \ 93 | # --maskconvkernel 3 \ 94 | 95 | # # --onlyupsample \ 96 | # # --maskconvkernel 3 \ 97 | # # --maskoffset 0 \ 98 | 99 | # # --onlyupsample \ 100 | # # --maskconvkernel 3 \ 101 | # # --maskoffset 0 \ 102 | # # --mask 103 | 104 | # # --onlyupsample \ 105 | # # --nosig \ 106 | 107 | 108 | # mkdir /home/kewang/website_michael/results/$2/ 109 | 110 | # cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 111 | 112 | # python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 113 | # /home/kewang/website_michael/results/$2/results_testing/ \ 114 | # --order bg mask original intermediate results real curves\ 115 | # --name $2-results_testing 116 | 117 | 118 | -------------------------------------------------------------------------------- /inference_scripts/Inference_Composite_masking_pixel.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | echo mask offset: $4 7 | 8 | echo $5 9 | 10 | if [ $5 == realhm ] 11 | then 12 | a=File_for_testing_composite_realhm 13 | elif [ $5 == realhr ] 14 | then 15 | a=File_for_testing_composite_realhr 16 | elif [ $5 == iharm ] 17 | then 18 | a=File_for_testing_composite_iharm 19 | elif [ $5 == LR ] 20 | then 21 | a=File_for_testing_composite_LR_2000 22 | 23 | elif [ $5 == LRself ] 24 | then 25 | a=File_for_testing_composite_LR_self_2000 26 | 27 | 28 | elif [ $5 == self ] 29 | then 30 | a=File_for_testing_self 31 | 32 | elif [ $5 == realself ] 33 | then 34 | a=File_for_testing_composite_realself 35 | 36 | elif [ $5 == adobe ] 37 | then 38 | a=File_for_testing_composite_adobereal 39 | 40 | else 41 | a=0 42 | fi 43 | 44 | echo Data stream: $a 45 | 46 | CUDA_VISIBLE_DEVICES=$3 47 | 48 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 49 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_realhm/ 50 | # /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_LR_self_2000/ 51 | 52 | CUDA_VISIBLE_DEVICES=$3 python PIH_test_compositeGAN_masking.py --datadir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/$a \ 53 | -g 0 \ 54 | --checkpoints $1 \ 55 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 56 | --bs 1 \ 57 | --composite \ 58 | --num-testing 500 \ 59 | --nocurve \ 60 | --piecewiselinear \ 61 | --masking \ 62 | --maskoffset $4 \ 63 | --swap \ 64 | --twoinputs \ 65 | --dim 64 \ 66 | # --lut \ 67 | # --lut-dim 24 \ 68 | # --effbool \ 69 | 70 | # --lowres \ 71 | # --effbool \ 72 | # --dim 64 \ 73 | # --lut \ 74 | # --lut-dim 16 \ 75 | 76 | 77 | # --vitbool \ 78 | 79 | # --pihnetbool \ 80 | # --pihnetbool \ 81 | # --lut \ 82 | 83 | # --lut \ 84 | # --onlyupsample \ 85 | # --maskconvkernel 3 \ 86 | 87 | # # --onlyupsample \ 88 | # # --maskconvkernel 3 \ 89 | # # --maskoffset 0 \ 90 | 91 | # # --onlyupsample \ 92 | # # --maskconvkernel 3 \ 93 | # # --maskoffset 0 \ 94 | # # --mask 95 | 96 | # # --onlyupsample \ 97 | # # --nosig \ 98 | 99 | 100 | mkdir /home/kewang/website_michael/results/$2/ 101 | 102 | cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 103 | 104 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 105 | /home/kewang/website_michael/results/$2/results_testing/ \ 106 | --order bg mask original intermediate results real curves\ 107 | --name $2-results_testing 108 | 109 | 110 | -------------------------------------------------------------------------------- /inference_scripts/Inference_Composite_unet.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | 7 | CUDA_VISIBLE_DEVICES=$3 8 | 9 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 10 | 11 | CUDA_VISIBLE_DEVICES=$3 python PIH_test_compositeGAN.py --datadir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/File_for_testing_composite_2000/ \ 12 | -g 0 \ 13 | --checkpoints $1 \ 14 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 15 | --bs 1 \ 16 | --composite \ 17 | --unet \ 18 | --num-testing 2000 \ 19 | 20 | mkdir /home/kewang/website_michael/results/$2/ 21 | 22 | cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 23 | 24 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing/ \ 25 | /home/kewang/website_michael/results/$2/results_testing/ \ 26 | --order bg mask original intermediate results real \ 27 | --name $2-results_testing 28 | 29 | # # Network hyperparameters 30 | # device=1 31 | # lr=1e-5 32 | # batch_size=16 33 | # date=202206070 34 | # name=iharmonysimplemodel 35 | 36 | # model_name=exp_${date}_batch_size_$((batch_size))_lr_${lr}_${name} 37 | 38 | # # Set folder names 39 | # dir_data=/mnt/localssd/Image_Harmonization_Dataset/ 40 | # dir_log=/home/kewang/sensei-fs-symlink/users/kewang/projects/PIH/PIH_ResNet/results/$model_name 41 | 42 | 43 | 44 | # CUDA_VISIBLE_DEVICES=$device python PIH_train.py --datadir $dir_data \ 45 | # -g 0 \ 46 | # --logdir $dir_log \ 47 | # --bs $batch_size \ 48 | # --lr $lr \ 49 | # --force_train_from_scratch \ 50 | # --ihd \ 51 | -------------------------------------------------------------------------------- /inference_scripts/Inference_iHarmony.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | echo Checkpoint: $1 4 | echo Model_Name: $2 5 | echo GPU: $3 6 | 7 | CUDA_VISIBLE_DEVICES=$3 8 | 9 | mkdir /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/ 10 | 11 | 12 | 13 | CUDA_VISIBLE_DEVICES=$3 python PIH_test.py --datadir /mnt/localssd/Image_Harmonization_Dataset/ \ 14 | -g 0 \ 15 | --checkpoints $1 \ 16 | --tmp_results /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing_iHarmony/ \ 17 | --bs 1 \ 18 | --num-testing 1000 \ 19 | --ihd \ 20 | 21 | 22 | mkdir /home/kewang/website_michael/results/$2/ 23 | 24 | cd /home/kewang/website_michael/ttools2-main/ttools2/scripts/ 25 | 26 | python /home/kewang/website_michael/ttools2-main/ttools2/scripts/image_gallery.py /home/kewang/sensei-fs-symlink/users/kewang/projects/data_processing/results_images/$2/results_testing_iHarmony/ \ 27 | /home/kewang/website_michael/results/$2/results_testing_iHarmony/ \ 28 | --order original mask intermediate results gt\ 29 | --name $2-results-testing-iHarmony 30 | 31 | # # Network hyperparameters 32 | # device=1 33 | # lr=1e-5 34 | # batch_size=16 35 | # date=202206070 36 | # name=iharmonysimplemodel 37 | 38 | # model_name=exp_${date}_batch_size_$((batch_size))_lr_${lr}_${name} 39 | 40 | # # Set folder names 41 | # dir_data=/mnt/localssd/Image_Harmonization_Dataset/ 42 | # dir_log=/home/kewang/sensei-fs-symlink/users/kewang/projects/PIH/PIH_ResNet/results/$model_name 43 | 44 | 45 | 46 | # CUDA_VISIBLE_DEVICES=$device python PIH_train.py --datadir $dir_data \ 47 | # -g 0 \ 48 | # --logdir $dir_log \ 49 | # --bs $batch_size \ 50 | # --lr $lr \ 51 | # --force_train_from_scratch \ 52 | # --ihd \ 53 | -------------------------------------------------------------------------------- /pretrained/pretrained.placeholder: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/pretrained/pretrained.placeholder -------------------------------------------------------------------------------- /results/results.placeholder: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/adobe/PIH/2823cccf0778c6ea213a3d366f03864ac8ab82e6/results/results.placeholder -------------------------------------------------------------------------------- /scripts/installation.sh: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | conda create -n pytorch_pih python=3.9 11 | conda activate pytorch_pih 12 | conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia 13 | pip install matplotlib 14 | pip install opencv-python 15 | pip install tqdm 16 | -------------------------------------------------------------------------------- /scripts/train_example.sh: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | 11 | # Network hyperparameters 12 | device=0 13 | lr=4e-5 14 | lrd=4e-5 15 | batch_size=8 16 | date=Demo_train_example 17 | reconweight=None 18 | training_ratio=1 19 | inputdimD=3 20 | recon_ratio=0.5 21 | recon_weight=0.92 ## Used here 22 | 23 | name=iharmony_${inputdimD}_ratio_${training_ratio}_${recon_ratio}_reconweight_${recon_weight} 24 | 25 | model_name=exp_${date}_batch_size_$((batch_size))_lr_${lr}_${name}_device_${device} 26 | 27 | # Set folder names 28 | dir_data=demo_data/train/ 29 | # dir_data=/mnt/localssd/LR_data/train/ 30 | 31 | 32 | CUDA_VISIBLE_DEVICES=$device python PIH_train.py --datadir $dir_data \ 33 | -g 0 \ 34 | --onlysaveg \ 35 | --bs $batch_size \ 36 | --lr $lr \ 37 | --lrd $lrd \ 38 | --force_train_from_scratch \ 39 | --tempdir \ 40 | $model_name \ 41 | --workers 8 \ 42 | --trainingratio ${training_ratio} \ 43 | --unetd \ 44 | --inputdimD ${inputdimD} \ 45 | --nocurve \ 46 | --reconratio ${recon_ratio} \ 47 | --piecewiselinear \ 48 | --pl-dim 32 \ 49 | --pairaugment \ 50 | --purepairaugment \ 51 | --lowdim \ 52 | --ganlossmask \ 53 | --reconwithgan \ 54 | --reconweight ${recon_weight} \ 55 | --masking \ 56 | --brush \ 57 | --maskoffset 0.5 \ 58 | --swap \ 59 | --onlyupsample \ 60 | --joint \ 61 | --lessskip \ 62 | --aggupsample \ 63 | --scheduler \ 64 | --colorjitter \ 65 | 66 | 67 | 68 | 69 | 70 | 71 | -------------------------------------------------------------------------------- /utils/efficientnet_v2.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | import collections.abc as container_abc 11 | from collections import OrderedDict 12 | from math import ceil, floor 13 | 14 | import torch 15 | import torch.nn as nn 16 | import torch.nn.functional as F 17 | from torch.utils import model_zoo 18 | 19 | 20 | def _pair(x): 21 | if isinstance(x, container_abc.Iterable): 22 | return x 23 | return (x, x) 24 | 25 | 26 | def torch_conv_out_spatial_shape(in_spatial_shape, kernel_size, stride): 27 | if in_spatial_shape is None: 28 | return None 29 | # in_spatial_shape -> [H,W] 30 | hin, win = _pair(in_spatial_shape) 31 | kh, kw = _pair(kernel_size) 32 | sh, sw = _pair(stride) 33 | 34 | # dilation and padding are ignored since they are always fixed in efficientnetV2 35 | hout = int(floor((hin - kh - 1) / sh + 1)) 36 | wout = int(floor((win - kw - 1) / sw + 1)) 37 | return hout, wout 38 | 39 | 40 | def get_activation(act_fn: str, **kwargs): 41 | if act_fn in ("silu", "swish"): 42 | return nn.SiLU(**kwargs) 43 | elif act_fn == "relu": 44 | return nn.ReLU(**kwargs) 45 | elif act_fn == "relu6": 46 | return nn.ReLU6(**kwargs) 47 | elif act_fn == "elu": 48 | return nn.ELU(**kwargs) 49 | elif act_fn == "leaky_relu": 50 | return nn.LeakyReLU(**kwargs) 51 | elif act_fn == "selu": 52 | return nn.SELU(**kwargs) 53 | elif act_fn == "mish": 54 | return nn.Mish(**kwargs) 55 | else: 56 | raise ValueError("Unsupported act_fn {}".format(act_fn)) 57 | 58 | 59 | def round_filters(filters, width_coefficient, depth_divisor=8): 60 | """Round number of filters based on depth multiplier.""" 61 | min_depth = depth_divisor 62 | filters *= width_coefficient 63 | new_filters = max( 64 | min_depth, int(filters + depth_divisor / 2) // depth_divisor * depth_divisor 65 | ) 66 | return int(new_filters) 67 | 68 | 69 | def round_repeats(repeats, depth_coefficient): 70 | """Round number of filters based on depth multiplier.""" 71 | return int(ceil(depth_coefficient * repeats)) 72 | 73 | 74 | class DropConnect(nn.Module): 75 | def __init__(self, rate=0.5): 76 | super(DropConnect, self).__init__() 77 | self.keep_prob = None 78 | self.set_rate(rate) 79 | 80 | def set_rate(self, rate): 81 | if not 0 <= rate < 1: 82 | raise ValueError("rate must be 0<=rate<1, got {} instead".format(rate)) 83 | self.keep_prob = 1 - rate 84 | 85 | def forward(self, x): 86 | if self.training: 87 | random_tensor = self.keep_prob + torch.rand( 88 | [x.size(0), 1, 1, 1], dtype=x.dtype, device=x.device 89 | ) 90 | binary_tensor = torch.floor(random_tensor) 91 | return torch.mul(torch.div(x, self.keep_prob), binary_tensor) 92 | else: 93 | return x 94 | 95 | 96 | class SamePaddingConv2d(nn.Module): 97 | def __init__( 98 | self, 99 | in_spatial_shape, 100 | in_channels, 101 | out_channels, 102 | kernel_size, 103 | stride, 104 | dilation=1, 105 | enforce_in_spatial_shape=False, 106 | **kwargs 107 | ): 108 | super(SamePaddingConv2d, self).__init__() 109 | 110 | self._in_spatial_shape = _pair(in_spatial_shape) 111 | # e.g. throw exception if input spatial shape does not match in_spatial_shape 112 | # when calling self.forward() 113 | self.enforce_in_spatial_shape = enforce_in_spatial_shape 114 | kernel_size = _pair(kernel_size) 115 | stride = _pair(stride) 116 | dilation = _pair(dilation) 117 | 118 | in_height, in_width = self._in_spatial_shape 119 | filter_height, filter_width = kernel_size 120 | stride_heigth, stride_width = stride 121 | dilation_height, dilation_width = dilation 122 | 123 | out_height = int(ceil(float(in_height) / float(stride_heigth))) 124 | out_width = int(ceil(float(in_width) / float(stride_width))) 125 | 126 | pad_along_height = max( 127 | (out_height - 1) * stride_heigth 128 | + filter_height 129 | + (filter_height - 1) * (dilation_height - 1) 130 | - in_height, 131 | 0, 132 | ) 133 | pad_along_width = max( 134 | (out_width - 1) * stride_width 135 | + filter_width 136 | + (filter_width - 1) * (dilation_width - 1) 137 | - in_width, 138 | 0, 139 | ) 140 | 141 | pad_top = pad_along_height // 2 142 | pad_bottom = pad_along_height - pad_top 143 | pad_left = pad_along_width // 2 144 | pad_right = pad_along_width - pad_left 145 | 146 | paddings = (pad_left, pad_right, pad_top, pad_bottom) 147 | if any(p > 0 for p in paddings): 148 | self.zero_pad = nn.ZeroPad2d(paddings) 149 | else: 150 | self.zero_pad = None 151 | self.conv = nn.Conv2d( 152 | in_channels=in_channels, 153 | out_channels=out_channels, 154 | kernel_size=kernel_size, 155 | stride=stride, 156 | dilation=dilation, 157 | **kwargs 158 | ) 159 | 160 | self._out_spatial_shape = (out_height, out_width) 161 | 162 | @property 163 | def out_spatial_shape(self): 164 | return self._out_spatial_shape 165 | 166 | def check_spatial_shape(self, x): 167 | if ( 168 | x.size(2) != self._in_spatial_shape[0] 169 | or x.size(3) != self._in_spatial_shape[1] 170 | ): 171 | raise ValueError( 172 | "Expected input spatial shape {}, got {} instead".format( 173 | self._in_spatial_shape, x.shape[2:] 174 | ) 175 | ) 176 | 177 | def forward(self, x): 178 | if self.enforce_in_spatial_shape: 179 | self.check_spatial_shape(x) 180 | if self.zero_pad is not None: 181 | x = self.zero_pad(x) 182 | x = self.conv(x) 183 | return x 184 | 185 | 186 | class SqueezeExcitate(nn.Module): 187 | def __init__(self, in_channels, se_size, activation=None): 188 | super(SqueezeExcitate, self).__init__() 189 | self.dim_reduce = nn.Conv2d( 190 | in_channels=in_channels, out_channels=se_size, kernel_size=1 191 | ) 192 | self.dim_restore = nn.Conv2d( 193 | in_channels=se_size, out_channels=in_channels, kernel_size=1 194 | ) 195 | self.activation = F.relu if activation is None else activation 196 | 197 | def forward(self, x): 198 | inp = x 199 | x = F.adaptive_avg_pool2d(x, (1, 1)) 200 | x = self.dim_reduce(x) 201 | x = self.activation(x) 202 | x = self.dim_restore(x) 203 | x = torch.sigmoid(x) 204 | return torch.mul(inp, x) 205 | 206 | 207 | class MBConvBlockV2(nn.Module): 208 | def __init__( 209 | self, 210 | in_channels, 211 | out_channels, 212 | kernel_size, 213 | stride, 214 | expansion_factor, 215 | act_fn, 216 | act_kwargs=None, 217 | bn_epsilon=None, 218 | bn_momentum=None, 219 | se_size=None, 220 | drop_connect_rate=None, 221 | bias=False, 222 | tf_style_conv=False, 223 | in_spatial_shape=None, 224 | ): 225 | 226 | super().__init__() 227 | 228 | if act_kwargs is None: 229 | act_kwargs = {} 230 | exp_channels = in_channels * expansion_factor 231 | 232 | self.ops_lst = [] 233 | 234 | # expansion convolution 235 | if expansion_factor != 1: 236 | self.expand_conv = nn.Conv2d( 237 | in_channels=in_channels, 238 | out_channels=exp_channels, 239 | kernel_size=1, 240 | bias=bias, 241 | ) 242 | 243 | self.expand_bn = nn.BatchNorm2d( 244 | num_features=exp_channels, eps=bn_epsilon, momentum=bn_momentum 245 | ) 246 | 247 | self.expand_act = get_activation(act_fn, **act_kwargs) 248 | self.ops_lst.extend([self.expand_conv, self.expand_bn, self.expand_act]) 249 | 250 | # depth-wise convolution 251 | if tf_style_conv: 252 | self.dp_conv = SamePaddingConv2d( 253 | in_spatial_shape=in_spatial_shape, 254 | in_channels=exp_channels, 255 | out_channels=exp_channels, 256 | kernel_size=kernel_size, 257 | stride=stride, 258 | groups=exp_channels, 259 | bias=bias, 260 | ) 261 | self.out_spatial_shape = self.dp_conv.out_spatial_shape 262 | else: 263 | self.dp_conv = nn.Conv2d( 264 | in_channels=exp_channels, 265 | out_channels=exp_channels, 266 | kernel_size=kernel_size, 267 | stride=stride, 268 | padding=1, 269 | groups=exp_channels, 270 | bias=bias, 271 | ) 272 | self.out_spatial_shape = torch_conv_out_spatial_shape( 273 | in_spatial_shape, kernel_size, stride 274 | ) 275 | 276 | self.dp_bn = nn.BatchNorm2d( 277 | num_features=exp_channels, eps=bn_epsilon, momentum=bn_momentum 278 | ) 279 | 280 | self.dp_act = get_activation(act_fn, **act_kwargs) 281 | self.ops_lst.extend([self.dp_conv, self.dp_bn, self.dp_act]) 282 | 283 | # Squeeze and Excitate 284 | if se_size is not None: 285 | self.se = SqueezeExcitate( 286 | exp_channels, se_size, activation=get_activation(act_fn, **act_kwargs) 287 | ) 288 | self.ops_lst.append(self.se) 289 | 290 | # projection layer 291 | self.project_conv = nn.Conv2d( 292 | in_channels=exp_channels, 293 | out_channels=out_channels, 294 | kernel_size=1, 295 | bias=bias, 296 | ) 297 | 298 | self.project_bn = nn.BatchNorm2d( 299 | num_features=out_channels, eps=bn_epsilon, momentum=bn_momentum 300 | ) 301 | 302 | # no activation function in projection layer 303 | 304 | self.ops_lst.extend([self.project_conv, self.project_bn]) 305 | 306 | self.skip_enabled = in_channels == out_channels and stride == 1 307 | 308 | if self.skip_enabled and drop_connect_rate is not None: 309 | self.drop_connect = DropConnect(drop_connect_rate) 310 | self.ops_lst.append(self.drop_connect) 311 | 312 | def forward(self, x): 313 | inp = x 314 | for op in self.ops_lst: 315 | x = op(x) 316 | if self.skip_enabled: 317 | return x + inp 318 | else: 319 | return x 320 | 321 | 322 | class FusedMBConvBlockV2(nn.Module): 323 | def __init__( 324 | self, 325 | in_channels, 326 | out_channels, 327 | kernel_size, 328 | stride, 329 | expansion_factor, 330 | act_fn, 331 | act_kwargs=None, 332 | bn_epsilon=None, 333 | bn_momentum=None, 334 | se_size=None, 335 | drop_connect_rate=None, 336 | bias=False, 337 | tf_style_conv=False, 338 | in_spatial_shape=None, 339 | ): 340 | 341 | super().__init__() 342 | 343 | if act_kwargs is None: 344 | act_kwargs = {} 345 | exp_channels = in_channels * expansion_factor 346 | 347 | self.ops_lst = [] 348 | 349 | # expansion convolution 350 | expansion_out_shape = in_spatial_shape 351 | if expansion_factor != 1: 352 | if tf_style_conv: 353 | self.expand_conv = SamePaddingConv2d( 354 | in_spatial_shape=in_spatial_shape, 355 | in_channels=in_channels, 356 | out_channels=exp_channels, 357 | kernel_size=kernel_size, 358 | stride=stride, 359 | bias=bias, 360 | ) 361 | expansion_out_shape = self.expand_conv.out_spatial_shape 362 | else: 363 | self.expand_conv = nn.Conv2d( 364 | in_channels=in_channels, 365 | out_channels=exp_channels, 366 | kernel_size=kernel_size, 367 | padding=1, 368 | stride=stride, 369 | bias=bias, 370 | ) 371 | expansion_out_shape = torch_conv_out_spatial_shape( 372 | in_spatial_shape, kernel_size, stride 373 | ) 374 | 375 | self.expand_bn = nn.BatchNorm2d( 376 | num_features=exp_channels, eps=bn_epsilon, momentum=bn_momentum 377 | ) 378 | 379 | self.expand_act = get_activation(act_fn, **act_kwargs) 380 | self.ops_lst.extend([self.expand_conv, self.expand_bn, self.expand_act]) 381 | 382 | # Squeeze and Excitate 383 | if se_size is not None: 384 | self.se = SqueezeExcitate( 385 | exp_channels, se_size, activation=get_activation(act_fn, **act_kwargs) 386 | ) 387 | self.ops_lst.append(self.se) 388 | 389 | # projection layer 390 | kernel_size = 1 if expansion_factor != 1 else kernel_size 391 | stride = 1 if expansion_factor != 1 else stride 392 | if tf_style_conv: 393 | self.project_conv = SamePaddingConv2d( 394 | in_spatial_shape=expansion_out_shape, 395 | in_channels=exp_channels, 396 | out_channels=out_channels, 397 | kernel_size=kernel_size, 398 | stride=stride, 399 | bias=bias, 400 | ) 401 | self.out_spatial_shape = self.project_conv.out_spatial_shape 402 | else: 403 | self.project_conv = nn.Conv2d( 404 | in_channels=exp_channels, 405 | out_channels=out_channels, 406 | kernel_size=kernel_size, 407 | stride=stride, 408 | padding=1 if kernel_size > 1 else 0, 409 | bias=bias, 410 | ) 411 | self.out_spatial_shape = torch_conv_out_spatial_shape( 412 | expansion_out_shape, kernel_size, stride 413 | ) 414 | 415 | self.project_bn = nn.BatchNorm2d( 416 | num_features=out_channels, eps=bn_epsilon, momentum=bn_momentum 417 | ) 418 | 419 | self.ops_lst.extend([self.project_conv, self.project_bn]) 420 | 421 | if expansion_factor == 1: 422 | self.project_act = get_activation(act_fn, **act_kwargs) 423 | self.ops_lst.append(self.project_act) 424 | 425 | self.skip_enabled = in_channels == out_channels and stride == 1 426 | 427 | if self.skip_enabled and drop_connect_rate is not None: 428 | self.drop_connect = DropConnect(drop_connect_rate) 429 | self.ops_lst.append(self.drop_connect) 430 | 431 | def forward(self, x): 432 | inp = x 433 | for op in self.ops_lst: 434 | x = op(x) 435 | if self.skip_enabled: 436 | return x + inp 437 | else: 438 | return x 439 | 440 | 441 | class EfficientNetV2(nn.Module): 442 | _models = { 443 | "b0": { 444 | "num_repeat": [1, 2, 2, 3, 5, 8], 445 | "kernel_size": [3, 3, 3, 3, 3, 3], 446 | "stride": [1, 2, 2, 2, 1, 2], 447 | "expand_ratio": [1, 4, 4, 4, 6, 6], 448 | "in_channel": [32, 16, 32, 48, 96, 112], 449 | "out_channel": [16, 32, 48, 96, 112, 192], 450 | "se_ratio": [None, None, None, 0.25, 0.25, 0.25], 451 | "conv_type": [1, 1, 1, 0, 0, 0], 452 | "is_feature_stage": [False, True, True, False, True, True], 453 | "width_coefficient": 1.0, 454 | "depth_coefficient": 1.0, 455 | "train_size": 192, 456 | "eval_size": 224, 457 | "dropout": 0.2, 458 | "weight_url": "https://api.onedrive.com/v1.0/shares/u!aHR0cHM6Ly8xZHJ2Lm1zL3UvcyFBdGlRcHc5VGNjZmlnUVBhWkZRcWNXR3dINmRLP2U9UUI5ZndH/root/content", 459 | "model_name": "efficientnet_v2_b0_21k_ft1k-a91e14c5.pth", 460 | }, 461 | "b1": { 462 | "num_repeat": [1, 2, 2, 3, 5, 8], 463 | "kernel_size": [3, 3, 3, 3, 3, 3], 464 | "stride": [1, 2, 2, 2, 1, 2], 465 | "expand_ratio": [1, 4, 4, 4, 6, 6], 466 | "in_channel": [32, 16, 32, 48, 96, 112], 467 | "out_channel": [16, 32, 48, 96, 112, 192], 468 | "se_ratio": [None, None, None, 0.25, 0.25, 0.25], 469 | "conv_type": [1, 1, 1, 0, 0, 0], 470 | "is_feature_stage": [False, True, True, False, True, True], 471 | "width_coefficient": 1.0, 472 | "depth_coefficient": 1.1, 473 | "train_size": 192, 474 | "eval_size": 240, 475 | "dropout": 0.2, 476 | "weight_url": "https://api.onedrive.com/v1.0/shares/u!aHR0cHM6Ly8xZHJ2Lm1zL3UvcyFBdGlRcHc5VGNjZmlnUVJnVGV5UndSY2J2amwtP2U9dTBiV1lO/root/content", 477 | "model_name": "efficientnet_v2_b1_21k_ft1k-58f4fb47.pth", 478 | }, 479 | "b2": { 480 | "num_repeat": [1, 2, 2, 3, 5, 8], 481 | "kernel_size": [3, 3, 3, 3, 3, 3], 482 | "stride": [1, 2, 2, 2, 1, 2], 483 | "expand_ratio": [1, 4, 4, 4, 6, 6], 484 | "in_channel": [32, 16, 32, 48, 96, 112], 485 | "out_channel": [16, 32, 48, 96, 112, 192], 486 | "se_ratio": [None, None, None, 0.25, 0.25, 0.25], 487 | "conv_type": [1, 1, 1, 0, 0, 0], 488 | "is_feature_stage": [False, True, True, False, True, True], 489 | "width_coefficient": 1.1, 490 | "depth_coefficient": 1.2, 491 | "train_size": 208, 492 | "eval_size": 260, 493 | "dropout": 0.3, 494 | "weight_url": "https://api.onedrive.com/v1.0/shares/u!aHR0cHM6Ly8xZHJ2Lm1zL3UvcyFBdGlRcHc5VGNjZmlnUVY4M2NySVFZbU41X0tGP2U9ZERZVmxK/root/content", 495 | "model_name": "efficientnet_v2_b2_21k_ft1k-db4ac0ee.pth", 496 | }, 497 | "b3": { 498 | "num_repeat": [1, 2, 2, 3, 5, 8], 499 | "kernel_size": [3, 3, 3, 3, 3, 3], 500 | "stride": [1, 2, 2, 2, 1, 2], 501 | "expand_ratio": [1, 4, 4, 4, 6, 6], 502 | "in_channel": [32, 16, 32, 48, 96, 112], 503 | "out_channel": [16, 32, 48, 96, 112, 192], 504 | "se_ratio": [None, None, None, 0.25, 0.25, 0.25], 505 | "conv_type": [1, 1, 1, 0, 0, 0], 506 | "is_feature_stage": [False, True, True, False, True, True], 507 | "width_coefficient": 1.2, 508 | "depth_coefficient": 1.4, 509 | "train_size": 240, 510 | "eval_size": 300, 511 | "dropout": 0.3, 512 | "weight_url": "https://api.onedrive.com/v1.0/shares/u!aHR0cHM6Ly8xZHJ2Lm1zL3UvcyFBdGlRcHc5VGNjZmlnUVpkamdZUzhhaDdtTTZLP2U9anA4VWN2/root/content", 513 | "model_name": "efficientnet_v2_b3_21k_ft1k-3da5874c.pth", 514 | }, 515 | "s": { 516 | "num_repeat": [2, 4, 4, 6, 9, 15], 517 | "kernel_size": [3, 3, 3, 3, 3, 3], 518 | "stride": [1, 2, 2, 2, 1, 2], 519 | "expand_ratio": [1, 4, 4, 4, 6, 6], 520 | "in_channel": [24, 24, 48, 64, 128, 160], 521 | "out_channel": [24, 48, 64, 128, 160, 256], 522 | "se_ratio": [None, None, None, 0.25, 0.25, 0.25], 523 | "conv_type": [1, 1, 1, 0, 0, 0], 524 | "is_feature_stage": [False, True, True, False, True, True], 525 | "width_coefficient": 1.0, 526 | "depth_coefficient": 1.0, 527 | "train_size": 300, 528 | "eval_size": 384, 529 | "dropout": 0.2, 530 | "weight_url": "https://api.onedrive.com/v1.0/shares/u!aHR0cHM6Ly8xZHJ2Lm1zL3UvcyFBdGlRcHc5VGNjZmllbFF5VWJOZzd0cmhBbm8/root/content", 531 | "model_name": "efficientnet_v2_s_21k_ft1k-dbb43f38.pth", 532 | }, 533 | "m": { 534 | "num_repeat": [2, 5, 5, 7, 14, 18, 5], 535 | "kernel_size": [3, 3, 3, 3, 3, 3, 3], 536 | "stride": [1, 2, 2, 2, 1, 2, 1], 537 | "expand_ratio": [1, 4, 4, 4, 6, 6, 6], 538 | "in_channel": [24, 24, 48, 64, 128, 160, 224], 539 | "out_channel": [24, 48, 64, 128, 160, 224, 512], 540 | "se_ratio": [None, None, None, 0.25, 0.25, 0.25, 0.25], 541 | "conv_type": [1, 1, 1, 0, 0, 0, 0], 542 | "is_feature_stage": [False, True, True, False, True, False, True], 543 | "width_coefficient": 1.0, 544 | "depth_coefficient": 1.0, 545 | "train_size": 384, 546 | "eval_size": 480, 547 | "dropout": 0.3, 548 | "weight_url": "https://api.onedrive.com/v1.0/shares/u!aHR0cHM6Ly8xZHJ2Lm1zL3UvcyFBdGlRcHc5VGNjZmllN1ZDazRFb0o1bnlyNUE/root/content", 549 | "model_name": "efficientnet_v2_m_21k_ft1k-da8e56c0.pth", 550 | }, 551 | "l": { 552 | "num_repeat": [4, 7, 7, 10, 19, 25, 7], 553 | "kernel_size": [3, 3, 3, 3, 3, 3, 3], 554 | "stride": [1, 2, 2, 2, 1, 2, 1], 555 | "expand_ratio": [1, 4, 4, 4, 6, 6, 6], 556 | "in_channel": [32, 32, 64, 96, 192, 224, 384], 557 | "out_channel": [32, 64, 96, 192, 224, 384, 640], 558 | "se_ratio": [None, None, None, 0.25, 0.25, 0.25, 0.25], 559 | "conv_type": [1, 1, 1, 0, 0, 0, 0], 560 | "is_feature_stage": [False, True, True, False, True, False, True], 561 | "feature_stages": [1, 2, 4, 6], 562 | "width_coefficient": 1.0, 563 | "depth_coefficient": 1.0, 564 | "train_size": 384, 565 | "eval_size": 480, 566 | "dropout": 0.4, 567 | "weight_url": "https://api.onedrive.com/v1.0/shares/u!aHR0cHM6Ly8xZHJ2Lm1zL3UvcyFBdGlRcHc5VGNjZmlmcmIyRHEtQTBhUTBhWVE/root/content", 568 | "model_name": "efficientnet_v2_l_21k_ft1k-08121eee.pth", 569 | }, 570 | "xl": { 571 | "num_repeat": [4, 8, 8, 16, 24, 32, 8], 572 | "kernel_size": [3, 3, 3, 3, 3, 3, 3], 573 | "stride": [1, 2, 2, 2, 1, 2, 1], 574 | "expand_ratio": [1, 4, 4, 4, 6, 6, 6], 575 | "in_channel": [32, 32, 64, 96, 192, 256, 512], 576 | "out_channel": [32, 64, 96, 192, 256, 512, 640], 577 | "se_ratio": [None, None, None, 0.25, 0.25, 0.25, 0.25], 578 | "conv_type": [1, 1, 1, 0, 0, 0, 0], 579 | "is_feature_stage": [False, True, True, False, True, False, True], 580 | "feature_stages": [1, 2, 4, 6], 581 | "width_coefficient": 1.0, 582 | "depth_coefficient": 1.0, 583 | "train_size": 384, 584 | "eval_size": 512, 585 | "dropout": 0.4, 586 | "weight_url": "https://api.onedrive.com/v1.0/shares/u!aHR0cHM6Ly8xZHJ2Lm1zL3UvcyFBdGlRcHc5VGNjZmlmVXQtRHJLa21taUkxWkE/root/content", 587 | "model_name": "efficientnet_v2_xl_21k_ft1k-1fcc9744.pth", 588 | }, 589 | } 590 | 591 | def __init__( 592 | self, 593 | model_name, 594 | in_channels=3, 595 | n_classes=1000, 596 | tf_style_conv=False, 597 | in_spatial_shape=None, 598 | activation="silu", 599 | activation_kwargs=None, 600 | bias=False, 601 | drop_connect_rate=0.2, 602 | dropout_rate=None, 603 | bn_epsilon=1e-3, 604 | bn_momentum=0.01, 605 | pretrained=False, 606 | progress=False, 607 | ): 608 | super().__init__() 609 | 610 | self.blocks = nn.ModuleList() 611 | self.model_name = model_name 612 | self.cfg = self._models[model_name] 613 | 614 | if tf_style_conv and in_spatial_shape is None: 615 | in_spatial_shape = self.cfg["eval_size"] 616 | 617 | activation_kwargs = {} if activation_kwargs is None else activation_kwargs 618 | dropout_rate = self.cfg["dropout"] if dropout_rate is None else dropout_rate 619 | _input_ch = in_channels 620 | 621 | self.feature_block_ids = [] 622 | 623 | # stem 624 | if tf_style_conv: 625 | self.stem_conv = SamePaddingConv2d( 626 | in_spatial_shape=in_spatial_shape, 627 | in_channels=in_channels, 628 | out_channels=round_filters( 629 | self.cfg["in_channel"][0], self.cfg["width_coefficient"] 630 | ), 631 | kernel_size=3, 632 | stride=2, 633 | bias=bias, 634 | ) 635 | in_spatial_shape = self.stem_conv.out_spatial_shape 636 | else: 637 | self.stem_conv = nn.Conv2d( 638 | in_channels=in_channels, 639 | out_channels=round_filters( 640 | self.cfg["in_channel"][0], self.cfg["width_coefficient"] 641 | ), 642 | kernel_size=3, 643 | stride=2, 644 | padding=1, 645 | bias=bias, 646 | ) 647 | 648 | self.stem_bn = nn.BatchNorm2d( 649 | num_features=round_filters( 650 | self.cfg["in_channel"][0], self.cfg["width_coefficient"] 651 | ), 652 | eps=bn_epsilon, 653 | momentum=bn_momentum, 654 | ) 655 | 656 | self.stem_act = get_activation(activation, **activation_kwargs) 657 | 658 | drop_connect_rates = self.get_dropconnect_rates(drop_connect_rate) 659 | 660 | stages = zip( 661 | *[ 662 | self.cfg[x] 663 | for x in [ 664 | "num_repeat", 665 | "kernel_size", 666 | "stride", 667 | "expand_ratio", 668 | "in_channel", 669 | "out_channel", 670 | "se_ratio", 671 | "conv_type", 672 | "is_feature_stage", 673 | ] 674 | ] 675 | ) 676 | 677 | idx = 0 678 | 679 | for stage_args in stages: 680 | ( 681 | num_repeat, 682 | kernel_size, 683 | stride, 684 | expand_ratio, 685 | in_channels, 686 | out_channels, 687 | se_ratio, 688 | conv_type, 689 | is_feature_stage, 690 | ) = stage_args 691 | 692 | in_channels = round_filters(in_channels, self.cfg["width_coefficient"]) 693 | out_channels = round_filters(out_channels, self.cfg["width_coefficient"]) 694 | num_repeat = round_repeats(num_repeat, self.cfg["depth_coefficient"]) 695 | 696 | conv_block = MBConvBlockV2 if conv_type == 0 else FusedMBConvBlockV2 697 | 698 | for _ in range(num_repeat): 699 | se_size = ( 700 | None if se_ratio is None else max(1, int(in_channels * se_ratio)) 701 | ) 702 | _b = conv_block( 703 | in_channels=in_channels, 704 | out_channels=out_channels, 705 | kernel_size=kernel_size, 706 | stride=stride, 707 | expansion_factor=expand_ratio, 708 | act_fn=activation, 709 | act_kwargs=activation_kwargs, 710 | bn_epsilon=bn_epsilon, 711 | bn_momentum=bn_momentum, 712 | se_size=se_size, 713 | drop_connect_rate=drop_connect_rates[idx], 714 | bias=bias, 715 | tf_style_conv=tf_style_conv, 716 | in_spatial_shape=in_spatial_shape, 717 | ) 718 | self.blocks.append(_b) 719 | idx += 1 720 | if tf_style_conv: 721 | in_spatial_shape = _b.out_spatial_shape 722 | in_channels = out_channels 723 | stride = 1 724 | 725 | if is_feature_stage: 726 | self.feature_block_ids.append(idx - 1) 727 | 728 | head_conv_out_channels = round_filters(1280, self.cfg["width_coefficient"]) 729 | 730 | self.head_conv = nn.Conv2d( 731 | in_channels=in_channels, 732 | out_channels=head_conv_out_channels, 733 | kernel_size=1, 734 | bias=bias, 735 | ) 736 | self.head_bn = nn.BatchNorm2d( 737 | num_features=head_conv_out_channels, eps=bn_epsilon, momentum=bn_momentum 738 | ) 739 | self.head_act = get_activation(activation, **activation_kwargs) 740 | 741 | self.dropout = nn.Dropout(p=dropout_rate) 742 | 743 | self.avpool = nn.AdaptiveAvgPool2d((1, 1)) 744 | self.fc = nn.Linear(head_conv_out_channels, n_classes) 745 | 746 | if pretrained: 747 | self._load_state(_input_ch, n_classes, progress, tf_style_conv) 748 | 749 | return 750 | 751 | def _load_state(self, in_channels, n_classes, progress, tf_style_conv): 752 | state_dict = model_zoo.load_url( 753 | self.cfg["weight_url"], progress=progress, file_name=self.cfg["model_name"] 754 | ) 755 | 756 | strict = True 757 | 758 | if not tf_style_conv: 759 | state_dict = OrderedDict( 760 | [ 761 | (k.replace(".conv.", "."), v) if ".conv." in k else (k, v) 762 | for k, v in state_dict.items() 763 | ] 764 | ) 765 | 766 | if in_channels != 3: 767 | if tf_style_conv: 768 | state_dict.pop("stem_conv.conv.weight") 769 | else: 770 | state_dict.pop("stem_conv.weight") 771 | strict = False 772 | 773 | if n_classes != 1000: 774 | state_dict.pop("fc.weight") 775 | state_dict.pop("fc.bias") 776 | strict = False 777 | 778 | self.load_state_dict(state_dict, strict=strict) 779 | print("Model weights loaded successfully.") 780 | 781 | def get_dropconnect_rates(self, drop_connect_rate): 782 | nr = self.cfg["num_repeat"] 783 | dc = self.cfg["depth_coefficient"] 784 | total = sum(round_repeats(nr[i], dc) for i in range(len(nr))) 785 | return [drop_connect_rate * i / total for i in range(total)] 786 | 787 | def get_features(self, x): 788 | x = self.stem_act(self.stem_bn(self.stem_conv(x))) 789 | 790 | features = [] 791 | feat_idx = 0 792 | for block_idx, block in enumerate(self.blocks): 793 | x = block(x) 794 | if block_idx == self.feature_block_ids[feat_idx]: 795 | features.append(x) 796 | feat_idx += 1 797 | 798 | return features 799 | 800 | def forward(self, x): 801 | x = self.stem_act(self.stem_bn(self.stem_conv(x))) 802 | for block in self.blocks: 803 | x = block(x) 804 | x = self.head_act(self.head_bn(self.head_conv(x))) 805 | x = self.dropout(torch.flatten(self.avpool(x), 1)) 806 | x = self.fc(x) 807 | 808 | return x 809 | -------------------------------------------------------------------------------- /utils/mobilenet_v3.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | """ 11 | Creates a MobileNetV3 Model as defined in: 12 | Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, Hartwig Adam. (2019). 13 | Searching for MobileNetV3 14 | arXiv preprint arXiv:1905.02244. 15 | """ 16 | 17 | import torch.nn as nn 18 | import math 19 | 20 | 21 | __all__ = ['mobilenetv3_large', 'mobilenetv3_small'] 22 | 23 | 24 | def _make_divisible(v, divisor, min_value=None): 25 | """ 26 | This function is taken from the original tf repo. 27 | It ensures that all layers have a channel number that is divisible by 8 28 | It can be seen here: 29 | https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py 30 | :param v: 31 | :param divisor: 32 | :param min_value: 33 | :return: 34 | """ 35 | if min_value is None: 36 | min_value = divisor 37 | new_v = max(min_value, int(v + divisor / 2) // divisor * divisor) 38 | # Make sure that round down does not go down by more than 10%. 39 | if new_v < 0.9 * v: 40 | new_v += divisor 41 | return new_v 42 | 43 | 44 | class h_sigmoid(nn.Module): 45 | def __init__(self, inplace=True): 46 | super(h_sigmoid, self).__init__() 47 | self.relu = nn.ReLU6(inplace=inplace) 48 | 49 | def forward(self, x): 50 | return self.relu(x + 3) / 6 51 | 52 | 53 | class h_swish(nn.Module): 54 | def __init__(self, inplace=True): 55 | super(h_swish, self).__init__() 56 | self.sigmoid = h_sigmoid(inplace=inplace) 57 | 58 | def forward(self, x): 59 | return x * self.sigmoid(x) 60 | 61 | 62 | class SELayer(nn.Module): 63 | def __init__(self, channel, reduction=4): 64 | super(SELayer, self).__init__() 65 | self.avg_pool = nn.AdaptiveAvgPool2d(1) 66 | self.fc = nn.Sequential( 67 | nn.Linear(channel, _make_divisible(channel // reduction, 8)), 68 | nn.ReLU(inplace=True), 69 | nn.Linear(_make_divisible(channel // reduction, 8), channel), 70 | h_sigmoid() 71 | ) 72 | 73 | def forward(self, x): 74 | b, c, _, _ = x.size() 75 | y = self.avg_pool(x).view(b, c) 76 | y = self.fc(y).view(b, c, 1, 1) 77 | return x * y 78 | 79 | 80 | def conv_3x3_bn(inp, oup, stride): 81 | return nn.Sequential( 82 | nn.Conv2d(inp, oup, 3, stride, 1, bias=False), 83 | nn.BatchNorm2d(oup), 84 | h_swish() 85 | ) 86 | 87 | 88 | def conv_1x1_bn(inp, oup): 89 | return nn.Sequential( 90 | nn.Conv2d(inp, oup, 1, 1, 0, bias=False), 91 | nn.BatchNorm2d(oup), 92 | h_swish() 93 | ) 94 | 95 | 96 | class InvertedResidual(nn.Module): 97 | def __init__(self, inp, hidden_dim, oup, kernel_size, stride, use_se, use_hs): 98 | super(InvertedResidual, self).__init__() 99 | assert stride in [1, 2] 100 | 101 | self.identity = stride == 1 and inp == oup 102 | 103 | if inp == hidden_dim: 104 | self.conv = nn.Sequential( 105 | # dw 106 | nn.Conv2d(hidden_dim, hidden_dim, kernel_size, stride, (kernel_size - 1) // 2, groups=hidden_dim, bias=False), 107 | nn.BatchNorm2d(hidden_dim), 108 | h_swish() if use_hs else nn.ReLU(inplace=True), 109 | # Squeeze-and-Excite 110 | SELayer(hidden_dim) if use_se else nn.Identity(), 111 | # pw-linear 112 | nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False), 113 | nn.BatchNorm2d(oup), 114 | ) 115 | else: 116 | self.conv = nn.Sequential( 117 | # pw 118 | nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False), 119 | nn.BatchNorm2d(hidden_dim), 120 | h_swish() if use_hs else nn.ReLU(inplace=True), 121 | # dw 122 | nn.Conv2d(hidden_dim, hidden_dim, kernel_size, stride, (kernel_size - 1) // 2, groups=hidden_dim, bias=False), 123 | nn.BatchNorm2d(hidden_dim), 124 | # Squeeze-and-Excite 125 | SELayer(hidden_dim) if use_se else nn.Identity(), 126 | h_swish() if use_hs else nn.ReLU(inplace=True), 127 | # pw-linear 128 | nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False), 129 | nn.BatchNorm2d(oup), 130 | ) 131 | 132 | def forward(self, x): 133 | if self.identity: 134 | return x + self.conv(x) 135 | else: 136 | return self.conv(x) 137 | 138 | 139 | class MobileNetV3(nn.Module): 140 | def __init__(self, cfgs, mode,input_f=3, num_classes=1000, width_mult=1.): 141 | super(MobileNetV3, self).__init__() 142 | # setting of inverted residual blocks 143 | self.cfgs = cfgs 144 | assert mode in ['large', 'small'] 145 | 146 | # building first layer 147 | input_channel = _make_divisible(16 * width_mult, 8) 148 | layers = [conv_3x3_bn(input_f, input_channel, 2)] 149 | # building inverted residual blocks 150 | block = InvertedResidual 151 | for k, t, c, use_se, use_hs, s in self.cfgs: 152 | output_channel = _make_divisible(c * width_mult, 8) 153 | exp_size = _make_divisible(input_channel * t, 8) 154 | layers.append(block(input_channel, exp_size, output_channel, k, s, use_se, use_hs)) 155 | input_channel = output_channel 156 | self.features = nn.Sequential(*layers) 157 | # building last several layers 158 | self.conv = conv_1x1_bn(input_channel, exp_size) 159 | self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) 160 | output_channel = {'large': 1280, 'small': 1024} 161 | output_channel = _make_divisible(output_channel[mode] * width_mult, 8) if width_mult > 1.0 else output_channel[mode] 162 | self.classifier = nn.Sequential( 163 | nn.Linear(exp_size, output_channel), 164 | h_swish(), 165 | nn.Dropout(0.2), 166 | nn.Linear(output_channel, num_classes), 167 | ) 168 | 169 | self._initialize_weights() 170 | 171 | def forward(self, x): 172 | x = self.features(x) 173 | x = self.conv(x) 174 | x = self.avgpool(x) 175 | x = x.view(x.size(0), -1) 176 | x = self.classifier(x) 177 | return x 178 | 179 | def _initialize_weights(self): 180 | for m in self.modules(): 181 | if isinstance(m, nn.Conv2d): 182 | n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels 183 | m.weight.data.normal_(0, math.sqrt(2. / n)) 184 | if m.bias is not None: 185 | m.bias.data.zero_() 186 | elif isinstance(m, nn.BatchNorm2d): 187 | m.weight.data.fill_(1) 188 | m.bias.data.zero_() 189 | elif isinstance(m, nn.Linear): 190 | m.weight.data.normal_(0, 0.01) 191 | m.bias.data.zero_() 192 | 193 | 194 | def mobilenetv3_large(**kwargs): 195 | """ 196 | Constructs a MobileNetV3-Large model 197 | """ 198 | cfgs = [ 199 | # k, t, c, SE, HS, s 200 | [3, 1, 16, 0, 0, 1], 201 | [3, 4, 24, 0, 0, 2], 202 | [3, 3, 24, 0, 0, 1], 203 | [5, 3, 40, 1, 0, 2], 204 | [5, 3, 40, 1, 0, 1], 205 | [5, 3, 40, 1, 0, 1], 206 | [3, 6, 80, 0, 1, 2], 207 | [3, 2.5, 80, 0, 1, 1], 208 | [3, 2.3, 80, 0, 1, 1], 209 | [3, 2.3, 80, 0, 1, 1], 210 | [3, 6, 112, 1, 1, 1], 211 | [3, 6, 112, 1, 1, 1], 212 | [5, 6, 160, 1, 1, 2], 213 | [5, 6, 160, 1, 1, 1], 214 | [5, 6, 160, 1, 1, 1] 215 | ] 216 | return MobileNetV3(cfgs, mode='large', **kwargs) 217 | 218 | 219 | 220 | 221 | 222 | 223 | 224 | 225 | 226 | def mobilenetv3_small(input_f = 3,num_classes=1000, **kwargs): 227 | """ 228 | Constructs a MobileNetV3-Small model 229 | """ 230 | cfgs = [ 231 | # k, t, c, SE, HS, s 232 | [3, 1, 16, 1, 0, 2], 233 | [3, 4.5, 24, 0, 0, 2], 234 | [3, 3.67, 24, 0, 0, 1], 235 | [5, 4, 40, 1, 1, 2], 236 | [5, 6, 40, 1, 1, 1], 237 | [5, 6, 40, 1, 1, 1], 238 | [5, 3, 48, 1, 1, 1], 239 | [5, 3, 48, 1, 1, 1], 240 | [5, 6, 96, 1, 1, 2], 241 | [5, 6, 96, 1, 1, 1], 242 | [5, 6, 96, 1, 1, 1], 243 | ] 244 | 245 | return MobileNetV3(cfgs,input_f=input_f,num_classes=num_classes, mode='small', **kwargs) -------------------------------------------------------------------------------- /utils/modules.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | import torch 11 | import torch.nn as nn 12 | 13 | 14 | class IBN(nn.Module): 15 | r"""Instance-Batch Normalization layer from 16 | `"Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net" 17 | ` 18 | Args: 19 | planes (int): Number of channels for the input tensor 20 | ratio (float): Ratio of instance normalization in the IBN layer 21 | """ 22 | 23 | def __init__(self, planes, ratio=0.5): 24 | super(IBN, self).__init__() 25 | self.half = int(planes * ratio) 26 | self.IN = nn.InstanceNorm2d(self.half, affine=True) 27 | self.BN = nn.BatchNorm2d(planes - self.half) 28 | 29 | def forward(self, x): 30 | split = torch.split(x, self.half, 1) 31 | out1 = self.IN(split[0].contiguous()) 32 | out2 = self.BN(split[1].contiguous()) 33 | out = torch.cat((out1, out2), 1) 34 | return out 35 | 36 | 37 | class SELayer(nn.Module): 38 | def __init__(self, channel, reduction=16): 39 | super(SELayer, self).__init__() 40 | self.avg_pool = nn.AdaptiveAvgPool2d(1) 41 | self.fc = nn.Sequential( 42 | nn.Linear(channel, int(channel / reduction), bias=False), 43 | nn.ReLU(inplace=True), 44 | nn.Linear(int(channel / reduction), channel, bias=False), 45 | nn.Sigmoid(), 46 | ) 47 | 48 | def forward(self, x): 49 | b, c, _, _ = x.size() 50 | y = self.avg_pool(x).view(b, c) 51 | y = self.fc(y).view(b, c, 1, 1) 52 | return x * y.expand_as(x) 53 | -------------------------------------------------------------------------------- /utils/networks.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | import torch 11 | import torch.nn as nn 12 | from torch.nn import init 13 | import functools 14 | from torch.optim import lr_scheduler 15 | 16 | 17 | ############################################################################### 18 | # Helper Functions 19 | ############################################################################### 20 | 21 | 22 | class Identity(nn.Module): 23 | def forward(self, x): 24 | return x 25 | 26 | 27 | def get_norm_layer(norm_type="instance"): 28 | """Return a normalization layer 29 | Parameters: 30 | norm_type (str) -- the name of the normalization layer: batch | instance | none 31 | For BatchNorm, we use learnable affine parameters and track running statistics (mean/stddev). 32 | For InstanceNorm, we do not use learnable affine parameters. We do not track running statistics. 33 | """ 34 | if norm_type == "batch": 35 | norm_layer = functools.partial( 36 | nn.BatchNorm2d, affine=True, track_running_stats=True 37 | ) 38 | elif norm_type == "instance": 39 | norm_layer = functools.partial( 40 | nn.InstanceNorm2d, affine=False, track_running_stats=False 41 | ) 42 | elif norm_type == "none": 43 | 44 | def norm_layer(x): 45 | return Identity() 46 | 47 | else: 48 | raise NotImplementedError("normalization layer [%s] is not found" % norm_type) 49 | return norm_layer 50 | 51 | 52 | def get_scheduler(optimizer, opt): 53 | """Return a learning rate scheduler 54 | Parameters: 55 | optimizer -- the optimizer of the network 56 | opt (option class) -- stores all the experiment flags; needs to be a subclass of BaseOptions． 57 | opt.lr_policy is the name of learning rate policy: linear | step | plateau | cosine 58 | For 'linear', we keep the same learning rate for the first epochs 59 | and linearly decay the rate to zero over the next epochs. 60 | For other schedulers (step, plateau, and cosine), we use the default PyTorch schedulers. 61 | See https://pytorch.org/docs/stable/optim.html for more details. 62 | """ 63 | if opt.lr_policy == "linear": 64 | 65 | def lambda_rule(epoch): 66 | lr_l = 1.0 - max(0, epoch + opt.epoch_count - opt.n_epochs) / float( 67 | opt.n_epochs_decay + 1 68 | ) 69 | return lr_l 70 | 71 | scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda_rule) 72 | elif opt.lr_policy == "step": 73 | scheduler = lr_scheduler.StepLR( 74 | optimizer, step_size=opt.lr_decay_iters, gamma=0.1 75 | ) 76 | elif opt.lr_policy == "plateau": 77 | scheduler = lr_scheduler.ReduceLROnPlateau( 78 | optimizer, mode="min", factor=0.2, threshold=0.01, patience=5 79 | ) 80 | elif opt.lr_policy == "cosine": 81 | scheduler = lr_scheduler.CosineAnnealingLR( 82 | optimizer, T_max=opt.n_epochs, eta_min=0 83 | ) 84 | else: 85 | return NotImplementedError( 86 | "learning rate policy [%s] is not implemented", opt.lr_policy 87 | ) 88 | return scheduler 89 | 90 | 91 | def init_weights(net, init_type="normal", init_gain=0.02): 92 | """Initialize network weights. 93 | Parameters: 94 | net (network) -- network to be initialized 95 | init_type (str) -- the name of an initialization method: normal | xavier | kaiming | orthogonal 96 | init_gain (float) -- scaling factor for normal, xavier and orthogonal. 97 | We use 'normal' in the original pix2pix and CycleGAN paper. But xavier and kaiming might 98 | work better for some applications. Feel free to try yourself. 99 | """ 100 | 101 | def init_func(m): # define the initialization function 102 | classname = m.__class__.__name__ 103 | if hasattr(m, "weight") and ( 104 | classname.find("Conv") != -1 or classname.find("Linear") != -1 105 | ): 106 | if init_type == "normal": 107 | init.normal_(m.weight.data, 0.0, init_gain) 108 | elif init_type == "xavier": 109 | init.xavier_normal_(m.weight.data, gain=init_gain) 110 | elif init_type == "kaiming": 111 | init.kaiming_normal_(m.weight.data, a=0, mode="fan_in") 112 | elif init_type == "orthogonal": 113 | init.orthogonal_(m.weight.data, gain=init_gain) 114 | else: 115 | raise NotImplementedError( 116 | "initialization method [%s] is not implemented" % init_type 117 | ) 118 | if hasattr(m, "bias") and m.bias is not None: 119 | init.constant_(m.bias.data, 0.0) 120 | elif ( 121 | classname.find("BatchNorm2d") != -1 122 | ): # BatchNorm Layer's weight is not a matrix; only normal distribution applies. 123 | init.normal_(m.weight.data, 1.0, init_gain) 124 | init.constant_(m.bias.data, 0.0) 125 | 126 | print("initialize network with %s" % init_type) 127 | net.apply(init_func) # apply the initialization function 128 | 129 | 130 | def init_net(net, init_type="normal", init_gain=0.02, gpu_ids=[]): 131 | """Initialize a network: 1. register CPU/GPU device (with multi-GPU support); 2. initialize the network weights 132 | Parameters: 133 | net (network) -- the network to be initialized 134 | init_type (str) -- the name of an initialization method: normal | xavier | kaiming | orthogonal 135 | gain (float) -- scaling factor for normal, xavier and orthogonal. 136 | gpu_ids (int list) -- which GPUs the network runs on: e.g., 0,1,2 137 | Return an initialized network. 138 | """ 139 | if len(gpu_ids) > 0: 140 | assert torch.cuda.is_available() 141 | net.to(gpu_ids[0]) 142 | net = torch.nn.DataParallel(net, gpu_ids) # multi-GPUs 143 | init_weights(net, init_type, init_gain=init_gain) 144 | return net 145 | 146 | 147 | def define_G( 148 | input_nc, 149 | output_nc, 150 | ngf, 151 | netG, 152 | norm="batch", 153 | use_dropout=False, 154 | init_type="normal", 155 | init_gain=0.02, 156 | gpu_ids=[], 157 | ): 158 | """Create a generator 159 | Parameters: 160 | input_nc (int) -- the number of channels in input images 161 | output_nc (int) -- the number of channels in output images 162 | ngf (int) -- the number of filters in the last conv layer 163 | netG (str) -- the architecture's name: resnet_9blocks | resnet_6blocks | unet_256 | unet_128 164 | norm (str) -- the name of normalization layers used in the network: batch | instance | none 165 | use_dropout (bool) -- if use dropout layers. 166 | init_type (str) -- the name of our initialization method. 167 | init_gain (float) -- scaling factor for normal, xavier and orthogonal. 168 | gpu_ids (int list) -- which GPUs the network runs on: e.g., 0,1,2 169 | Returns a generator 170 | Our current implementation provides two types of generators: 171 | U-Net: [unet_128] (for 128x128 input images) and [unet_256] (for 256x256 input images) 172 | The original U-Net paper: https://arxiv.org/abs/1505.04597 173 | Resnet-based generator: [resnet_6blocks] (with 6 Resnet blocks) and [resnet_9blocks] (with 9 Resnet blocks) 174 | Resnet-based generator consists of several Resnet blocks between a few downsampling/upsampling operations. 175 | We adapt Torch code from Justin Johnson's neural style transfer project (https://github.com/jcjohnson/fast-neural-style). 176 | The generator has been initialized by . It uses RELU for non-linearity. 177 | """ 178 | net = None 179 | norm_layer = get_norm_layer(norm_type=norm) 180 | 181 | if netG == "resnet_9blocks": 182 | net = ResnetGenerator( 183 | input_nc, 184 | output_nc, 185 | ngf, 186 | norm_layer=norm_layer, 187 | use_dropout=use_dropout, 188 | n_blocks=9, 189 | ) 190 | elif netG == "resnet_6blocks": 191 | net = ResnetGenerator( 192 | input_nc, 193 | output_nc, 194 | ngf, 195 | norm_layer=norm_layer, 196 | use_dropout=use_dropout, 197 | n_blocks=6, 198 | ) 199 | elif netG == "unet_128": 200 | net = UnetGenerator( 201 | input_nc, output_nc, 7, ngf, norm_layer=norm_layer, use_dropout=use_dropout 202 | ) 203 | elif netG == "unet_256": 204 | net = UnetGenerator( 205 | input_nc, output_nc, 8, ngf, norm_layer=norm_layer, use_dropout=use_dropout 206 | ) 207 | else: 208 | raise NotImplementedError("Generator model name [%s] is not recognized" % netG) 209 | return init_net(net, init_type, init_gain, gpu_ids) 210 | 211 | 212 | def define_D( 213 | input_nc, 214 | ndf, 215 | netD, 216 | n_layers_D=3, 217 | norm="batch", 218 | init_type="normal", 219 | init_gain=0.02, 220 | gpu_ids=[], 221 | ): 222 | """Create a discriminator 223 | Parameters: 224 | input_nc (int) -- the number of channels in input images 225 | ndf (int) -- the number of filters in the first conv layer 226 | netD (str) -- the architecture's name: basic | n_layers | pixel 227 | n_layers_D (int) -- the number of conv layers in the discriminator; effective when netD=='n_layers' 228 | norm (str) -- the type of normalization layers used in the network. 229 | init_type (str) -- the name of the initialization method. 230 | init_gain (float) -- scaling factor for normal, xavier and orthogonal. 231 | gpu_ids (int list) -- which GPUs the network runs on: e.g., 0,1,2 232 | Returns a discriminator 233 | Our current implementation provides three types of discriminators: 234 | [basic]: 'PatchGAN' classifier described in the original pix2pix paper. 235 | It can classify whether 70×70 overlapping patches are real or fake. 236 | Such a patch-level discriminator architecture has fewer parameters 237 | than a full-image discriminator and can work on arbitrarily-sized images 238 | in a fully convolutional fashion. 239 | [n_layers]: With this mode, you can specify the number of conv layers in the discriminator 240 | with the parameter (default=3 as used in [basic] (PatchGAN).) 241 | [pixel]: 1x1 PixelGAN discriminator can classify whether a pixel is real or not. 242 | It encourages greater color diversity but has no effect on spatial statistics. 243 | The discriminator has been initialized by . It uses Leakly RELU for non-linearity. 244 | """ 245 | net = None 246 | norm_layer = get_norm_layer(norm_type=norm) 247 | 248 | if netD == "basic": # default PatchGAN classifier 249 | net = NLayerDiscriminator(input_nc, ndf, n_layers=3, norm_layer=norm_layer) 250 | elif netD == "n_layers": # more options 251 | net = NLayerDiscriminator(input_nc, ndf, n_layers_D, norm_layer=norm_layer) 252 | elif netD == "pixel": # classify if each pixel is real or fake 253 | net = PixelDiscriminator(input_nc, ndf, norm_layer=norm_layer) 254 | else: 255 | raise NotImplementedError( 256 | "Discriminator model name [%s] is not recognized" % netD 257 | ) 258 | return init_net(net, init_type, init_gain, gpu_ids) 259 | 260 | 261 | ############################################################################## 262 | # Classes 263 | ############################################################################## 264 | class GANLoss(nn.Module): 265 | """Define different GAN objectives. 266 | The GANLoss class abstracts away the need to create the target label tensor 267 | that has the same size as the input. 268 | """ 269 | 270 | def __init__( 271 | self, 272 | gan_mode, 273 | gan_loss_mask=False, 274 | target_real_label=1.0, 275 | target_fake_label=0.0, 276 | ): 277 | """Initialize the GANLoss class. 278 | Parameters: 279 | gan_mode (str) - - the type of GAN objective. It currently supports vanilla, lsgan, and wgangp. 280 | target_real_label (bool) - - label for a real image 281 | target_fake_label (bool) - - label of a fake image 282 | Note: Do not use sigmoid as the last layer of Discriminator. 283 | LSGAN needs no sigmoid. vanilla GANs will handle it with BCEWithLogitsLoss. 284 | """ 285 | super(GANLoss, self).__init__() 286 | self.register_buffer("real_label", torch.tensor(target_real_label)) 287 | self.register_buffer("fake_label", torch.tensor(target_fake_label)) 288 | self.gan_mode = gan_mode 289 | self.gan_loss_mask = gan_loss_mask 290 | if gan_mode == "lsgan": 291 | self.loss = nn.MSELoss() 292 | elif gan_mode == "vanilla": 293 | self.loss = nn.BCEWithLogitsLoss() 294 | elif gan_mode in ["wgangp"]: 295 | self.loss = None 296 | else: 297 | raise NotImplementedError("gan mode %s not implemented" % gan_mode) 298 | 299 | def get_target_tensor(self, prediction, target_is_real, mask=None): 300 | """Create label tensors with the same size as the input. 301 | Parameters: 302 | prediction (tensor) - - tpyically the prediction from a discriminator 303 | target_is_real (bool) - - if the ground truth label is for real images or fake images 304 | Returns: 305 | A label tensor filled with ground truth label, and with the size of the input 306 | """ 307 | 308 | if target_is_real: 309 | target_tensor = self.real_label 310 | return target_tensor.expand_as(prediction) 311 | 312 | else: 313 | if self.gan_loss_mask: 314 | return 1 - mask 315 | else: 316 | target_tensor = self.fake_label 317 | return target_tensor.expand_as(prediction) 318 | 319 | def __call__(self, prediction, target_is_real, mask=None): 320 | """Calculate loss given Discriminator's output and grount truth labels. 321 | Parameters: 322 | prediction (tensor) - - tpyically the prediction output from a discriminator 323 | target_is_real (bool) - - if the ground truth label is for real images or fake images 324 | Returns: 325 | the calculated loss. 326 | """ 327 | if self.gan_mode in ["lsgan", "vanilla"]: 328 | target_tensor = self.get_target_tensor(prediction, target_is_real, mask) 329 | loss = self.loss(prediction, target_tensor) 330 | elif self.gan_mode == "wgangp": 331 | if target_is_real: 332 | loss = -prediction.mean() 333 | else: 334 | loss = prediction.mean() 335 | return loss 336 | 337 | 338 | def cal_gradient_penalty( 339 | netD, real_data, fake_data, device, type="mixed", constant=1.0, lambda_gp=10.0 340 | ): 341 | """Calculate the gradient penalty loss, used in WGAN-GP paper https://arxiv.org/abs/1704.00028 342 | Arguments: 343 | netD (network) -- discriminator network 344 | real_data (tensor array) -- real images 345 | fake_data (tensor array) -- generated images from the generator 346 | device (str) -- GPU / CPU: from torch.device('cuda:{}'.format(self.gpu_ids[0])) if self.gpu_ids else torch.device('cpu') 347 | type (str) -- if we mix real and fake data or not [real | fake | mixed]. 348 | constant (float) -- the constant used in formula ( ||gradient||_2 - constant)^2 349 | lambda_gp (float) -- weight for this loss 350 | Returns the gradient penalty loss 351 | """ 352 | if lambda_gp > 0.0: 353 | if ( 354 | type == "real" 355 | ): # either use real images, fake images, or a linear interpolation of two. 356 | interpolatesv = real_data 357 | elif type == "fake": 358 | interpolatesv = fake_data 359 | elif type == "mixed": 360 | alpha = torch.rand(real_data.shape[0], 1, device=device) 361 | alpha = ( 362 | alpha.expand( 363 | real_data.shape[0], real_data.nelement() // real_data.shape[0] 364 | ) 365 | .contiguous() 366 | .view(*real_data.shape) 367 | ) 368 | interpolatesv = alpha * real_data + ((1 - alpha) * fake_data) 369 | else: 370 | raise NotImplementedError("{} not implemented".format(type)) 371 | interpolatesv.requires_grad_(True) 372 | disc_interpolates = netD(interpolatesv) 373 | gradients = torch.autograd.grad( 374 | outputs=disc_interpolates, 375 | inputs=interpolatesv, 376 | grad_outputs=torch.ones(disc_interpolates.size()).to(device), 377 | create_graph=True, 378 | retain_graph=True, 379 | only_inputs=True, 380 | ) 381 | gradients = gradients[0].view(real_data.size(0), -1) # flat the data 382 | gradient_penalty = ( 383 | ((gradients + 1e-16).norm(2, dim=1) - constant) ** 2 384 | ).mean() * lambda_gp # added eps 385 | return gradient_penalty, gradients 386 | else: 387 | return 0.0, None 388 | 389 | 390 | class ResnetGenerator(nn.Module): 391 | """Resnet-based generator that consists of Resnet blocks between a few downsampling/upsampling operations. 392 | We adapt Torch code and idea from Justin Johnson's neural style transfer project(https://github.com/jcjohnson/fast-neural-style) 393 | """ 394 | 395 | def __init__( 396 | self, 397 | input_nc, 398 | output_nc, 399 | ngf=64, 400 | norm_layer=nn.BatchNorm2d, 401 | use_dropout=False, 402 | n_blocks=6, 403 | padding_type="reflect", 404 | ): 405 | """Construct a Resnet-based generator 406 | Parameters: 407 | input_nc (int) -- the number of channels in input images 408 | output_nc (int) -- the number of channels in output images 409 | ngf (int) -- the number of filters in the last conv layer 410 | norm_layer -- normalization layer 411 | use_dropout (bool) -- if use dropout layers 412 | n_blocks (int) -- the number of ResNet blocks 413 | padding_type (str) -- the name of padding layer in conv layers: reflect | replicate | zero 414 | """ 415 | assert n_blocks >= 0 416 | super(ResnetGenerator, self).__init__() 417 | if type(norm_layer) == functools.partial: 418 | use_bias = norm_layer.func == nn.InstanceNorm2d 419 | else: 420 | use_bias = norm_layer == nn.InstanceNorm2d 421 | 422 | model = [ 423 | nn.ReflectionPad2d(3), 424 | nn.Conv2d(input_nc, ngf, kernel_size=7, padding=0, bias=use_bias), 425 | norm_layer(ngf), 426 | nn.ReLU(True), 427 | ] 428 | 429 | n_downsampling = 2 430 | for i in range(n_downsampling): # add downsampling layers 431 | mult = 2**i 432 | model += [ 433 | nn.Conv2d( 434 | ngf * mult, 435 | ngf * mult * 2, 436 | kernel_size=3, 437 | stride=2, 438 | padding=1, 439 | bias=use_bias, 440 | ), 441 | norm_layer(ngf * mult * 2), 442 | nn.ReLU(True), 443 | ] 444 | 445 | mult = 2**n_downsampling 446 | for i in range(n_blocks): # add ResNet blocks 447 | 448 | model += [ 449 | ResnetBlock( 450 | ngf * mult, 451 | padding_type=padding_type, 452 | norm_layer=norm_layer, 453 | use_dropout=use_dropout, 454 | use_bias=use_bias, 455 | ) 456 | ] 457 | 458 | for i in range(n_downsampling): # add upsampling layers 459 | mult = 2 ** (n_downsampling - i) 460 | model += [ 461 | nn.ConvTranspose2d( 462 | ngf * mult, 463 | int(ngf * mult / 2), 464 | kernel_size=3, 465 | stride=2, 466 | padding=1, 467 | output_padding=1, 468 | bias=use_bias, 469 | ), 470 | norm_layer(int(ngf * mult / 2)), 471 | nn.ReLU(True), 472 | ] 473 | model += [nn.ReflectionPad2d(3)] 474 | model += [nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0)] 475 | model += [nn.Tanh()] 476 | 477 | self.model = nn.Sequential(*model) 478 | 479 | def forward(self, input): 480 | """Standard forward""" 481 | return self.model(input) 482 | 483 | 484 | class ResnetBlock(nn.Module): 485 | """Define a Resnet block""" 486 | 487 | def __init__(self, dim, padding_type, norm_layer, use_dropout, use_bias): 488 | """Initialize the Resnet block 489 | A resnet block is a conv block with skip connections 490 | We construct a conv block with build_conv_block function, 491 | and implement skip connections in function. 492 | Original Resnet paper: https://arxiv.org/pdf/1512.03385.pdf 493 | """ 494 | super(ResnetBlock, self).__init__() 495 | self.conv_block = self.build_conv_block( 496 | dim, padding_type, norm_layer, use_dropout, use_bias 497 | ) 498 | 499 | def build_conv_block(self, dim, padding_type, norm_layer, use_dropout, use_bias): 500 | """Construct a convolutional block. 501 | Parameters: 502 | dim (int) -- the number of channels in the conv layer. 503 | padding_type (str) -- the name of padding layer: reflect | replicate | zero 504 | norm_layer -- normalization layer 505 | use_dropout (bool) -- if use dropout layers. 506 | use_bias (bool) -- if the conv layer uses bias or not 507 | Returns a conv block (with a conv layer, a normalization layer, and a non-linearity layer (ReLU)) 508 | """ 509 | conv_block = [] 510 | p = 0 511 | if padding_type == "reflect": 512 | conv_block += [nn.ReflectionPad2d(1)] 513 | elif padding_type == "replicate": 514 | conv_block += [nn.ReplicationPad2d(1)] 515 | elif padding_type == "zero": 516 | p = 1 517 | else: 518 | raise NotImplementedError("padding [%s] is not implemented" % padding_type) 519 | 520 | conv_block += [ 521 | nn.Conv2d(dim, dim, kernel_size=3, padding=p, bias=use_bias), 522 | norm_layer(dim), 523 | nn.ReLU(True), 524 | ] 525 | if use_dropout: 526 | conv_block += [nn.Dropout(0.5)] 527 | 528 | p = 0 529 | if padding_type == "reflect": 530 | conv_block += [nn.ReflectionPad2d(1)] 531 | elif padding_type == "replicate": 532 | conv_block += [nn.ReplicationPad2d(1)] 533 | elif padding_type == "zero": 534 | p = 1 535 | else: 536 | raise NotImplementedError("padding [%s] is not implemented" % padding_type) 537 | conv_block += [ 538 | nn.Conv2d(dim, dim, kernel_size=3, padding=p, bias=use_bias), 539 | norm_layer(dim), 540 | ] 541 | 542 | return nn.Sequential(*conv_block) 543 | 544 | def forward(self, x): 545 | """Forward function (with skip connections)""" 546 | out = x + self.conv_block(x) # add skip connections 547 | return out 548 | 549 | 550 | class UnetGenerator(nn.Module): 551 | """Create a Unet-based generator""" 552 | 553 | def __init__( 554 | self, 555 | input_nc, 556 | output_nc, 557 | num_downs, 558 | ngf=64, 559 | norm_layer=nn.BatchNorm2d, 560 | use_dropout=False, 561 | ): 562 | """Construct a Unet generator 563 | Parameters: 564 | input_nc (int) -- the number of channels in input images 565 | output_nc (int) -- the number of channels in output images 566 | num_downs (int) -- the number of downsamplings in UNet. For example, # if |num_downs| == 7, 567 | image of size 128x128 will become of size 1x1 # at the bottleneck 568 | ngf (int) -- the number of filters in the last conv layer 569 | norm_layer -- normalization layer 570 | We construct the U-Net from the innermost layer to the outermost layer. 571 | It is a recursive process. 572 | """ 573 | super(UnetGenerator, self).__init__() 574 | # construct unet structure 575 | unet_block = UnetSkipConnectionBlock( 576 | ngf * 8, 577 | ngf * 8, 578 | input_nc=None, 579 | submodule=None, 580 | norm_layer=norm_layer, 581 | innermost=True, 582 | ) # add the innermost layer 583 | for i in range(num_downs - 5): # add intermediate layers with ngf * 8 filters 584 | unet_block = UnetSkipConnectionBlock( 585 | ngf * 8, 586 | ngf * 8, 587 | input_nc=None, 588 | submodule=unet_block, 589 | norm_layer=norm_layer, 590 | use_dropout=use_dropout, 591 | ) 592 | # gradually reduce the number of filters from ngf * 8 to ngf 593 | unet_block = UnetSkipConnectionBlock( 594 | ngf * 4, ngf * 8, input_nc=None, submodule=unet_block, norm_layer=norm_layer 595 | ) 596 | unet_block = UnetSkipConnectionBlock( 597 | ngf * 2, ngf * 4, input_nc=None, submodule=unet_block, norm_layer=norm_layer 598 | ) 599 | unet_block = UnetSkipConnectionBlock( 600 | ngf, ngf * 2, input_nc=None, submodule=unet_block, norm_layer=norm_layer 601 | ) 602 | self.model = UnetSkipConnectionBlock( 603 | output_nc, 604 | ngf, 605 | input_nc=input_nc, 606 | submodule=unet_block, 607 | outermost=True, 608 | norm_layer=norm_layer, 609 | ) # add the outermost layer 610 | 611 | def forward(self, input): 612 | """Standard forward""" 613 | return self.model(input) 614 | 615 | 616 | class UnetSkipConnectionBlock(nn.Module): 617 | """Defines the Unet submodule with skip connection. 618 | X -------------------identity---------------------- 619 | |-- downsampling -- |submodule| -- upsampling --| 620 | """ 621 | 622 | def __init__( 623 | self, 624 | outer_nc, 625 | inner_nc, 626 | input_nc=None, 627 | submodule=None, 628 | outermost=False, 629 | innermost=False, 630 | norm_layer=nn.BatchNorm2d, 631 | use_dropout=False, 632 | ): 633 | """Construct a Unet submodule with skip connections. 634 | Parameters: 635 | outer_nc (int) -- the number of filters in the outer conv layer 636 | inner_nc (int) -- the number of filters in the inner conv layer 637 | input_nc (int) -- the number of channels in input images/features 638 | submodule (UnetSkipConnectionBlock) -- previously defined submodules 639 | outermost (bool) -- if this module is the outermost module 640 | innermost (bool) -- if this module is the innermost module 641 | norm_layer -- normalization layer 642 | use_dropout (bool) -- if use dropout layers. 643 | """ 644 | super(UnetSkipConnectionBlock, self).__init__() 645 | self.outermost = outermost 646 | if type(norm_layer) == functools.partial: 647 | use_bias = norm_layer.func == nn.InstanceNorm2d 648 | else: 649 | use_bias = norm_layer == nn.InstanceNorm2d 650 | if input_nc is None: 651 | input_nc = outer_nc 652 | downconv = nn.Conv2d( 653 | input_nc, inner_nc, kernel_size=4, stride=2, padding=1, bias=use_bias 654 | ) 655 | downrelu = nn.LeakyReLU(0.2, True) 656 | downnorm = norm_layer(inner_nc) 657 | uprelu = nn.ReLU(True) 658 | upnorm = norm_layer(outer_nc) 659 | 660 | if outermost: 661 | upconv = nn.ConvTranspose2d( 662 | inner_nc * 2, outer_nc, kernel_size=4, stride=2, padding=1 663 | ) 664 | down = [downconv] 665 | up = [uprelu, upconv, nn.Tanh()] 666 | model = down + [submodule] + up 667 | elif innermost: 668 | upconv = nn.ConvTranspose2d( 669 | inner_nc, outer_nc, kernel_size=4, stride=2, padding=1, bias=use_bias 670 | ) 671 | down = [downrelu, downconv] 672 | up = [uprelu, upconv, upnorm] 673 | model = down + up 674 | else: 675 | upconv = nn.ConvTranspose2d( 676 | inner_nc * 2, 677 | outer_nc, 678 | kernel_size=4, 679 | stride=2, 680 | padding=1, 681 | bias=use_bias, 682 | ) 683 | down = [downrelu, downconv, downnorm] 684 | up = [uprelu, upconv, upnorm] 685 | 686 | if use_dropout: 687 | model = down + [submodule] + up + [nn.Dropout(0.5)] 688 | else: 689 | model = down + [submodule] + up 690 | 691 | self.model = nn.Sequential(*model) 692 | 693 | def forward(self, x): 694 | if self.outermost: 695 | return self.model(x) 696 | else: # add skip connections 697 | return torch.cat([x, self.model(x)], 1) 698 | 699 | 700 | class NLayerDiscriminator(nn.Module): 701 | """Defines a PatchGAN discriminator""" 702 | 703 | def __init__(self, input_nc, ndf=64, n_layers=3, norm_layer=nn.BatchNorm2d): 704 | """Construct a PatchGAN discriminator 705 | Parameters: 706 | input_nc (int) -- the number of channels in input images 707 | ndf (int) -- the number of filters in the last conv layer 708 | n_layers (int) -- the number of conv layers in the discriminator 709 | norm_layer -- normalization layer 710 | """ 711 | super(NLayerDiscriminator, self).__init__() 712 | if ( 713 | type(norm_layer) == functools.partial 714 | ): # no need to use bias as BatchNorm2d has affine parameters 715 | use_bias = norm_layer.func == nn.InstanceNorm2d 716 | else: 717 | use_bias = norm_layer == nn.InstanceNorm2d 718 | 719 | kw = 4 720 | padw = 1 721 | sequence = [ 722 | nn.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw), 723 | nn.LeakyReLU(0.2, True), 724 | ] 725 | nf_mult = 1 726 | nf_mult_prev = 1 727 | for n in range(1, n_layers): # gradually increase the number of filters 728 | nf_mult_prev = nf_mult 729 | nf_mult = min(2**n, 8) 730 | sequence += [ 731 | nn.Conv2d( 732 | ndf * nf_mult_prev, 733 | ndf * nf_mult, 734 | kernel_size=kw, 735 | stride=2, 736 | padding=padw, 737 | bias=use_bias, 738 | ), 739 | norm_layer(ndf * nf_mult), 740 | nn.LeakyReLU(0.2, True), 741 | ] 742 | 743 | nf_mult_prev = nf_mult 744 | nf_mult = min(2**n_layers, 8) 745 | sequence += [ 746 | nn.Conv2d( 747 | ndf * nf_mult_prev, 748 | ndf * nf_mult, 749 | kernel_size=kw, 750 | stride=1, 751 | padding=padw, 752 | bias=use_bias, 753 | ), 754 | norm_layer(ndf * nf_mult), 755 | nn.LeakyReLU(0.2, True), 756 | ] 757 | 758 | sequence += [ 759 | nn.Conv2d(ndf * nf_mult, 1, kernel_size=kw, stride=1, padding=padw) 760 | ] # output 1 channel prediction map 761 | self.model = nn.Sequential(*sequence) 762 | 763 | def forward(self, input): 764 | """Standard forward.""" 765 | return self.model(input) 766 | 767 | 768 | class PixelDiscriminator(nn.Module): 769 | """Defines a 1x1 PatchGAN discriminator (pixelGAN)""" 770 | 771 | def __init__(self, input_nc, ndf=64, norm_layer=nn.BatchNorm2d): 772 | """Construct a 1x1 PatchGAN discriminator 773 | Parameters: 774 | input_nc (int) -- the number of channels in input images 775 | ndf (int) -- the number of filters in the last conv layer 776 | norm_layer -- normalization layer 777 | """ 778 | super(PixelDiscriminator, self).__init__() 779 | if ( 780 | type(norm_layer) == functools.partial 781 | ): # no need to use bias as BatchNorm2d has affine parameters 782 | use_bias = norm_layer.func == nn.InstanceNorm2d 783 | else: 784 | use_bias = norm_layer == nn.InstanceNorm2d 785 | 786 | self.net = [ 787 | nn.Conv2d(input_nc, ndf, kernel_size=1, stride=1, padding=0), 788 | nn.LeakyReLU(0.2, True), 789 | nn.Conv2d(ndf, ndf * 2, kernel_size=1, stride=1, padding=0, bias=use_bias), 790 | norm_layer(ndf * 2), 791 | nn.LeakyReLU(0.2, True), 792 | nn.Conv2d(ndf * 2, 1, kernel_size=1, stride=1, padding=0, bias=use_bias), 793 | ] 794 | 795 | self.net = nn.Sequential(*self.net) 796 | 797 | def forward(self, input): 798 | """Standard forward.""" 799 | return self.net(input) 800 | -------------------------------------------------------------------------------- /utils/resnet.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | import torch.nn as nn 11 | import torch.utils.model_zoo as model_zoo 12 | import torch 13 | import torch.nn.functional as f 14 | # from vit_pytorch import ViT 15 | 16 | from utils.efficientnet_v2 import EfficientNetV2 17 | from utils.mobilenet_v3 import MobileNetV3, mobilenetv3_small 18 | 19 | __all__ = [ 20 | "ResNet", 21 | "resnet18", 22 | "resnet34", 23 | "resnet50", 24 | "resnet101", 25 | "resnet152", 26 | "resnext50_32x4d", 27 | "resnext101_32x8d", 28 | ] 29 | 30 | 31 | model_urls = { 32 | "resnet18": "https://download.pytorch.org/models/resnet18-5c106cde.pth", 33 | "resnet34": "https://download.pytorch.org/models/resnet34-333f7ec4.pth", 34 | "resnet50": "https://download.pytorch.org/models/resnet50-19c8e357.pth", 35 | "resnet101": "https://download.pytorch.org/models/resnet101-5d3b4d8f.pth", 36 | "resnet152": "https://download.pytorch.org/models/resnet152-b121ed2d.pth", 37 | } 38 | 39 | 40 | def conv3x3(in_planes, out_planes, stride=1, groups=1): 41 | """3x3 convolution with padding""" 42 | return nn.Conv2d( 43 | in_planes, 44 | out_planes, 45 | kernel_size=3, 46 | stride=stride, 47 | padding=1, 48 | groups=groups, 49 | bias=False, 50 | ) 51 | 52 | 53 | def conv1x1(in_planes, out_planes, stride=1): 54 | """1x1 convolution""" 55 | return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) 56 | 57 | 58 | class BasicBlock(nn.Module): 59 | expansion = 1 60 | 61 | def __init__( 62 | self, 63 | inplanes, 64 | planes, 65 | stride=1, 66 | downsample=None, 67 | groups=1, 68 | base_width=64, 69 | norm_layer=None, 70 | ): 71 | super(BasicBlock, self).__init__() 72 | if norm_layer is None: 73 | norm_layer = nn.BatchNorm2d 74 | if groups != 1 or base_width != 64: 75 | raise ValueError("BasicBlock only supports groups=1 and base_width=64") 76 | # Both self.conv1 and self.downsample layers downsample the input when stride != 1 77 | self.conv1 = conv3x3(inplanes, planes, stride) 78 | self.bn1 = norm_layer(planes) 79 | self.relu = nn.ReLU(inplace=False) 80 | self.conv2 = conv3x3(planes, planes) 81 | self.bn2 = norm_layer(planes) 82 | self.downsample = downsample 83 | self.stride = stride 84 | 85 | def forward(self, x): 86 | identity = x 87 | 88 | out = self.conv1(x) 89 | out = self.bn1(out) 90 | out = self.relu(out) 91 | 92 | out = self.conv2(out) 93 | out = self.bn2(out) 94 | 95 | if self.downsample is not None: 96 | identity = self.downsample(x) 97 | 98 | out = out + identity 99 | out = self.relu(out) 100 | 101 | return out 102 | 103 | 104 | class Bottleneck(nn.Module): 105 | expansion = 4 106 | 107 | def __init__( 108 | self, 109 | inplanes, 110 | planes, 111 | stride=1, 112 | downsample=None, 113 | groups=1, 114 | base_width=64, 115 | norm_layer=None, 116 | ): 117 | super(Bottleneck, self).__init__() 118 | if norm_layer is None: 119 | norm_layer = nn.BatchNorm2d 120 | width = int(planes * (base_width / 64.0)) * groups 121 | # Both self.conv2 and self.downsample layers downsample the input when stride != 1 122 | self.conv1 = conv1x1(inplanes, width) 123 | self.bn1 = norm_layer(width) 124 | self.conv2 = conv3x3(width, width, stride, groups) 125 | self.bn2 = norm_layer(width) 126 | self.conv3 = conv1x1(width, planes * self.expansion) 127 | self.bn3 = norm_layer(planes * self.expansion) 128 | self.relu = nn.ReLU(inplace=False) 129 | self.downsample = downsample 130 | self.stride = stride 131 | 132 | def forward(self, x): 133 | identity = x 134 | 135 | out = self.conv1(x) 136 | out = self.bn1(out) 137 | out = self.relu(out) 138 | 139 | out = self.conv2(out) 140 | out = self.bn2(out) 141 | out = self.relu(out) 142 | 143 | out = self.conv3(out) 144 | out = self.bn3(out) 145 | 146 | if self.downsample is not None: 147 | identity = self.downsample(x) 148 | 149 | out = out + identity 150 | out = self.relu(out) 151 | 152 | return out 153 | 154 | 155 | class ResNet(nn.Module): 156 | def __init__( 157 | self, 158 | block, 159 | layers, 160 | input_f=2, 161 | num_classes=1000, 162 | zero_init_residual=False, 163 | groups=1, 164 | width_per_group=64, 165 | norm_layer=None, 166 | sigmoid=False, 167 | ): 168 | super(ResNet, self).__init__() 169 | if norm_layer is None: 170 | norm_layer = nn.BatchNorm2d 171 | 172 | self.inplanes = 64 173 | self.groups = groups 174 | self.base_width = width_per_group 175 | print(input_f) 176 | self.conv1 = nn.Conv2d( 177 | input_f, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False 178 | ) 179 | self.bn1 = norm_layer(self.inplanes) 180 | self.relu = nn.ReLU(inplace=False) 181 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) 182 | self.layer1 = self._make_layer(block, 64, layers[0], norm_layer=norm_layer) 183 | self.layer2 = self._make_layer( 184 | block, 128, layers[1], stride=2, norm_layer=norm_layer 185 | ) 186 | self.layer3 = self._make_layer( 187 | block, 256, layers[2], stride=2, norm_layer=norm_layer 188 | ) 189 | self.layer4 = self._make_layer( 190 | block, 512, layers[3], stride=2, norm_layer=norm_layer 191 | ) 192 | self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) 193 | self.fc = nn.Linear(512 * block.expansion, num_classes) 194 | self.sigmoid = sigmoid 195 | for m in self.modules(): 196 | if isinstance(m, nn.Conv2d): 197 | nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu") 198 | elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): 199 | nn.init.constant_(m.weight, 1) 200 | nn.init.constant_(m.bias, 0) 201 | 202 | # Zero-initialize the last BN in each residual branch, 203 | # so that the residual branch starts with zeros, and each residual block behaves like an identity. 204 | # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 205 | if zero_init_residual: 206 | for m in self.modules(): 207 | if isinstance(m, Bottleneck): 208 | nn.init.constant_(m.bn3.weight, 0) 209 | elif isinstance(m, BasicBlock): 210 | nn.init.constant_(m.bn2.weight, 0) 211 | 212 | def _make_layer(self, block, planes, blocks, stride=1, norm_layer=None): 213 | if norm_layer is None: 214 | norm_layer = nn.BatchNorm2d 215 | downsample = None 216 | if stride != 1 or self.inplanes != planes * block.expansion: 217 | downsample = nn.Sequential( 218 | conv1x1(self.inplanes, planes * block.expansion, stride), 219 | norm_layer(planes * block.expansion), 220 | ) 221 | 222 | layers = [] 223 | layers.append( 224 | block( 225 | self.inplanes, 226 | planes, 227 | stride, 228 | downsample, 229 | self.groups, 230 | self.base_width, 231 | norm_layer, 232 | ) 233 | ) 234 | self.inplanes = planes * block.expansion 235 | for _ in range(1, blocks): 236 | layers.append( 237 | block( 238 | self.inplanes, 239 | planes, 240 | groups=self.groups, 241 | base_width=self.base_width, 242 | norm_layer=norm_layer, 243 | ) 244 | ) 245 | 246 | return nn.Sequential(*layers) 247 | 248 | def forward(self, x): 249 | x = self.conv1(x) 250 | x = self.bn1(x) 251 | x = self.relu(x) 252 | x = self.maxpool(x) 253 | 254 | x = self.layer1(x) 255 | x = self.layer2(x) 256 | x = self.layer3(x) 257 | x = self.layer4(x) 258 | 259 | x = self.avgpool(x) 260 | x1 = x.view(x.size(0), -1) 261 | x = self.fc(x1) 262 | if self.sigmoid: 263 | x = nn.Sigmoid()(x) 264 | else: 265 | pass 266 | # x = f.normalize(x, p=2, dim=1) 267 | return x, x1 268 | 269 | 270 | class ResNet_PIH(nn.Module): 271 | def __init__( 272 | self, 273 | block, 274 | layers, 275 | input_f=2, 276 | zero_init_residual=False, 277 | groups=1, 278 | width_per_group=64, 279 | norm_layer=None, 280 | ): 281 | super(ResNet_PIH, self).__init__() 282 | if norm_layer is None: 283 | norm_layer = nn.BatchNorm2d 284 | 285 | self.inplanes = 64 286 | self.groups = groups 287 | self.base_width = width_per_group 288 | print(input_f) 289 | self.conv1 = nn.Conv2d( 290 | input_f, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False 291 | ) 292 | self.bn1 = norm_layer(self.inplanes) 293 | self.relu = nn.ReLU(inplace=False) 294 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) 295 | self.layer1 = self._make_layer(block, 64, layers[0], norm_layer=norm_layer) 296 | self.layer2 = self._make_layer( 297 | block, 128, layers[1], stride=2, norm_layer=norm_layer 298 | ) 299 | self.layer3 = self._make_layer( 300 | block, 256, layers[2], stride=2, norm_layer=norm_layer 301 | ) 302 | self.layer4 = self._make_layer( 303 | block, 512, layers[3], stride=2, norm_layer=norm_layer 304 | ) 305 | self.avgpool = nn.AdaptiveAvgPool2d((3, 3)) 306 | for m in self.modules(): 307 | if isinstance(m, nn.Conv2d): 308 | nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu") 309 | elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): 310 | nn.init.constant_(m.weight, 1) 311 | nn.init.constant_(m.bias, 0) 312 | 313 | # Zero-initialize the last BN in each residual branch, 314 | # so that the residual branch starts with zeros, and each residual block behaves like an identity. 315 | # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 316 | if zero_init_residual: 317 | for m in self.modules(): 318 | if isinstance(m, Bottleneck): 319 | nn.init.constant_(m.bn3.weight, 0) 320 | elif isinstance(m, BasicBlock): 321 | nn.init.constant_(m.bn2.weight, 0) 322 | 323 | def _make_layer(self, block, planes, blocks, stride=1, norm_layer=None): 324 | if norm_layer is None: 325 | norm_layer = nn.BatchNorm2d 326 | downsample = None 327 | if stride != 1 or self.inplanes != planes * block.expansion: 328 | downsample = nn.Sequential( 329 | conv1x1(self.inplanes, planes * block.expansion, stride), 330 | norm_layer(planes * block.expansion), 331 | ) 332 | 333 | layers = [] 334 | layers.append( 335 | block( 336 | self.inplanes, 337 | planes, 338 | stride, 339 | downsample, 340 | self.groups, 341 | self.base_width, 342 | norm_layer, 343 | ) 344 | ) 345 | self.inplanes = planes * block.expansion 346 | for _ in range(1, blocks): 347 | layers.append( 348 | block( 349 | self.inplanes, 350 | planes, 351 | groups=self.groups, 352 | base_width=self.base_width, 353 | norm_layer=norm_layer, 354 | ) 355 | ) 356 | 357 | return nn.Sequential(*layers) 358 | 359 | def forward(self, x): 360 | x = self.conv1(x) 361 | x = self.bn1(x) 362 | x = self.relu(x) 363 | x = self.maxpool(x) 364 | 365 | x = self.layer1(x) 366 | x = self.layer2(x) 367 | x = self.layer3(x) 368 | x = self.layer4(x) 369 | 370 | x = self.avgpool(x) 371 | x = x.view(x.size(0), -1) 372 | 373 | return x 374 | 375 | 376 | 377 | 378 | 379 | class EffNetV2(nn.Module): 380 | def __init__( 381 | self, pretrained=False, input_f=7, num_classes=1000, sigmoid=False, **kwargs 382 | ): 383 | super(EffNetV2, self).__init__() 384 | self.num_classes = num_classes 385 | 386 | self.EffNet = EfficientNetV2( 387 | "s", in_channels=input_f, n_classes=self.num_classes, pretrained=False 388 | ) 389 | self.sigmoid = sigmoid 390 | 391 | def forward(self, x): 392 | 393 | feature_output = self.EffNet(x) 394 | if self.sigmoid: 395 | feature_output = nn.Sigmoid()(feature_output) 396 | 397 | return feature_output, 0 398 | 399 | 400 | 401 | 402 | class MobileNetV3(nn.Module): 403 | def __init__( 404 | self, pretrained=False, input_f=7, num_classes=1000, sigmoid=False, **kwargs 405 | ): 406 | super(MobileNetV3, self).__init__() 407 | self.num_classes = num_classes 408 | 409 | self.EffNet = mobilenetv3_small(input_f,num_classes) 410 | self.sigmoid = sigmoid 411 | 412 | def forward(self, x): 413 | 414 | feature_output = self.EffNet(x) 415 | if self.sigmoid: 416 | feature_output = nn.Sigmoid()(feature_output) 417 | 418 | return feature_output, 0 419 | 420 | 421 | class PIHNet(nn.Module): 422 | def __init__( 423 | self, pretrained=False, input_f=7, num_classes=1000, sigmoid=False, **kwargs 424 | ): 425 | super(PIHNet, self).__init__() 426 | self.model_bg = ResNet_PIH(BasicBlock, [3, 4, 6, 3], input_f=4, **kwargs) 427 | 428 | self.model_fg = ResNet_PIH(BasicBlock, [3, 4, 6, 3], input_f=4, **kwargs) 429 | 430 | self.model_cp = ResNet_PIH(BasicBlock, [3, 4, 6, 3], input_f=4, **kwargs) 431 | 432 | self.classifier = nn.Sequential( 433 | nn.Linear(512 * 3 * 3, 1024), 434 | nn.ReLU(True), 435 | nn.Linear(1024, num_classes), 436 | ) 437 | self.sigmoid = sigmoid 438 | 439 | def forward(self, x): 440 | input_image = x[:, :3, ...] 441 | bg_image = x[:, 3:6, ...] 442 | mask_image = x[:, 6:, ...] 443 | feature_bg = self.model_bg(torch.cat((1 - mask_image, bg_image), 1)) 444 | feature_fg = self.model_fg(torch.cat((mask_image, input_image * mask_image), 1)) 445 | feature_cp = self.model_cp(torch.cat((mask_image, input_image), 1)) 446 | 447 | feature_all = feature_bg + feature_fg + feature_cp 448 | feature_output = self.classifier(feature_all) 449 | if self.sigmoid: 450 | feature_output = nn.Sigmoid()(feature_output) 451 | return feature_output, 0 452 | 453 | 454 | def resnet18(pretrained=False, input_f=4, num_classes=1000, **kwargs): 455 | """Constructs a ResNet-18 model. 456 | Args: 457 | pretrained (bool): If True, returns a model pre-trained on ImageNet 458 | """ 459 | model = ResNet( 460 | BasicBlock, [2, 2, 2, 2], input_f=input_f, num_classes=num_classes, **kwargs 461 | ) 462 | if pretrained: 463 | model.load_state_dict(model_zoo.load_url(model_urls["resnet18"])) 464 | return model 465 | 466 | 467 | def resnet18_m(pretrained=False, num_classes=1000, **kwargs): 468 | """Constructs a ResNet-18 model. 469 | Args: 470 | pretrained (bool): If True, returns a model pre-trained on ImageNet 471 | """ 472 | model = ResNet( 473 | BasicBlock, [2, 2, 2, 2], input_f=1, num_classes=num_classes, **kwargs 474 | ) 475 | if pretrained: 476 | model.load_state_dict(model_zoo.load_url(model_urls["resnet18"])) 477 | return model 478 | 479 | 480 | def resnet34(pretrained=False, input_f=4, num_classes=1000, sigmoid=False, **kwargs): 481 | """Constructs a ResNet-34 model. 482 | Args: 483 | pretrained (bool): If True, returns a model pre-trained on ImageNet 484 | """ 485 | model = ResNet( 486 | BasicBlock, 487 | [3, 4, 6, 3], 488 | input_f=input_f, 489 | num_classes=num_classes, 490 | sigmoid=sigmoid, 491 | **kwargs 492 | ) 493 | if pretrained: 494 | model.load_state_dict(model_zoo.load_url(model_urls["resnet34"])) 495 | return model 496 | 497 | 498 | def resnet50(pretrained=False, input_f=4, num_classes=1000, sigmoid=False, **kwargs): 499 | """Constructs a ResNet-50 model. 500 | Args: 501 | pretrained (bool): If True, returns a model pre-trained on ImageNet 502 | """ 503 | model = ResNet( 504 | Bottleneck, 505 | [3, 4, 6, 3], 506 | num_classes=num_classes, 507 | input_f=input_f, 508 | sigmoid=sigmoid, 509 | **kwargs 510 | ) 511 | if pretrained: 512 | model.load_state_dict(model_zoo.load_url(model_urls["resnet50"])) 513 | return model 514 | 515 | 516 | def resnet101(pretrained=False, input_f=4, num_classes=1000, sigmoid=False, **kwargs): 517 | """Constructs a ResNet-101 model. 518 | Args: 519 | pretrained (bool): If True, returns a model pre-trained on ImageNet 520 | """ 521 | # print("Using Resnet 101") 522 | model = ResNet(Bottleneck, [3, 4, 23, 3], 523 | num_classes=num_classes, 524 | input_f=input_f, 525 | sigmoid=sigmoid, **kwargs) 526 | if pretrained: 527 | model.load_state_dict(model_zoo.load_url(model_urls["resnet101"])) 528 | return model 529 | 530 | 531 | def resnet152(pretrained=False, **kwargs): 532 | """Constructs a ResNet-152 model. 533 | Args: 534 | pretrained (bool): If True, returns a model pre-trained on ImageNet 535 | """ 536 | model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs) 537 | if pretrained: 538 | model.load_state_dict(model_zoo.load_url(model_urls["resnet152"])) 539 | return model 540 | 541 | 542 | def resnext50_32x4d(pretrained=False, **kwargs): 543 | model = ResNet(Bottleneck, [3, 4, 6, 3], groups=32, width_per_group=4, **kwargs) 544 | # if pretrained: 545 | # model.load_state_dict(model_zoo.load_url(model_urls['resnext50_32x4d'])) 546 | return model 547 | 548 | 549 | def resnext101_32x8d(pretrained=False, **kwargs): 550 | model = ResNet(Bottleneck, [3, 4, 23, 3], groups=32, width_per_group=8, **kwargs) 551 | # if pretrained: 552 | # model.load_state_dict(model_zoo.load_url(model_urls['resnext101_32x8d'])) 553 | return model 554 | -------------------------------------------------------------------------------- /utils/resnet_ibn.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Adobe. All rights reserved. 2 | # This file is licensed to you under the Apache License, Version 2.0 (the "License"); 3 | # you may not use this file except in compliance with the License. You may obtain a copy 4 | # of the License at http://www.apache.org/licenses/LICENSE-2.0 5 | 6 | # Unless required by applicable law or agreed to in writing, software distributed under 7 | # the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR REPRESENTATIONS 8 | # OF ANY KIND, either express or implied. See the License for the specific language 9 | # governing permissions and limitations under the License. 10 | import math 11 | import warnings 12 | 13 | import torch 14 | import torch.nn as nn 15 | 16 | from utils.modules import IBN 17 | 18 | 19 | __all__ = [ 20 | "ResNet_IBN", 21 | "resnet18_ibn_a", 22 | "resnet34_ibn_a", 23 | "resnet50_ibn_a", 24 | "resnet101_ibn_a", 25 | "resnet152_ibn_a", 26 | "resnet18_ibn_b", 27 | "resnet34_ibn_b", 28 | "resnet50_ibn_b", 29 | "resnet101_ibn_b", 30 | "resnet152_ibn_b", 31 | ] 32 | 33 | 34 | model_urls = { 35 | "resnet18_ibn_a": "https://github.com/XingangPan/IBN-Net/releases/download/v1.0/resnet18_ibn_a-2f571257.pth", 36 | "resnet34_ibn_a": "https://github.com/XingangPan/IBN-Net/releases/download/v1.0/resnet34_ibn_a-94bc1577.pth", 37 | "resnet50_ibn_a": "https://github.com/XingangPan/IBN-Net/releases/download/v1.0/resnet50_ibn_a-d9d0bb7b.pth", 38 | "resnet101_ibn_a": "https://github.com/XingangPan/IBN-Net/releases/download/v1.0/resnet101_ibn_a-59ea0ac6.pth", 39 | "resnet18_ibn_b": "https://github.com/XingangPan/IBN-Net/releases/download/v1.0/resnet18_ibn_b-bc2f3c11.pth", 40 | "resnet34_ibn_b": "https://github.com/XingangPan/IBN-Net/releases/download/v1.0/resnet34_ibn_b-04134c37.pth", 41 | "resnet50_ibn_b": "https://github.com/XingangPan/IBN-Net/releases/download/v1.0/resnet50_ibn_b-9ca61e85.pth", 42 | "resnet101_ibn_b": "https://github.com/XingangPan/IBN-Net/releases/download/v1.0/resnet101_ibn_b-c55f6dba.pth", 43 | } 44 | 45 | 46 | class BasicBlock_IBN(nn.Module): 47 | expansion = 1 48 | 49 | def __init__(self, inplanes, planes, ibn=None, stride=1, downsample=None): 50 | super(BasicBlock_IBN, self).__init__() 51 | self.conv1 = nn.Conv2d( 52 | inplanes, planes, kernel_size=3, stride=stride, padding=1, bias=False 53 | ) 54 | if ibn == "a": 55 | self.bn1 = IBN(planes) 56 | else: 57 | self.bn1 = nn.BatchNorm2d(planes) 58 | self.relu = nn.ReLU(inplace=True) 59 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1, bias=False) 60 | self.bn2 = nn.BatchNorm2d(planes) 61 | self.IN = nn.InstanceNorm2d(planes, affine=True) if ibn == "b" else None 62 | self.downsample = downsample 63 | self.stride = stride 64 | 65 | def forward(self, x): 66 | residual = x 67 | 68 | out = self.conv1(x) 69 | out = self.bn1(out) 70 | out = self.relu(out) 71 | 72 | out = self.conv2(out) 73 | out = self.bn2(out) 74 | 75 | if self.downsample is not None: 76 | residual = self.downsample(x) 77 | 78 | out += residual 79 | if self.IN is not None: 80 | out = self.IN(out) 81 | out = self.relu(out) 82 | 83 | return out 84 | 85 | 86 | class Bottleneck_IBN(nn.Module): 87 | expansion = 4 88 | 89 | def __init__(self, inplanes, planes, ibn=None, stride=1, downsample=None): 90 | super(Bottleneck_IBN, self).__init__() 91 | self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) 92 | if ibn == "a": 93 | self.bn1 = IBN(planes) 94 | else: 95 | self.bn1 = nn.BatchNorm2d(planes) 96 | self.conv2 = nn.Conv2d( 97 | planes, planes, kernel_size=3, stride=stride, padding=1, bias=False 98 | ) 99 | self.bn2 = nn.BatchNorm2d(planes) 100 | self.conv3 = nn.Conv2d( 101 | planes, planes * self.expansion, kernel_size=1, bias=False 102 | ) 103 | self.bn3 = nn.BatchNorm2d(planes * self.expansion) 104 | self.IN = nn.InstanceNorm2d(planes * 4, affine=True) if ibn == "b" else None 105 | self.relu = nn.ReLU(inplace=True) 106 | self.downsample = downsample 107 | self.stride = stride 108 | 109 | def forward(self, x): 110 | residual = x 111 | 112 | out = self.conv1(x) 113 | out = self.bn1(out) 114 | out = self.relu(out) 115 | 116 | out = self.conv2(out) 117 | out = self.bn2(out) 118 | out = self.relu(out) 119 | 120 | out = self.conv3(out) 121 | out = self.bn3(out) 122 | 123 | if self.downsample is not None: 124 | residual = self.downsample(x) 125 | 126 | out += residual 127 | if self.IN is not None: 128 | out = self.IN(out) 129 | out = self.relu(out) 130 | 131 | return out 132 | 133 | 134 | class ResNet_IBN(nn.Module): 135 | def __init__( 136 | self, 137 | block, 138 | layers, 139 | ibn_cfg=("a", "a", "a", None), 140 | input_f=7, 141 | num_classes=1000, 142 | sigmoid=False, 143 | ): 144 | self.inplanes = 64 145 | self.sigmoid = sigmoid 146 | super(ResNet_IBN, self).__init__() 147 | self.conv1 = nn.Conv2d( 148 | input_f, 64, kernel_size=7, stride=2, padding=3, bias=False 149 | ) 150 | if ibn_cfg[0] == "b": 151 | self.bn1 = nn.InstanceNorm2d(64, affine=True) 152 | else: 153 | self.bn1 = nn.BatchNorm2d(64) 154 | self.relu = nn.ReLU(inplace=True) 155 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) 156 | self.layer1 = self._make_layer(block, 64, layers[0], ibn=ibn_cfg[0]) 157 | self.layer2 = self._make_layer(block, 128, layers[1], stride=2, ibn=ibn_cfg[1]) 158 | self.layer3 = self._make_layer(block, 256, layers[2], stride=2, ibn=ibn_cfg[2]) 159 | self.layer4 = self._make_layer(block, 512, layers[3], stride=2, ibn=ibn_cfg[3]) 160 | self.avgpool = nn.AvgPool2d(7) 161 | self.fc = nn.Linear(512 * 4 * block.expansion, num_classes) 162 | 163 | for m in self.modules(): 164 | if isinstance(m, nn.Conv2d): 165 | n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels 166 | m.weight.data.normal_(0, math.sqrt(2.0 / n)) 167 | elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.InstanceNorm2d): 168 | m.weight.data.fill_(1) 169 | m.bias.data.zero_() 170 | 171 | def _make_layer(self, block, planes, blocks, stride=1, ibn=None): 172 | downsample = None 173 | if stride != 1 or self.inplanes != planes * block.expansion: 174 | downsample = nn.Sequential( 175 | nn.Conv2d( 176 | self.inplanes, 177 | planes * block.expansion, 178 | kernel_size=1, 179 | stride=stride, 180 | bias=False, 181 | ), 182 | nn.BatchNorm2d(planes * block.expansion), 183 | ) 184 | 185 | layers = [] 186 | layers.append( 187 | block( 188 | self.inplanes, planes, None if ibn == "b" else ibn, stride, downsample 189 | ) 190 | ) 191 | self.inplanes = planes * block.expansion 192 | for i in range(1, blocks): 193 | layers.append( 194 | block( 195 | self.inplanes, 196 | planes, 197 | None if (ibn == "b" and i < blocks - 1) else ibn, 198 | ) 199 | ) 200 | 201 | return nn.Sequential(*layers) 202 | 203 | def forward(self, x): 204 | x = self.conv1(x) 205 | x = self.bn1(x) 206 | x = self.relu(x) 207 | x = self.maxpool(x) 208 | 209 | x = self.layer1(x) 210 | x = self.layer2(x) 211 | x = self.layer3(x) 212 | x = self.layer4(x) 213 | 214 | x = self.avgpool(x) 215 | x = x.view(x.size(0), -1) 216 | x = self.fc(x) 217 | 218 | if self.sigmoid: 219 | x = nn.Sigmoid()(x) 220 | else: 221 | pass 222 | 223 | return x, 0 224 | 225 | 226 | def resnet18_ibn_a(pretrained=False, **kwargs): 227 | """Constructs a ResNet-18-IBN-a model. 228 | Args: 229 | pretrained (bool): If True, returns a model pre-trained on ImageNet 230 | """ 231 | model = ResNet_IBN( 232 | block=BasicBlock_IBN, 233 | layers=[2, 2, 2, 2], 234 | ibn_cfg=("a", "a", "a", None), 235 | **kwargs 236 | ) 237 | if pretrained: 238 | model.load_state_dict( 239 | torch.hub.load_state_dict_from_url(model_urls["resnet18_ibn_a"]) 240 | ) 241 | return model 242 | 243 | 244 | def resnet34_ibn_a(pretrained=False, **kwargs): 245 | """Constructs a ResNet-34-IBN-a model. 246 | Args: 247 | pretrained (bool): If True, returns a model pre-trained on ImageNet 248 | """ 249 | model = ResNet_IBN( 250 | block=BasicBlock_IBN, 251 | layers=[3, 4, 6, 3], 252 | ibn_cfg=("a", "a", "a", None), 253 | **kwargs 254 | ) 255 | if pretrained: 256 | model.load_state_dict( 257 | torch.hub.load_state_dict_from_url(model_urls["resnet34_ibn_a"]) 258 | ) 259 | return model 260 | 261 | 262 | def resnet50_ibn_a(pretrained=False, **kwargs): 263 | """Constructs a ResNet-50-IBN-a model. 264 | Args: 265 | pretrained (bool): If True, returns a model pre-trained on ImageNet 266 | """ 267 | model = ResNet_IBN( 268 | block=Bottleneck_IBN, 269 | layers=[3, 4, 6, 3], 270 | ibn_cfg=("a", "a", "a", None), 271 | **kwargs 272 | ) 273 | if pretrained: 274 | model.load_state_dict( 275 | torch.hub.load_state_dict_from_url(model_urls["resnet50_ibn_a"]) 276 | ) 277 | return model 278 | 279 | 280 | def resnet101_ibn_a(pretrained=False, **kwargs): 281 | """Constructs a ResNet-101-IBN-a model. 282 | Args: 283 | pretrained (bool): If True, returns a model pre-trained on ImageNet 284 | """ 285 | model = ResNet_IBN( 286 | block=Bottleneck_IBN, 287 | layers=[3, 4, 23, 3], 288 | ibn_cfg=("a", "a", "a", None), 289 | **kwargs 290 | ) 291 | if pretrained: 292 | model.load_state_dict( 293 | torch.hub.load_state_dict_from_url(model_urls["resnet101_ibn_a"]) 294 | ) 295 | return model 296 | 297 | 298 | def resnet152_ibn_a(pretrained=False, **kwargs): 299 | """Constructs a ResNet-152-IBN-a model. 300 | Args: 301 | pretrained (bool): If True, returns a model pre-trained on ImageNet 302 | """ 303 | model = ResNet_IBN( 304 | block=Bottleneck_IBN, 305 | layers=[3, 8, 36, 3], 306 | ibn_cfg=("a", "a", "a", None), 307 | **kwargs 308 | ) 309 | if pretrained: 310 | warnings.warn("Pretrained model not available for ResNet-152-IBN-a!") 311 | return model 312 | 313 | 314 | def resnet18_ibn_b(pretrained=False, **kwargs): 315 | """Constructs a ResNet-18-IBN-b model. 316 | Args: 317 | pretrained (bool): If True, returns a model pre-trained on ImageNet 318 | """ 319 | model = ResNet_IBN( 320 | block=BasicBlock_IBN, 321 | layers=[2, 2, 2, 2], 322 | ibn_cfg=("b", "b", None, None), 323 | **kwargs 324 | ) 325 | if pretrained: 326 | model.load_state_dict( 327 | torch.hub.load_state_dict_from_url(model_urls["resnet18_ibn_b"]) 328 | ) 329 | return model 330 | 331 | 332 | def resnet34_ibn_b(pretrained=False, **kwargs): 333 | """Constructs a ResNet-34-IBN-b model. 334 | Args: 335 | pretrained (bool): If True, returns a model pre-trained on ImageNet 336 | """ 337 | model = ResNet_IBN( 338 | block=BasicBlock_IBN, 339 | layers=[3, 4, 6, 3], 340 | ibn_cfg=("b", "b", None, None), 341 | **kwargs 342 | ) 343 | if pretrained: 344 | model.load_state_dict( 345 | torch.hub.load_state_dict_from_url(model_urls["resnet34_ibn_b"]) 346 | ) 347 | return model 348 | 349 | 350 | def resnet50_ibn_b( 351 | pretrained=False, input_f=4, num_classes=1000, sigmoid=False, **kwargs 352 | ): 353 | """Constructs a ResNet-50-IBN-b model. 354 | Args: 355 | pretrained (bool): If True, returns a model pre-trained on ImageNet 356 | """ 357 | model = ResNet_IBN( 358 | block=Bottleneck_IBN, 359 | layers=[3, 4, 6, 3], 360 | ibn_cfg=("b", "b", None, None), 361 | input_f=input_f, 362 | num_classes=num_classes, 363 | sigmoid=sigmoid, 364 | **kwargs 365 | ) 366 | return model 367 | 368 | 369 | def resnet101_ibn_b(pretrained=False, **kwargs): 370 | """Constructs a ResNet-101-IBN-b model. 371 | Args: 372 | pretrained (bool): If True, returns a model pre-trained on ImageNet 373 | """ 374 | model = ResNet_IBN( 375 | block=Bottleneck_IBN, 376 | layers=[3, 4, 23, 3], 377 | ibn_cfg=("b", "b", None, None), 378 | **kwargs 379 | ) 380 | if pretrained: 381 | model.load_state_dict( 382 | torch.hub.load_state_dict_from_url(model_urls["resnet101_ibn_b"]) 383 | ) 384 | return model 385 | 386 | 387 | def resnet152_ibn_b(pretrained=False, **kwargs): 388 | """Constructs a ResNet-152-IBN-b model. 389 | Args: 390 | pretrained (bool): If True, returns a model pre-trained on ImageNet 391 | """ 392 | model = ResNet_IBN( 393 | block=Bottleneck_IBN, 394 | layers=[3, 8, 36, 3], 395 | ibn_cfg=("b", "b", None, None), 396 | **kwargs 397 | ) 398 | if pretrained: 399 | warnings.warn("Pretrained model not available for ResNet-152-IBN-b!") 400 | return model 401 | -------------------------------------------------------------------------------- /utils/unet/__init__.py: -------------------------------------------------------------------------------- 1 | from .unet_model import UNet 2 | -------------------------------------------------------------------------------- /utils/unet/unet_model.py: -------------------------------------------------------------------------------- 1 | """ Full assembly of the parts to form the complete network """ 2 | 3 | from .unet_parts import * 4 | 5 | 6 | class UNet(nn.Module): 7 | def __init__(self, n_channels, n_classes, bilinear=False): 8 | super(UNet, self).__init__() 9 | self.n_channels = n_channels 10 | self.n_classes = n_classes 11 | self.bilinear = bilinear 12 | 13 | self.inc = DoubleConv(n_channels, 64) 14 | self.down1 = Down(64, 128) 15 | self.down2 = Down(128, 256) 16 | self.down3 = Down(256, 512) 17 | factor = 2 if bilinear else 1 18 | self.down4 = Down(512, 1024 // factor) 19 | self.up1 = Up(1024, 512 // factor, bilinear) 20 | self.up2 = Up(512, 256 // factor, bilinear) 21 | self.up3 = Up(256, 128 // factor, bilinear) 22 | self.up4 = Up(128, 64, bilinear) 23 | self.outc = OutConv(64, n_classes) 24 | 25 | def forward(self, x): 26 | x1 = self.inc(x) 27 | x2 = self.down1(x1) 28 | x3 = self.down2(x2) 29 | x4 = self.down3(x3) 30 | x5 = self.down4(x4) 31 | x = self.up1(x5, x4) 32 | x = self.up2(x, x3) 33 | x = self.up3(x, x2) 34 | x = self.up4(x, x1) 35 | logits = self.outc(x) 36 | return logits 37 | 38 | 39 | 40 | -------------------------------------------------------------------------------- /utils/unet/unet_parts.py: -------------------------------------------------------------------------------- 1 | """ Parts of the U-Net model """ 2 | 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | 7 | 8 | class DoubleConv(nn.Module): 9 | """(convolution => [BN] => ReLU) * 2""" 10 | 11 | def __init__(self, in_channels, out_channels, mid_channels=None): 12 | super().__init__() 13 | if not mid_channels: 14 | mid_channels = out_channels 15 | self.double_conv = nn.Sequential( 16 | nn.Conv2d(in_channels, mid_channels, kernel_size=3, padding=1, bias=False), 17 | nn.BatchNorm2d(mid_channels), 18 | nn.ReLU(inplace=True), 19 | nn.Conv2d(mid_channels, out_channels, kernel_size=3, padding=1, bias=False), 20 | nn.BatchNorm2d(out_channels), 21 | nn.ReLU(inplace=True) 22 | ) 23 | 24 | def forward(self, x): 25 | return self.double_conv(x) 26 | 27 | 28 | class Down(nn.Module): 29 | """Downscaling with maxpool then double conv""" 30 | 31 | def __init__(self, in_channels, out_channels): 32 | super().__init__() 33 | self.maxpool_conv = nn.Sequential( 34 | nn.MaxPool2d(2), 35 | DoubleConv(in_channels, out_channels) 36 | ) 37 | 38 | def forward(self, x): 39 | return self.maxpool_conv(x) 40 | 41 | 42 | class Up(nn.Module): 43 | """Upscaling then double conv""" 44 | 45 | def __init__(self, in_channels, out_channels, bilinear=True): 46 | super().__init__() 47 | 48 | # if bilinear, use the normal convolutions to reduce the number of channels 49 | if bilinear: 50 | self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True) 51 | self.conv = DoubleConv(in_channels, out_channels, in_channels // 2) 52 | else: 53 | self.up = nn.ConvTranspose2d(in_channels, in_channels // 2, kernel_size=2, stride=2) 54 | self.conv = DoubleConv(in_channels, out_channels) 55 | 56 | def forward(self, x1, x2): 57 | x1 = self.up(x1) 58 | # input is CHW 59 | diffY = x2.size()[2] - x1.size()[2] 60 | diffX = x2.size()[3] - x1.size()[3] 61 | 62 | x1 = F.pad(x1, [diffX // 2, diffX - diffX // 2, 63 | diffY // 2, diffY - diffY // 2]) 64 | # if you have padding issues, see 65 | # https://github.com/HaiyongJiang/U-Net-Pytorch-Unstructured-Buggy/commit/0e854509c2cea854e247a9c615f175f76fbb2e3a 66 | # https://github.com/xiaopeng-liao/Pytorch-UNet/commit/8ebac70e633bac59fc22bb5195e513d5832fb3bd 67 | x = torch.cat([x2, x1], dim=1) 68 | return self.conv(x) 69 | 70 | 71 | class OutConv(nn.Module): 72 | def __init__(self, in_channels, out_channels): 73 | super(OutConv, self).__init__() 74 | self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1) 75 | self.sigmoid = nn.Sigmoid() 76 | def forward(self, x): 77 | return self.sigmoid(self.conv(x)) 78 | --------------------------------------------------------------------------------