├── Deep_Fake_Demo.ipynb ├── LICENSE.md ├── README.md ├── animate.py ├── augmentation.py ├── config ├── bair-256.yaml ├── fashion-256.yaml ├── mgif-256.yaml ├── nemo-256.yaml ├── taichi-256.yaml ├── taichi-adv-256.yaml ├── vox-256.yaml └── vox-adv-256.yaml ├── crop-video.py ├── data ├── bair256.csv ├── taichi-loading │ ├── README.md │ ├── load_videos.py │ └── taichi-metadata.csv └── taichi256.csv ├── demo.py ├── frames_dataset.py ├── logger.py ├── modules ├── dense_motion.py ├── discriminator.py ├── generator.py ├── keypoint_detector.py ├── model.py └── util.py ├── reconstruction.py ├── requirements.txt ├── run.py ├── sup-mat ├── absolute-demo.gif ├── download.gif ├── fashion-teaser.gif ├── mgif-teaser.gif ├── relative-demo.gif └── vox-teaser.gif ├── sync_batchnorm ├── __init__.py ├── batchnorm.py ├── comm.py ├── replicate.py └── unittest.py └── train.py /LICENSE.md: -------------------------------------------------------------------------------- 1 | ## creative commons 2 | 3 | # Attribution-NonCommercial 4.0 International 4 | 5 | Creative Commons Corporation (“Creative Commons”) is not a law firm and does not provide legal services or legal advice. Distribution of Creative Commons public licenses does not create a lawyer-client or other relationship. Creative Commons makes its licenses and related information available on an “as-is” basis. Creative Commons gives no warranties regarding its licenses, any material licensed under their terms and conditions, or any related information. Creative Commons disclaims all liability for damages resulting from their use to the fullest extent possible. 6 | 7 | ### Using Creative Commons Public Licenses 8 | 9 | Creative Commons public licenses provide a standard set of terms and conditions that creators and other rights holders may use to share original works of authorship and other material subject to copyright and certain other rights specified in the public license below. The following considerations are for informational purposes only, are not exhaustive, and do not form part of our licenses. 10 | 11 | * __Considerations for licensors:__ Our public licenses are intended for use by those authorized to give the public permission to use material in ways otherwise restricted by copyright and certain other rights. Our licenses are irrevocable. Licensors should read and understand the terms and conditions of the license they choose before applying it. Licensors should also secure all rights necessary before applying our licenses so that the public can reuse the material as expected. Licensors should clearly mark any material not subject to the license. This includes other CC-licensed material, or material used under an exception or limitation to copyright. [More considerations for licensors](http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensors). 12 | 13 | * __Considerations for the public:__ By using one of our public licenses, a licensor grants the public permission to use the licensed material under specified terms and conditions. If the licensor’s permission is not necessary for any reason–for example, because of any applicable exception or limitation to copyright–then that use is not regulated by the license. Our licenses grant only permissions under copyright and certain other rights that a licensor has authority to grant. Use of the licensed material may still be restricted for other reasons, including because others have copyright or other rights in the material. A licensor may make special requests, such as asking that all changes be marked or described. Although not required by our licenses, you are encouraged to respect those requests where reasonable. [More considerations for the public](http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensees). 14 | 15 | ## Creative Commons Attribution-NonCommercial 4.0 International Public License 16 | 17 | By exercising the Licensed Rights (defined below), You accept and agree to be bound by the terms and conditions of this Creative Commons Attribution-NonCommercial 4.0 International Public License ("Public License"). To the extent this Public License may be interpreted as a contract, You are granted the Licensed Rights in consideration of Your acceptance of these terms and conditions, and the Licensor grants You such rights in consideration of benefits the Licensor receives from making the Licensed Material available under these terms and conditions. 18 | 19 | ### Section 1 – Definitions. 20 | 21 | a. __Adapted Material__ means material subject to Copyright and Similar Rights that is derived from or based upon the Licensed Material and in which the Licensed Material is translated, altered, arranged, transformed, or otherwise modified in a manner requiring permission under the Copyright and Similar Rights held by the Licensor. For purposes of this Public License, where the Licensed Material is a musical work, performance, or sound recording, Adapted Material is always produced where the Licensed Material is synched in timed relation with a moving image. 22 | 23 | b. __Adapter's License__ means the license You apply to Your Copyright and Similar Rights in Your contributions to Adapted Material in accordance with the terms and conditions of this Public License. 24 | 25 | c. __Copyright and Similar Rights__ means copyright and/or similar rights closely related to copyright including, without limitation, performance, broadcast, sound recording, and Sui Generis Database Rights, without regard to how the rights are labeled or categorized. For purposes of this Public License, the rights specified in Section 2(b)(1)-(2) are not Copyright and Similar Rights. 26 | 27 | d. __Effective Technological Measures__ means those measures that, in the absence of proper authority, may not be circumvented under laws fulfilling obligations under Article 11 of the WIPO Copyright Treaty adopted on December 20, 1996, and/or similar international agreements. 28 | 29 | e. __Exceptions and Limitations__ means fair use, fair dealing, and/or any other exception or limitation to Copyright and Similar Rights that applies to Your use of the Licensed Material. 30 | 31 | f. __Licensed Material__ means the artistic or literary work, database, or other material to which the Licensor applied this Public License. 32 | 33 | g. __Licensed Rights__ means the rights granted to You subject to the terms and conditions of this Public License, which are limited to all Copyright and Similar Rights that apply to Your use of the Licensed Material and that the Licensor has authority to license. 34 | 35 | h. __Licensor__ means the individual(s) or entity(ies) granting rights under this Public License. 36 | 37 | i. __NonCommercial__ means not primarily intended for or directed towards commercial advantage or monetary compensation. For purposes of this Public License, the exchange of the Licensed Material for other material subject to Copyright and Similar Rights by digital file-sharing or similar means is NonCommercial provided there is no payment of monetary compensation in connection with the exchange. 38 | 39 | j. __Share__ means to provide material to the public by any means or process that requires permission under the Licensed Rights, such as reproduction, public display, public performance, distribution, dissemination, communication, or importation, and to make material available to the public including in ways that members of the public may access the material from a place and at a time individually chosen by them. 40 | 41 | k. __Sui Generis Database Rights__ means rights other than copyright resulting from Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases, as amended and/or succeeded, as well as other essentially equivalent rights anywhere in the world. 42 | 43 | l. __You__ means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning. 44 | 45 | ### Section 2 – Scope. 46 | 47 | a. ___License grant.___ 48 | 49 | 1. Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty-free, non-sublicensable, non-exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to: 50 | 51 | A. reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only; and 52 | 53 | B. produce, reproduce, and Share Adapted Material for NonCommercial purposes only. 54 | 55 | 2. __Exceptions and Limitations.__ For the avoidance of doubt, where Exceptions and Limitations apply to Your use, this Public License does not apply, and You do not need to comply with its terms and conditions. 56 | 57 | 3. __Term.__ The term of this Public License is specified in Section 6(a). 58 | 59 | 4. __Media and formats; technical modifications allowed.__ The Licensor authorizes You to exercise the Licensed Rights in all media and formats whether now known or hereafter created, and to make technical modifications necessary to do so. The Licensor waives and/or agrees not to assert any right or authority to forbid You from making technical modifications necessary to exercise the Licensed Rights, including technical modifications necessary to circumvent Effective Technological Measures. For purposes of this Public License, simply making modifications authorized by this Section 2(a)(4) never produces Adapted Material. 60 | 61 | 5. __Downstream recipients.__ 62 | 63 | A. __Offer from the Licensor – Licensed Material.__ Every recipient of the Licensed Material automatically receives an offer from the Licensor to exercise the Licensed Rights under the terms and conditions of this Public License. 64 | 65 | B. __No downstream restrictions.__ You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, the Licensed Material if doing so restricts exercise of the Licensed Rights by any recipient of the Licensed Material. 66 | 67 | 6. __No endorsement.__ Nothing in this Public License constitutes or may be construed as permission to assert or imply that You are, or that Your use of the Licensed Material is, connected with, or sponsored, endorsed, or granted official status by, the Licensor or others designated to receive attribution as provided in Section 3(a)(1)(A)(i). 68 | 69 | b. ___Other rights.___ 70 | 71 | 1. Moral rights, such as the right of integrity, are not licensed under this Public License, nor are publicity, privacy, and/or other similar personality rights; however, to the extent possible, the Licensor waives and/or agrees not to assert any such rights held by the Licensor to the limited extent necessary to allow You to exercise the Licensed Rights, but not otherwise. 72 | 73 | 2. Patent and trademark rights are not licensed under this Public License. 74 | 75 | 3. To the extent possible, the Licensor waives any right to collect royalties from You for the exercise of the Licensed Rights, whether directly or through a collecting society under any voluntary or waivable statutory or compulsory licensing scheme. In all other cases the Licensor expressly reserves any right to collect such royalties, including when the Licensed Material is used other than for NonCommercial purposes. 76 | 77 | ### Section 3 – License Conditions. 78 | 79 | Your exercise of the Licensed Rights is expressly made subject to the following conditions. 80 | 81 | a. ___Attribution.___ 82 | 83 | 1. If You Share the Licensed Material (including in modified form), You must: 84 | 85 | A. retain the following if it is supplied by the Licensor with the Licensed Material: 86 | 87 | i. identification of the creator(s) of the Licensed Material and any others designated to receive attribution, in any reasonable manner requested by the Licensor (including by pseudonym if designated); 88 | 89 | ii. a copyright notice; 90 | 91 | iii. a notice that refers to this Public License; 92 | 93 | iv. a notice that refers to the disclaimer of warranties; 94 | 95 | v. a URI or hyperlink to the Licensed Material to the extent reasonably practicable; 96 | 97 | B. indicate if You modified the Licensed Material and retain an indication of any previous modifications; and 98 | 99 | C. indicate the Licensed Material is licensed under this Public License, and include the text of, or the URI or hyperlink to, this Public License. 100 | 101 | 2. You may satisfy the conditions in Section 3(a)(1) in any reasonable manner based on the medium, means, and context in which You Share the Licensed Material. For example, it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information. 102 | 103 | 3. If requested by the Licensor, You must remove any of the information required by Section 3(a)(1)(A) to the extent reasonably practicable. 104 | 105 | 4. If You Share Adapted Material You produce, the Adapter's License You apply must not prevent recipients of the Adapted Material from complying with this Public License. 106 | 107 | ### Section 4 – Sui Generis Database Rights. 108 | 109 | Where the Licensed Rights include Sui Generis Database Rights that apply to Your use of the Licensed Material: 110 | 111 | a. for the avoidance of doubt, Section 2(a)(1) grants You the right to extract, reuse, reproduce, and Share all or a substantial portion of the contents of the database for NonCommercial purposes only; 112 | 113 | b. if You include all or a substantial portion of the database contents in a database in which You have Sui Generis Database Rights, then the database in which You have Sui Generis Database Rights (but not its individual contents) is Adapted Material; and 114 | 115 | c. You must comply with the conditions in Section 3(a) if You Share all or a substantial portion of the contents of the database. 116 | 117 | For the avoidance of doubt, this Section 4 supplements and does not replace Your obligations under this Public License where the Licensed Rights include other Copyright and Similar Rights. 118 | 119 | ### Section 5 – Disclaimer of Warranties and Limitation of Liability. 120 | 121 | a. __Unless otherwise separately undertaken by the Licensor, to the extent possible, the Licensor offers the Licensed Material as-is and as-available, and makes no representations or warranties of any kind concerning the Licensed Material, whether express, implied, statutory, or other. This includes, without limitation, warranties of title, merchantability, fitness for a particular purpose, non-infringement, absence of latent or other defects, accuracy, or the presence or absence of errors, whether or not known or discoverable. Where disclaimers of warranties are not allowed in full or in part, this disclaimer may not apply to You.__ 122 | 123 | b. __To the extent possible, in no event will the Licensor be liable to You on any legal theory (including, without limitation, negligence) or otherwise for any direct, special, indirect, incidental, consequential, punitive, exemplary, or other losses, costs, expenses, or damages arising out of this Public License or use of the Licensed Material, even if the Licensor has been advised of the possibility of such losses, costs, expenses, or damages. Where a limitation of liability is not allowed in full or in part, this limitation may not apply to You.__ 124 | 125 | c. The disclaimer of warranties and limitation of liability provided above shall be interpreted in a manner that, to the extent possible, most closely approximates an absolute disclaimer and waiver of all liability. 126 | 127 | ### Section 6 – Term and Termination. 128 | 129 | a. This Public License applies for the term of the Copyright and Similar Rights licensed here. However, if You fail to comply with this Public License, then Your rights under this Public License terminate automatically. 130 | 131 | b. Where Your right to use the Licensed Material has terminated under Section 6(a), it reinstates: 132 | 133 | 1. automatically as of the date the violation is cured, provided it is cured within 30 days of Your discovery of the violation; or 134 | 135 | 2. upon express reinstatement by the Licensor. 136 | 137 | For the avoidance of doubt, this Section 6(b) does not affect any right the Licensor may have to seek remedies for Your violations of this Public License. 138 | 139 | c. For the avoidance of doubt, the Licensor may also offer the Licensed Material under separate terms or conditions or stop distributing the Licensed Material at any time; however, doing so will not terminate this Public License. 140 | 141 | d. Sections 1, 5, 6, 7, and 8 survive termination of this Public License. 142 | 143 | ### Section 7 – Other Terms and Conditions. 144 | 145 | a. The Licensor shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed. 146 | 147 | b. Any arrangements, understandings, or agreements regarding the Licensed Material not stated herein are separate from and independent of the terms and conditions of this Public License. 148 | 149 | ### Section 8 – Interpretation. 150 | 151 | a. For the avoidance of doubt, this Public License does not, and shall not be interpreted to, reduce, limit, restrict, or impose conditions on any use of the Licensed Material that could lawfully be made without permission under this Public License. 152 | 153 | b. To the extent possible, if any provision of this Public License is deemed unenforceable, it shall be automatically reformed to the minimum extent necessary to make it enforceable. If the provision cannot be reformed, it shall be severed from this Public License without affecting the enforceability of the remaining terms and conditions. 154 | 155 | c. No term or condition of this Public License will be waived and no failure to comply consented to unless expressly agreed to by the Licensor. 156 | 157 | d. Nothing in this Public License constitutes or may be interpreted as a limitation upon, or waiver of, any privileges and immunities that apply to the Licensor or You, including from the legal processes of any jurisdiction or authority. 158 | 159 | > Creative Commons is not a party to its public licenses. Notwithstanding, Creative Commons may elect to apply one of its public licenses to material it publishes and in those instances will be considered the “Licensor.” Except for the limited purpose of indicating that material is shared under a Creative Commons public license or as otherwise permitted by the Creative Commons policies published at [creativecommons.org/policies](http://creativecommons.org/policies), Creative Commons does not authorize the use of the trademark “Creative Commons” or any other trademark or logo of Creative Commons without its prior written consent including, without limitation, in connection with any unauthorized modifications to any of its public licenses or any other arrangements, understandings, or agreements concerning use of licensed material. For the avoidance of doubt, this paragraph does not form part of the public licenses. 160 | > 161 | > Creative Commons may be contacted at creativecommons.org 162 | 163 | --------------------------- LICENSE FOR Synchronized-BatchNorm-PyTorch -------------------------------- 164 | 165 | MIT License 166 | 167 | Copyright (c) 2018 Jiayuan MAO 168 | 169 | Permission is hereby granted, free of charge, to any person obtaining a copy 170 | of this software and associated documentation files (the "Software"), to deal 171 | in the Software without restriction, including without limitation the rights 172 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 173 | copies of the Software, and to permit persons to whom the Software is 174 | furnished to do so, subject to the following conditions: 175 | 176 | The above copyright notice and this permission notice shall be included in all 177 | copies or substantial portions of the Software. 178 | 179 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 180 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 181 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 182 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 183 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 184 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 185 | SOFTWARE. 186 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ## ⚡ First Order Motion Model for Image Animation ⚡ 2 | 3 | This repository contains source code for the paper [First Order Motion Model for Image Animation](https://papers.nips.cc/paper/8935-first-order-motion-model-for-image-animation) by Aliaksandr Siarohin, [Stéphane Lathuilière](http://stelat.eu), [Sergey Tulyakov](http://stulyakov.com), [Elisa Ricci](http://elisaricci.eu/) and [Nicu Sebe](http://disi.unitn.it/~sebe/). 4 | This repository is taken from a Youtube video by [Two Minute Papers](https://www.youtube.com/watch?v=mUfJOQKdtAk&t=17s) 5 | 6 | Original Aliaksandr Siarohin Repo: [Github](https://github.com/AliaksandrSiarohin/first-order-model)
7 | Two Minute Papers Youtube Video Link : [Youtube-Video](https://www.youtube.com/watch?v=mUfJOQKdtAk&t=17s) 8 | 9 | ### 📝 Model Output 10 | ![Screenshot](https://github.com/snehitvaddi/Deep-Fake_First_Order_Model/blob/master/sup-mat/vox-teaser.gif) 11 | ### 🖼 Example on Custom Data 12 | 13 | ### 🔬 COLAB DEMO 14 | You can run this code from [GOOGLE COLAB](https://colab.research.google.com/drive/11YHTBYpBDoG28RwmVKj2VYt2jBblMh1M?usp=sharing) 15 | ### 📌 Installation 16 | This code supports ```python3```. To install the dependencies run: 17 | ``` 18 | pip install -r requirements.txt 19 | ``` 20 | ### 🕶 Pre-trained checkpoint 21 | Checkpoints can be found under following link: [google-drive](https://drive.google.com/open?id=1PyQJmkdCsAkOYwUyaj_l-l0as-iLDgeH) or [yandex-disk](https://yadi.sk/d/lEw8uRm140L_eQ). 22 | 23 | ### ⚡ Animation Demo 24 | To run a demo, download checkpoint and run the following command: 25 | ``` 26 | python demo.py --config config/dataset_name.yaml --driving_video path/to/driving --source_image path/to/source --checkpoint path/to/checkpoint --relative --adapt_scale 27 | ``` 28 | The result will be stored in ```result.mp4```. 29 | 30 | The driving videos and source images should be cropped before it can be used in our method. To obtain some semi-automatic crop suggestions you can use ```python crop-video.py --inp some_youtube_video.mp4```. It will generate commands for crops using ffmpeg. In order to use the script, face-alligment library is needed: 31 | ``` 32 | git clone https://github.com/1adrianb/face-alignment 33 | cd face-alignment 34 | pip install -r requirements.txt 35 | python setup.py install 36 | 37 | ### ⚖ Training on your own dataset 38 | 1) Resize all the videos to the same size e.g 256x256, the videos can be in '.gif', '.mp4' or folder with images. 39 | We recommend the later, for each video make a separate folder with all the frames in '.png' format. This format is loss-less, and it has better i/o performance. 40 | 41 | 2) Create a folder ```data/dataset_name``` with 2 subfolders ```train``` and ```test```, put training videos in the ```train``` and testing in the ```test```. 42 | 43 | 3) Create a config ```config/dataset_name.yaml```, in dataset_params specify the root dir the ```root_dir: data/dataset_name```. Also adjust the number of epoch in train_params. 44 | 45 | -------------------------------------------------------------------------------- /animate.py: -------------------------------------------------------------------------------- 1 | import os 2 | from tqdm import tqdm 3 | 4 | import torch 5 | from torch.utils.data import DataLoader 6 | 7 | from frames_dataset import PairedDataset 8 | from logger import Logger, Visualizer 9 | import imageio 10 | from scipy.spatial import ConvexHull 11 | import numpy as np 12 | 13 | from sync_batchnorm import DataParallelWithCallback 14 | 15 | 16 | def normalize_kp(kp_source, kp_driving, kp_driving_initial, adapt_movement_scale=False, 17 | use_relative_movement=False, use_relative_jacobian=False): 18 | if adapt_movement_scale: 19 | source_area = ConvexHull(kp_source['value'][0].data.cpu().numpy()).volume 20 | driving_area = ConvexHull(kp_driving_initial['value'][0].data.cpu().numpy()).volume 21 | adapt_movement_scale = np.sqrt(source_area) / np.sqrt(driving_area) 22 | else: 23 | adapt_movement_scale = 1 24 | 25 | kp_new = {k: v for k, v in kp_driving.items()} 26 | 27 | if use_relative_movement: 28 | kp_value_diff = (kp_driving['value'] - kp_driving_initial['value']) 29 | kp_value_diff *= adapt_movement_scale 30 | kp_new['value'] = kp_value_diff + kp_source['value'] 31 | 32 | if use_relative_jacobian: 33 | jacobian_diff = torch.matmul(kp_driving['jacobian'], torch.inverse(kp_driving_initial['jacobian'])) 34 | kp_new['jacobian'] = torch.matmul(jacobian_diff, kp_source['jacobian']) 35 | 36 | return kp_new 37 | 38 | 39 | def animate(config, generator, kp_detector, checkpoint, log_dir, dataset): 40 | log_dir = os.path.join(log_dir, 'animation') 41 | png_dir = os.path.join(log_dir, 'png') 42 | animate_params = config['animate_params'] 43 | 44 | dataset = PairedDataset(initial_dataset=dataset, number_of_pairs=animate_params['num_pairs']) 45 | dataloader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=1) 46 | 47 | if checkpoint is not None: 48 | Logger.load_cpk(checkpoint, generator=generator, kp_detector=kp_detector) 49 | else: 50 | raise AttributeError("Checkpoint should be specified for mode='animate'.") 51 | 52 | if not os.path.exists(log_dir): 53 | os.makedirs(log_dir) 54 | 55 | if not os.path.exists(png_dir): 56 | os.makedirs(png_dir) 57 | 58 | if torch.cuda.is_available(): 59 | generator = DataParallelWithCallback(generator) 60 | kp_detector = DataParallelWithCallback(kp_detector) 61 | 62 | generator.eval() 63 | kp_detector.eval() 64 | 65 | for it, x in tqdm(enumerate(dataloader)): 66 | with torch.no_grad(): 67 | predictions = [] 68 | visualizations = [] 69 | 70 | driving_video = x['driving_video'] 71 | source_frame = x['source_video'][:, :, 0, :, :] 72 | 73 | kp_source = kp_detector(source_frame) 74 | kp_driving_initial = kp_detector(driving_video[:, :, 0]) 75 | 76 | for frame_idx in range(driving_video.shape[2]): 77 | driving_frame = driving_video[:, :, frame_idx] 78 | kp_driving = kp_detector(driving_frame) 79 | kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving, 80 | kp_driving_initial=kp_driving_initial, **animate_params['normalization_params']) 81 | out = generator(source_frame, kp_source=kp_source, kp_driving=kp_norm) 82 | 83 | out['kp_driving'] = kp_driving 84 | out['kp_source'] = kp_source 85 | out['kp_norm'] = kp_norm 86 | 87 | del out['sparse_deformed'] 88 | 89 | predictions.append(np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0]) 90 | 91 | visualization = Visualizer(**config['visualizer_params']).visualize(source=source_frame, 92 | driving=driving_frame, out=out) 93 | visualization = visualization 94 | visualizations.append(visualization) 95 | 96 | predictions = np.concatenate(predictions, axis=1) 97 | result_name = "-".join([x['driving_name'][0], x['source_name'][0]]) 98 | imageio.imsave(os.path.join(png_dir, result_name + '.png'), (255 * predictions).astype(np.uint8)) 99 | 100 | image_name = result_name + animate_params['format'] 101 | imageio.mimsave(os.path.join(log_dir, image_name), visualizations) 102 | -------------------------------------------------------------------------------- /augmentation.py: -------------------------------------------------------------------------------- 1 | """ 2 | Code from https://github.com/hassony2/torch_videovision 3 | """ 4 | 5 | import numbers 6 | 7 | import random 8 | import numpy as np 9 | import PIL 10 | 11 | from skimage.transform import resize, rotate 12 | from skimage.util import pad 13 | import torchvision 14 | 15 | import warnings 16 | 17 | from skimage import img_as_ubyte, img_as_float 18 | 19 | 20 | def crop_clip(clip, min_h, min_w, h, w): 21 | if isinstance(clip[0], np.ndarray): 22 | cropped = [img[min_h:min_h + h, min_w:min_w + w, :] for img in clip] 23 | 24 | elif isinstance(clip[0], PIL.Image.Image): 25 | cropped = [ 26 | img.crop((min_w, min_h, min_w + w, min_h + h)) for img in clip 27 | ] 28 | else: 29 | raise TypeError('Expected numpy.ndarray or PIL.Image' + 30 | 'but got list of {0}'.format(type(clip[0]))) 31 | return cropped 32 | 33 | 34 | def pad_clip(clip, h, w): 35 | im_h, im_w = clip[0].shape[:2] 36 | pad_h = (0, 0) if h < im_h else ((h - im_h) // 2, (h - im_h + 1) // 2) 37 | pad_w = (0, 0) if w < im_w else ((w - im_w) // 2, (w - im_w + 1) // 2) 38 | 39 | return pad(clip, ((0, 0), pad_h, pad_w, (0, 0)), mode='edge') 40 | 41 | 42 | def resize_clip(clip, size, interpolation='bilinear'): 43 | if isinstance(clip[0], np.ndarray): 44 | if isinstance(size, numbers.Number): 45 | im_h, im_w, im_c = clip[0].shape 46 | # Min spatial dim already matches minimal size 47 | if (im_w <= im_h and im_w == size) or (im_h <= im_w 48 | and im_h == size): 49 | return clip 50 | new_h, new_w = get_resize_sizes(im_h, im_w, size) 51 | size = (new_w, new_h) 52 | else: 53 | size = size[1], size[0] 54 | 55 | scaled = [ 56 | resize(img, size, order=1 if interpolation == 'bilinear' else 0, preserve_range=True, 57 | mode='constant', anti_aliasing=True) for img in clip 58 | ] 59 | elif isinstance(clip[0], PIL.Image.Image): 60 | if isinstance(size, numbers.Number): 61 | im_w, im_h = clip[0].size 62 | # Min spatial dim already matches minimal size 63 | if (im_w <= im_h and im_w == size) or (im_h <= im_w 64 | and im_h == size): 65 | return clip 66 | new_h, new_w = get_resize_sizes(im_h, im_w, size) 67 | size = (new_w, new_h) 68 | else: 69 | size = size[1], size[0] 70 | if interpolation == 'bilinear': 71 | pil_inter = PIL.Image.NEAREST 72 | else: 73 | pil_inter = PIL.Image.BILINEAR 74 | scaled = [img.resize(size, pil_inter) for img in clip] 75 | else: 76 | raise TypeError('Expected numpy.ndarray or PIL.Image' + 77 | 'but got list of {0}'.format(type(clip[0]))) 78 | return scaled 79 | 80 | 81 | def get_resize_sizes(im_h, im_w, size): 82 | if im_w < im_h: 83 | ow = size 84 | oh = int(size * im_h / im_w) 85 | else: 86 | oh = size 87 | ow = int(size * im_w / im_h) 88 | return oh, ow 89 | 90 | 91 | class RandomFlip(object): 92 | def __init__(self, time_flip=False, horizontal_flip=False): 93 | self.time_flip = time_flip 94 | self.horizontal_flip = horizontal_flip 95 | 96 | def __call__(self, clip): 97 | if random.random() < 0.5 and self.time_flip: 98 | return clip[::-1] 99 | if random.random() < 0.5 and self.horizontal_flip: 100 | return [np.fliplr(img) for img in clip] 101 | 102 | return clip 103 | 104 | 105 | class RandomResize(object): 106 | """Resizes a list of (H x W x C) numpy.ndarray to the final size 107 | The larger the original image is, the more times it takes to 108 | interpolate 109 | Args: 110 | interpolation (str): Can be one of 'nearest', 'bilinear' 111 | defaults to nearest 112 | size (tuple): (widht, height) 113 | """ 114 | 115 | def __init__(self, ratio=(3. / 4., 4. / 3.), interpolation='nearest'): 116 | self.ratio = ratio 117 | self.interpolation = interpolation 118 | 119 | def __call__(self, clip): 120 | scaling_factor = random.uniform(self.ratio[0], self.ratio[1]) 121 | 122 | if isinstance(clip[0], np.ndarray): 123 | im_h, im_w, im_c = clip[0].shape 124 | elif isinstance(clip[0], PIL.Image.Image): 125 | im_w, im_h = clip[0].size 126 | 127 | new_w = int(im_w * scaling_factor) 128 | new_h = int(im_h * scaling_factor) 129 | new_size = (new_w, new_h) 130 | resized = resize_clip( 131 | clip, new_size, interpolation=self.interpolation) 132 | 133 | return resized 134 | 135 | 136 | class RandomCrop(object): 137 | """Extract random crop at the same location for a list of videos 138 | Args: 139 | size (sequence or int): Desired output size for the 140 | crop in format (h, w) 141 | """ 142 | 143 | def __init__(self, size): 144 | if isinstance(size, numbers.Number): 145 | size = (size, size) 146 | 147 | self.size = size 148 | 149 | def __call__(self, clip): 150 | """ 151 | Args: 152 | img (PIL.Image or numpy.ndarray): List of videos to be cropped 153 | in format (h, w, c) in numpy.ndarray 154 | Returns: 155 | PIL.Image or numpy.ndarray: Cropped list of videos 156 | """ 157 | h, w = self.size 158 | if isinstance(clip[0], np.ndarray): 159 | im_h, im_w, im_c = clip[0].shape 160 | elif isinstance(clip[0], PIL.Image.Image): 161 | im_w, im_h = clip[0].size 162 | else: 163 | raise TypeError('Expected numpy.ndarray or PIL.Image' + 164 | 'but got list of {0}'.format(type(clip[0]))) 165 | 166 | clip = pad_clip(clip, h, w) 167 | im_h, im_w = clip.shape[1:3] 168 | x1 = 0 if h == im_h else random.randint(0, im_w - w) 169 | y1 = 0 if w == im_w else random.randint(0, im_h - h) 170 | cropped = crop_clip(clip, y1, x1, h, w) 171 | 172 | return cropped 173 | 174 | 175 | class RandomRotation(object): 176 | """Rotate entire clip randomly by a random angle within 177 | given bounds 178 | Args: 179 | degrees (sequence or int): Range of degrees to select from 180 | If degrees is a number instead of sequence like (min, max), 181 | the range of degrees, will be (-degrees, +degrees). 182 | """ 183 | 184 | def __init__(self, degrees): 185 | if isinstance(degrees, numbers.Number): 186 | if degrees < 0: 187 | raise ValueError('If degrees is a single number,' 188 | 'must be positive') 189 | degrees = (-degrees, degrees) 190 | else: 191 | if len(degrees) != 2: 192 | raise ValueError('If degrees is a sequence,' 193 | 'it must be of len 2.') 194 | 195 | self.degrees = degrees 196 | 197 | def __call__(self, clip): 198 | """ 199 | Args: 200 | img (PIL.Image or numpy.ndarray): List of videos to be cropped 201 | in format (h, w, c) in numpy.ndarray 202 | Returns: 203 | PIL.Image or numpy.ndarray: Cropped list of videos 204 | """ 205 | angle = random.uniform(self.degrees[0], self.degrees[1]) 206 | if isinstance(clip[0], np.ndarray): 207 | rotated = [rotate(image=img, angle=angle, preserve_range=True) for img in clip] 208 | elif isinstance(clip[0], PIL.Image.Image): 209 | rotated = [img.rotate(angle) for img in clip] 210 | else: 211 | raise TypeError('Expected numpy.ndarray or PIL.Image' + 212 | 'but got list of {0}'.format(type(clip[0]))) 213 | 214 | return rotated 215 | 216 | 217 | class ColorJitter(object): 218 | """Randomly change the brightness, contrast and saturation and hue of the clip 219 | Args: 220 | brightness (float): How much to jitter brightness. brightness_factor 221 | is chosen uniformly from [max(0, 1 - brightness), 1 + brightness]. 222 | contrast (float): How much to jitter contrast. contrast_factor 223 | is chosen uniformly from [max(0, 1 - contrast), 1 + contrast]. 224 | saturation (float): How much to jitter saturation. saturation_factor 225 | is chosen uniformly from [max(0, 1 - saturation), 1 + saturation]. 226 | hue(float): How much to jitter hue. hue_factor is chosen uniformly from 227 | [-hue, hue]. Should be >=0 and <= 0.5. 228 | """ 229 | 230 | def __init__(self, brightness=0, contrast=0, saturation=0, hue=0): 231 | self.brightness = brightness 232 | self.contrast = contrast 233 | self.saturation = saturation 234 | self.hue = hue 235 | 236 | def get_params(self, brightness, contrast, saturation, hue): 237 | if brightness > 0: 238 | brightness_factor = random.uniform( 239 | max(0, 1 - brightness), 1 + brightness) 240 | else: 241 | brightness_factor = None 242 | 243 | if contrast > 0: 244 | contrast_factor = random.uniform( 245 | max(0, 1 - contrast), 1 + contrast) 246 | else: 247 | contrast_factor = None 248 | 249 | if saturation > 0: 250 | saturation_factor = random.uniform( 251 | max(0, 1 - saturation), 1 + saturation) 252 | else: 253 | saturation_factor = None 254 | 255 | if hue > 0: 256 | hue_factor = random.uniform(-hue, hue) 257 | else: 258 | hue_factor = None 259 | return brightness_factor, contrast_factor, saturation_factor, hue_factor 260 | 261 | def __call__(self, clip): 262 | """ 263 | Args: 264 | clip (list): list of PIL.Image 265 | Returns: 266 | list PIL.Image : list of transformed PIL.Image 267 | """ 268 | if isinstance(clip[0], np.ndarray): 269 | brightness, contrast, saturation, hue = self.get_params( 270 | self.brightness, self.contrast, self.saturation, self.hue) 271 | 272 | # Create img transform function sequence 273 | img_transforms = [] 274 | if brightness is not None: 275 | img_transforms.append(lambda img: torchvision.transforms.functional.adjust_brightness(img, brightness)) 276 | if saturation is not None: 277 | img_transforms.append(lambda img: torchvision.transforms.functional.adjust_saturation(img, saturation)) 278 | if hue is not None: 279 | img_transforms.append(lambda img: torchvision.transforms.functional.adjust_hue(img, hue)) 280 | if contrast is not None: 281 | img_transforms.append(lambda img: torchvision.transforms.functional.adjust_contrast(img, contrast)) 282 | random.shuffle(img_transforms) 283 | img_transforms = [img_as_ubyte, torchvision.transforms.ToPILImage()] + img_transforms + [np.array, 284 | img_as_float] 285 | 286 | with warnings.catch_warnings(): 287 | warnings.simplefilter("ignore") 288 | jittered_clip = [] 289 | for img in clip: 290 | jittered_img = img 291 | for func in img_transforms: 292 | jittered_img = func(jittered_img) 293 | jittered_clip.append(jittered_img.astype('float32')) 294 | elif isinstance(clip[0], PIL.Image.Image): 295 | brightness, contrast, saturation, hue = self.get_params( 296 | self.brightness, self.contrast, self.saturation, self.hue) 297 | 298 | # Create img transform function sequence 299 | img_transforms = [] 300 | if brightness is not None: 301 | img_transforms.append(lambda img: torchvision.transforms.functional.adjust_brightness(img, brightness)) 302 | if saturation is not None: 303 | img_transforms.append(lambda img: torchvision.transforms.functional.adjust_saturation(img, saturation)) 304 | if hue is not None: 305 | img_transforms.append(lambda img: torchvision.transforms.functional.adjust_hue(img, hue)) 306 | if contrast is not None: 307 | img_transforms.append(lambda img: torchvision.transforms.functional.adjust_contrast(img, contrast)) 308 | random.shuffle(img_transforms) 309 | 310 | # Apply to all videos 311 | jittered_clip = [] 312 | for img in clip: 313 | for func in img_transforms: 314 | jittered_img = func(img) 315 | jittered_clip.append(jittered_img) 316 | 317 | else: 318 | raise TypeError('Expected numpy.ndarray or PIL.Image' + 319 | 'but got list of {0}'.format(type(clip[0]))) 320 | return jittered_clip 321 | 322 | 323 | class AllAugmentationTransform: 324 | def __init__(self, resize_param=None, rotation_param=None, flip_param=None, crop_param=None, jitter_param=None): 325 | self.transforms = [] 326 | 327 | if flip_param is not None: 328 | self.transforms.append(RandomFlip(**flip_param)) 329 | 330 | if rotation_param is not None: 331 | self.transforms.append(RandomRotation(**rotation_param)) 332 | 333 | if resize_param is not None: 334 | self.transforms.append(RandomResize(**resize_param)) 335 | 336 | if crop_param is not None: 337 | self.transforms.append(RandomCrop(**crop_param)) 338 | 339 | if jitter_param is not None: 340 | self.transforms.append(ColorJitter(**jitter_param)) 341 | 342 | def __call__(self, clip): 343 | for t in self.transforms: 344 | clip = t(clip) 345 | return clip 346 | -------------------------------------------------------------------------------- /config/bair-256.yaml: -------------------------------------------------------------------------------- 1 | dataset_params: 2 | root_dir: data/bair 3 | frame_shape: [256, 256, 3] 4 | id_sampling: False 5 | augmentation_params: 6 | flip_param: 7 | horizontal_flip: True 8 | time_flip: True 9 | jitter_param: 10 | brightness: 0.1 11 | contrast: 0.1 12 | saturation: 0.1 13 | hue: 0.1 14 | 15 | 16 | model_params: 17 | common_params: 18 | num_kp: 10 19 | num_channels: 3 20 | estimate_jacobian: True 21 | kp_detector_params: 22 | temperature: 0.1 23 | block_expansion: 32 24 | max_features: 1024 25 | scale_factor: 0.25 26 | num_blocks: 5 27 | generator_params: 28 | block_expansion: 64 29 | max_features: 512 30 | num_down_blocks: 2 31 | num_bottleneck_blocks: 6 32 | estimate_occlusion_map: True 33 | dense_motion_params: 34 | block_expansion: 64 35 | max_features: 1024 36 | num_blocks: 5 37 | scale_factor: 0.25 38 | discriminator_params: 39 | scales: [1] 40 | block_expansion: 32 41 | max_features: 512 42 | num_blocks: 4 43 | sn: True 44 | 45 | train_params: 46 | num_epochs: 20 47 | num_repeats: 1 48 | epoch_milestones: [12, 18] 49 | lr_generator: 2.0e-4 50 | lr_discriminator: 2.0e-4 51 | lr_kp_detector: 2.0e-4 52 | batch_size: 36 53 | scales: [1, 0.5, 0.25, 0.125] 54 | checkpoint_freq: 10 55 | transform_params: 56 | sigma_affine: 0.05 57 | sigma_tps: 0.005 58 | points_tps: 5 59 | loss_weights: 60 | generator_gan: 1 61 | discriminator_gan: 1 62 | feature_matching: [10, 10, 10, 10] 63 | perceptual: [10, 10, 10, 10, 10] 64 | equivariance_value: 10 65 | equivariance_jacobian: 10 66 | 67 | reconstruction_params: 68 | num_videos: 1000 69 | format: '.mp4' 70 | 71 | animate_params: 72 | num_pairs: 50 73 | format: '.mp4' 74 | normalization_params: 75 | adapt_movement_scale: False 76 | use_relative_movement: True 77 | use_relative_jacobian: True 78 | 79 | visualizer_params: 80 | kp_size: 5 81 | draw_border: True 82 | colormap: 'gist_rainbow' 83 | -------------------------------------------------------------------------------- /config/fashion-256.yaml: -------------------------------------------------------------------------------- 1 | dataset_params: 2 | root_dir: data/fashion-png 3 | frame_shape: [256, 256, 3] 4 | id_sampling: False 5 | augmentation_params: 6 | flip_param: 7 | horizontal_flip: True 8 | time_flip: True 9 | jitter_param: 10 | hue: 0.1 11 | 12 | model_params: 13 | common_params: 14 | num_kp: 10 15 | num_channels: 3 16 | estimate_jacobian: True 17 | kp_detector_params: 18 | temperature: 0.1 19 | block_expansion: 32 20 | max_features: 1024 21 | scale_factor: 0.25 22 | num_blocks: 5 23 | generator_params: 24 | block_expansion: 64 25 | max_features: 512 26 | num_down_blocks: 2 27 | num_bottleneck_blocks: 6 28 | estimate_occlusion_map: True 29 | dense_motion_params: 30 | block_expansion: 64 31 | max_features: 1024 32 | num_blocks: 5 33 | scale_factor: 0.25 34 | discriminator_params: 35 | scales: [1] 36 | block_expansion: 32 37 | max_features: 512 38 | num_blocks: 4 39 | 40 | train_params: 41 | num_epochs: 100 42 | num_repeats: 50 43 | epoch_milestones: [60, 90] 44 | lr_generator: 2.0e-4 45 | lr_discriminator: 2.0e-4 46 | lr_kp_detector: 2.0e-4 47 | batch_size: 27 48 | scales: [1, 0.5, 0.25, 0.125] 49 | checkpoint_freq: 50 50 | transform_params: 51 | sigma_affine: 0.05 52 | sigma_tps: 0.005 53 | points_tps: 5 54 | loss_weights: 55 | generator_gan: 1 56 | discriminator_gan: 1 57 | feature_matching: [10, 10, 10, 10] 58 | perceptual: [10, 10, 10, 10, 10] 59 | equivariance_value: 10 60 | equivariance_jacobian: 10 61 | 62 | reconstruction_params: 63 | num_videos: 1000 64 | format: '.mp4' 65 | 66 | animate_params: 67 | num_pairs: 50 68 | format: '.mp4' 69 | normalization_params: 70 | adapt_movement_scale: False 71 | use_relative_movement: True 72 | use_relative_jacobian: True 73 | 74 | visualizer_params: 75 | kp_size: 5 76 | draw_border: True 77 | colormap: 'gist_rainbow' 78 | -------------------------------------------------------------------------------- /config/mgif-256.yaml: -------------------------------------------------------------------------------- 1 | dataset_params: 2 | root_dir: data/moving-gif 3 | frame_shape: [256, 256, 3] 4 | id_sampling: False 5 | augmentation_params: 6 | flip_param: 7 | horizontal_flip: True 8 | time_flip: True 9 | crop_param: 10 | size: [256, 256] 11 | resize_param: 12 | ratio: [0.9, 1.1] 13 | jitter_param: 14 | hue: 0.5 15 | 16 | model_params: 17 | common_params: 18 | num_kp: 10 19 | num_channels: 3 20 | estimate_jacobian: True 21 | kp_detector_params: 22 | temperature: 0.1 23 | block_expansion: 32 24 | max_features: 1024 25 | scale_factor: 0.25 26 | num_blocks: 5 27 | single_jacobian_map: True 28 | generator_params: 29 | block_expansion: 64 30 | max_features: 512 31 | num_down_blocks: 2 32 | num_bottleneck_blocks: 6 33 | estimate_occlusion_map: True 34 | dense_motion_params: 35 | block_expansion: 64 36 | max_features: 1024 37 | num_blocks: 5 38 | scale_factor: 0.25 39 | discriminator_params: 40 | scales: [1] 41 | block_expansion: 32 42 | max_features: 512 43 | num_blocks: 4 44 | sn: True 45 | 46 | train_params: 47 | num_epochs: 100 48 | num_repeats: 25 49 | epoch_milestones: [60, 90] 50 | lr_generator: 2.0e-4 51 | lr_discriminator: 2.0e-4 52 | lr_kp_detector: 2.0e-4 53 | 54 | batch_size: 36 55 | scales: [1, 0.5, 0.25, 0.125] 56 | checkpoint_freq: 100 57 | transform_params: 58 | sigma_affine: 0.05 59 | sigma_tps: 0.005 60 | points_tps: 5 61 | loss_weights: 62 | generator_gan: 1 63 | discriminator_gan: 1 64 | feature_matching: [10, 10, 10, 10] 65 | perceptual: [10, 10, 10, 10, 10] 66 | equivariance_value: 10 67 | equivariance_jacobian: 10 68 | 69 | reconstruction_params: 70 | num_videos: 1000 71 | format: '.mp4' 72 | 73 | animate_params: 74 | num_pairs: 50 75 | format: '.mp4' 76 | normalization_params: 77 | adapt_movement_scale: False 78 | use_relative_movement: True 79 | use_relative_jacobian: True 80 | 81 | visualizer_params: 82 | kp_size: 5 83 | draw_border: True 84 | colormap: 'gist_rainbow' 85 | -------------------------------------------------------------------------------- /config/nemo-256.yaml: -------------------------------------------------------------------------------- 1 | dataset_params: 2 | root_dir: data/nemo-png 3 | frame_shape: [256, 256, 3] 4 | id_sampling: False 5 | augmentation_params: 6 | flip_param: 7 | horizontal_flip: True 8 | time_flip: True 9 | 10 | model_params: 11 | common_params: 12 | num_kp: 10 13 | num_channels: 3 14 | estimate_jacobian: True 15 | kp_detector_params: 16 | temperature: 0.1 17 | block_expansion: 32 18 | max_features: 1024 19 | scale_factor: 0.25 20 | num_blocks: 5 21 | generator_params: 22 | block_expansion: 64 23 | max_features: 512 24 | num_down_blocks: 2 25 | num_bottleneck_blocks: 6 26 | estimate_occlusion_map: True 27 | dense_motion_params: 28 | block_expansion: 64 29 | max_features: 1024 30 | num_blocks: 5 31 | scale_factor: 0.25 32 | discriminator_params: 33 | scales: [1] 34 | block_expansion: 32 35 | max_features: 512 36 | num_blocks: 4 37 | sn: True 38 | 39 | train_params: 40 | num_epochs: 100 41 | num_repeats: 8 42 | epoch_milestones: [60, 90] 43 | lr_generator: 2.0e-4 44 | lr_discriminator: 2.0e-4 45 | lr_kp_detector: 2.0e-4 46 | batch_size: 36 47 | scales: [1, 0.5, 0.25, 0.125] 48 | checkpoint_freq: 50 49 | transform_params: 50 | sigma_affine: 0.05 51 | sigma_tps: 0.005 52 | points_tps: 5 53 | loss_weights: 54 | generator_gan: 1 55 | discriminator_gan: 1 56 | feature_matching: [10, 10, 10, 10] 57 | perceptual: [10, 10, 10, 10, 10] 58 | equivariance_value: 10 59 | equivariance_jacobian: 10 60 | 61 | reconstruction_params: 62 | num_videos: 1000 63 | format: '.mp4' 64 | 65 | animate_params: 66 | num_pairs: 50 67 | format: '.mp4' 68 | normalization_params: 69 | adapt_movement_scale: False 70 | use_relative_movement: True 71 | use_relative_jacobian: True 72 | 73 | visualizer_params: 74 | kp_size: 5 75 | draw_border: True 76 | colormap: 'gist_rainbow' 77 | -------------------------------------------------------------------------------- /config/taichi-256.yaml: -------------------------------------------------------------------------------- 1 | # Dataset parameters 2 | # Each dataset should contain 2 folders train and test 3 | # Each video can be represented as: 4 | # - an image of concatenated frames 5 | # - '.mp4' or '.gif' 6 | # - folder with all frames from a specific video 7 | # In case of Taichi. Same (youtube) video can be splitted in many parts (chunks). Each part has a following 8 | # format (id)#other#info.mp4. For example '12335#adsbf.mp4' has an id 12335. In case of TaiChi id stands for youtube 9 | # video id. 10 | dataset_params: 11 | # Path to data, data can be stored in several formats: .mp4 or .gif videos, stacked .png images or folders with frames. 12 | root_dir: data/taichi-png 13 | # Image shape, needed for staked .png format. 14 | frame_shape: [256, 256, 3] 15 | # In case of TaiChi single video can be splitted in many chunks, or the maybe several videos for single person. 16 | # In this case epoch can be a pass over different videos (if id_sampling=True) or over different chunks (if id_sampling=False) 17 | # If the name of the video '12335#adsbf.mp4' the id is assumed to be 12335 18 | id_sampling: True 19 | # List with pairs for animation, None for random pairs 20 | pairs_list: data/taichi256.csv 21 | # Augmentation parameters see augmentation.py for all posible augmentations 22 | augmentation_params: 23 | flip_param: 24 | horizontal_flip: True 25 | time_flip: True 26 | jitter_param: 27 | brightness: 0.1 28 | contrast: 0.1 29 | saturation: 0.1 30 | hue: 0.1 31 | 32 | # Defines model architecture 33 | model_params: 34 | common_params: 35 | # Number of keypoint 36 | num_kp: 10 37 | # Number of channels per image 38 | num_channels: 3 39 | # Using first or zero order model 40 | estimate_jacobian: True 41 | kp_detector_params: 42 | # Softmax temperature for keypoint heatmaps 43 | temperature: 0.1 44 | # Number of features mutliplier 45 | block_expansion: 32 46 | # Maximum allowed number of features 47 | max_features: 1024 48 | # Number of block in Unet. Can be increased or decreased depending or resolution. 49 | num_blocks: 5 50 | # Keypioint is predicted on smaller images for better performance, 51 | # scale_factor=0.25 means that 256x256 image will be resized to 64x64 52 | scale_factor: 0.25 53 | generator_params: 54 | # Number of features mutliplier 55 | block_expansion: 64 56 | # Maximum allowed number of features 57 | max_features: 512 58 | # Number of downsampling blocks in Jonson architecture. 59 | # Can be increased or decreased depending or resolution. 60 | num_down_blocks: 2 61 | # Number of ResBlocks in Jonson architecture. 62 | num_bottleneck_blocks: 6 63 | # Use occlusion map or not 64 | estimate_occlusion_map: True 65 | 66 | dense_motion_params: 67 | # Number of features mutliplier 68 | block_expansion: 64 69 | # Maximum allowed number of features 70 | max_features: 1024 71 | # Number of block in Unet. Can be increased or decreased depending or resolution. 72 | num_blocks: 5 73 | # Dense motion is predicted on smaller images for better performance, 74 | # scale_factor=0.25 means that 256x256 image will be resized to 64x64 75 | scale_factor: 0.25 76 | discriminator_params: 77 | # Discriminator can be multiscale, if you want 2 discriminator on original 78 | # resolution and half of the original, specify scales: [1, 0.5] 79 | scales: [1] 80 | # Number of features mutliplier 81 | block_expansion: 32 82 | # Maximum allowed number of features 83 | max_features: 512 84 | # Number of blocks. Can be increased or decreased depending or resolution. 85 | num_blocks: 4 86 | 87 | # Parameters of training 88 | train_params: 89 | # Number of training epochs 90 | num_epochs: 100 91 | # For better i/o performance when number of videos is small number of epochs can be multiplied by this number. 92 | # Thus effectivlly with num_repeats=100 each epoch is 100 times larger. 93 | num_repeats: 150 94 | # Drop learning rate by 10 times after this epochs 95 | epoch_milestones: [60, 90] 96 | # Initial learing rate for all modules 97 | lr_generator: 2.0e-4 98 | lr_discriminator: 2.0e-4 99 | lr_kp_detector: 2.0e-4 100 | batch_size: 30 101 | # Scales for perceptual pyramide loss. If scales = [1, 0.5, 0.25, 0.125] and image resolution is 256x256, 102 | # than the loss will be computer on resolutions 256x256, 128x128, 64x64, 32x32. 103 | scales: [1, 0.5, 0.25, 0.125] 104 | # Save checkpoint this frequently. If checkpoint_freq=50, checkpoint will be saved every 50 epochs. 105 | checkpoint_freq: 50 106 | # Parameters of transform for equivariance loss 107 | transform_params: 108 | # Sigma for affine part 109 | sigma_affine: 0.05 110 | # Sigma for deformation part 111 | sigma_tps: 0.005 112 | # Number of point in the deformation grid 113 | points_tps: 5 114 | loss_weights: 115 | # Weight for LSGAN loss in generator, 0 for no adversarial loss. 116 | generator_gan: 0 117 | # Weight for LSGAN loss in discriminator 118 | discriminator_gan: 1 119 | # Weights for feature matching loss, the number should be the same as number of blocks in discriminator. 120 | feature_matching: [10, 10, 10, 10] 121 | # Weights for perceptual loss. 122 | perceptual: [10, 10, 10, 10, 10] 123 | # Weights for value equivariance. 124 | equivariance_value: 10 125 | # Weights for jacobian equivariance. 126 | equivariance_jacobian: 10 127 | 128 | # Parameters of reconstruction 129 | reconstruction_params: 130 | # Maximum number of videos for reconstruction 131 | num_videos: 1000 132 | # Format for visualization, note that results will be also stored in staked .png. 133 | format: '.mp4' 134 | 135 | # Parameters of animation 136 | animate_params: 137 | # Maximum number of pairs for animation, the pairs will be either taken from pairs_list or random. 138 | num_pairs: 50 139 | # Format for visualization, note that results will be also stored in staked .png. 140 | format: '.mp4' 141 | # Normalization of diriving keypoints 142 | normalization_params: 143 | # Increase or decrease relative movement scale depending on the size of the object 144 | adapt_movement_scale: False 145 | # Apply only relative displacement of the keypoint 146 | use_relative_movement: True 147 | # Apply only relative change in jacobian 148 | use_relative_jacobian: True 149 | 150 | # Visualization parameters 151 | visualizer_params: 152 | # Draw keypoints of this size, increase or decrease depending on resolution 153 | kp_size: 5 154 | # Draw white border around images 155 | draw_border: True 156 | # Color map for keypoints 157 | colormap: 'gist_rainbow' 158 | -------------------------------------------------------------------------------- /config/taichi-adv-256.yaml: -------------------------------------------------------------------------------- 1 | # Dataset parameters 2 | dataset_params: 3 | # Path to data, data can be stored in several formats: .mp4 or .gif videos, stacked .png images or folders with frames. 4 | root_dir: data/taichi-png 5 | # Image shape, needed for staked .png format. 6 | frame_shape: [256, 256, 3] 7 | # In case of TaiChi single video can be splitted in many chunks, or the maybe several videos for single person. 8 | # In this case epoch can be a pass over different videos (if id_sampling=True) or over different chunks (if id_sampling=False) 9 | # If the name of the video '12335#adsbf.mp4' the id is assumed to be 12335 10 | id_sampling: True 11 | # List with pairs for animation, None for random pairs 12 | pairs_list: data/taichi256.csv 13 | # Augmentation parameters see augmentation.py for all posible augmentations 14 | augmentation_params: 15 | flip_param: 16 | horizontal_flip: True 17 | time_flip: True 18 | jitter_param: 19 | brightness: 0.1 20 | contrast: 0.1 21 | saturation: 0.1 22 | hue: 0.1 23 | 24 | # Defines model architecture 25 | model_params: 26 | common_params: 27 | # Number of keypoint 28 | num_kp: 10 29 | # Number of channels per image 30 | num_channels: 3 31 | # Using first or zero order model 32 | estimate_jacobian: True 33 | kp_detector_params: 34 | # Softmax temperature for keypoint heatmaps 35 | temperature: 0.1 36 | # Number of features mutliplier 37 | block_expansion: 32 38 | # Maximum allowed number of features 39 | max_features: 1024 40 | # Number of block in Unet. Can be increased or decreased depending or resolution. 41 | num_blocks: 5 42 | # Keypioint is predicted on smaller images for better performance, 43 | # scale_factor=0.25 means that 256x256 image will be resized to 64x64 44 | scale_factor: 0.25 45 | generator_params: 46 | # Number of features mutliplier 47 | block_expansion: 64 48 | # Maximum allowed number of features 49 | max_features: 512 50 | # Number of downsampling blocks in Jonson architecture. 51 | # Can be increased or decreased depending or resolution. 52 | num_down_blocks: 2 53 | # Number of ResBlocks in Jonson architecture. 54 | num_bottleneck_blocks: 6 55 | # Use occlusion map or not 56 | estimate_occlusion_map: True 57 | 58 | dense_motion_params: 59 | # Number of features mutliplier 60 | block_expansion: 64 61 | # Maximum allowed number of features 62 | max_features: 1024 63 | # Number of block in Unet. Can be increased or decreased depending or resolution. 64 | num_blocks: 5 65 | # Dense motion is predicted on smaller images for better performance, 66 | # scale_factor=0.25 means that 256x256 image will be resized to 64x64 67 | scale_factor: 0.25 68 | discriminator_params: 69 | # Discriminator can be multiscale, if you want 2 discriminator on original 70 | # resolution and half of the original, specify scales: [1, 0.5] 71 | scales: [1] 72 | # Number of features mutliplier 73 | block_expansion: 32 74 | # Maximum allowed number of features 75 | max_features: 512 76 | # Number of blocks. Can be increased or decreased depending or resolution. 77 | num_blocks: 4 78 | use_kp: True 79 | 80 | # Parameters of training 81 | train_params: 82 | # Number of training epochs 83 | num_epochs: 150 84 | # For better i/o performance when number of videos is small number of epochs can be multiplied by this number. 85 | # Thus effectivlly with num_repeats=100 each epoch is 100 times larger. 86 | num_repeats: 150 87 | # Drop learning rate by 10 times after this epochs 88 | epoch_milestones: [] 89 | # Initial learing rate for all modules 90 | lr_generator: 2.0e-4 91 | lr_discriminator: 2.0e-4 92 | lr_kp_detector: 0 93 | batch_size: 27 94 | # Scales for perceptual pyramide loss. If scales = [1, 0.5, 0.25, 0.125] and image resolution is 256x256, 95 | # than the loss will be computer on resolutions 256x256, 128x128, 64x64, 32x32. 96 | scales: [1, 0.5, 0.25, 0.125] 97 | # Save checkpoint this frequently. If checkpoint_freq=50, checkpoint will be saved every 50 epochs. 98 | checkpoint_freq: 50 99 | # Parameters of transform for equivariance loss 100 | transform_params: 101 | # Sigma for affine part 102 | sigma_affine: 0.05 103 | # Sigma for deformation part 104 | sigma_tps: 0.005 105 | # Number of point in the deformation grid 106 | points_tps: 5 107 | loss_weights: 108 | # Weight for LSGAN loss in generator 109 | generator_gan: 1 110 | # Weight for LSGAN loss in discriminator 111 | discriminator_gan: 1 112 | # Weights for feature matching loss, the number should be the same as number of blocks in discriminator. 113 | feature_matching: [10, 10, 10, 10] 114 | # Weights for perceptual loss. 115 | perceptual: [10, 10, 10, 10, 10] 116 | # Weights for value equivariance. 117 | equivariance_value: 10 118 | # Weights for jacobian equivariance. 119 | equivariance_jacobian: 10 120 | 121 | # Parameters of reconstruction 122 | reconstruction_params: 123 | # Maximum number of videos for reconstruction 124 | num_videos: 1000 125 | # Format for visualization, note that results will be also stored in staked .png. 126 | format: '.mp4' 127 | 128 | # Parameters of animation 129 | animate_params: 130 | # Maximum number of pairs for animation, the pairs will be either taken from pairs_list or random. 131 | num_pairs: 50 132 | # Format for visualization, note that results will be also stored in staked .png. 133 | format: '.mp4' 134 | # Normalization of diriving keypoints 135 | normalization_params: 136 | # Increase or decrease relative movement scale depending on the size of the object 137 | adapt_movement_scale: False 138 | # Apply only relative displacement of the keypoint 139 | use_relative_movement: True 140 | # Apply only relative change in jacobian 141 | use_relative_jacobian: True 142 | 143 | # Visualization parameters 144 | visualizer_params: 145 | # Draw keypoints of this size, increase or decrease depending on resolution 146 | kp_size: 5 147 | # Draw white border around images 148 | draw_border: True 149 | # Color map for keypoints 150 | colormap: 'gist_rainbow' 151 | -------------------------------------------------------------------------------- /config/vox-256.yaml: -------------------------------------------------------------------------------- 1 | dataset_params: 2 | root_dir: data/vox-png 3 | frame_shape: [256, 256, 3] 4 | id_sampling: True 5 | pairs_list: data/vox256.csv 6 | augmentation_params: 7 | flip_param: 8 | horizontal_flip: True 9 | time_flip: True 10 | jitter_param: 11 | brightness: 0.1 12 | contrast: 0.1 13 | saturation: 0.1 14 | hue: 0.1 15 | 16 | 17 | model_params: 18 | common_params: 19 | num_kp: 10 20 | num_channels: 3 21 | estimate_jacobian: True 22 | kp_detector_params: 23 | temperature: 0.1 24 | block_expansion: 32 25 | max_features: 1024 26 | scale_factor: 0.25 27 | num_blocks: 5 28 | generator_params: 29 | block_expansion: 64 30 | max_features: 512 31 | num_down_blocks: 2 32 | num_bottleneck_blocks: 6 33 | estimate_occlusion_map: True 34 | dense_motion_params: 35 | block_expansion: 64 36 | max_features: 1024 37 | num_blocks: 5 38 | scale_factor: 0.25 39 | discriminator_params: 40 | scales: [1] 41 | block_expansion: 32 42 | max_features: 512 43 | num_blocks: 4 44 | sn: True 45 | 46 | train_params: 47 | num_epochs: 100 48 | num_repeats: 75 49 | epoch_milestones: [60, 90] 50 | lr_generator: 2.0e-4 51 | lr_discriminator: 2.0e-4 52 | lr_kp_detector: 2.0e-4 53 | batch_size: 40 54 | scales: [1, 0.5, 0.25, 0.125] 55 | checkpoint_freq: 50 56 | transform_params: 57 | sigma_affine: 0.05 58 | sigma_tps: 0.005 59 | points_tps: 5 60 | loss_weights: 61 | generator_gan: 0 62 | discriminator_gan: 1 63 | feature_matching: [10, 10, 10, 10] 64 | perceptual: [10, 10, 10, 10, 10] 65 | equivariance_value: 10 66 | equivariance_jacobian: 10 67 | 68 | reconstruction_params: 69 | num_videos: 1000 70 | format: '.mp4' 71 | 72 | animate_params: 73 | num_pairs: 50 74 | format: '.mp4' 75 | normalization_params: 76 | adapt_movement_scale: False 77 | use_relative_movement: True 78 | use_relative_jacobian: True 79 | 80 | visualizer_params: 81 | kp_size: 5 82 | draw_border: True 83 | colormap: 'gist_rainbow' 84 | -------------------------------------------------------------------------------- /config/vox-adv-256.yaml: -------------------------------------------------------------------------------- 1 | dataset_params: 2 | root_dir: data/vox-png 3 | frame_shape: [256, 256, 3] 4 | id_sampling: True 5 | pairs_list: data/vox256.csv 6 | augmentation_params: 7 | flip_param: 8 | horizontal_flip: True 9 | time_flip: True 10 | jitter_param: 11 | brightness: 0.1 12 | contrast: 0.1 13 | saturation: 0.1 14 | hue: 0.1 15 | 16 | 17 | model_params: 18 | common_params: 19 | num_kp: 10 20 | num_channels: 3 21 | estimate_jacobian: True 22 | kp_detector_params: 23 | temperature: 0.1 24 | block_expansion: 32 25 | max_features: 1024 26 | scale_factor: 0.25 27 | num_blocks: 5 28 | generator_params: 29 | block_expansion: 64 30 | max_features: 512 31 | num_down_blocks: 2 32 | num_bottleneck_blocks: 6 33 | estimate_occlusion_map: True 34 | dense_motion_params: 35 | block_expansion: 64 36 | max_features: 1024 37 | num_blocks: 5 38 | scale_factor: 0.25 39 | discriminator_params: 40 | scales: [1] 41 | block_expansion: 32 42 | max_features: 512 43 | num_blocks: 4 44 | use_kp: True 45 | 46 | 47 | train_params: 48 | num_epochs: 150 49 | num_repeats: 75 50 | epoch_milestones: [] 51 | lr_generator: 2.0e-4 52 | lr_discriminator: 2.0e-4 53 | lr_kp_detector: 2.0e-4 54 | batch_size: 36 55 | scales: [1, 0.5, 0.25, 0.125] 56 | checkpoint_freq: 50 57 | transform_params: 58 | sigma_affine: 0.05 59 | sigma_tps: 0.005 60 | points_tps: 5 61 | loss_weights: 62 | generator_gan: 1 63 | discriminator_gan: 1 64 | feature_matching: [10, 10, 10, 10] 65 | perceptual: [10, 10, 10, 10, 10] 66 | equivariance_value: 10 67 | equivariance_jacobian: 10 68 | 69 | reconstruction_params: 70 | num_videos: 1000 71 | format: '.mp4' 72 | 73 | animate_params: 74 | num_pairs: 50 75 | format: '.mp4' 76 | normalization_params: 77 | adapt_movement_scale: False 78 | use_relative_movement: True 79 | use_relative_jacobian: True 80 | 81 | visualizer_params: 82 | kp_size: 5 83 | draw_border: True 84 | colormap: 'gist_rainbow' 85 | -------------------------------------------------------------------------------- /crop-video.py: -------------------------------------------------------------------------------- 1 | import face_alignment 2 | import skimage.io 3 | import numpy 4 | from argparse import ArgumentParser 5 | from skimage import img_as_ubyte 6 | from skimage.transform import resize 7 | from tqdm import tqdm 8 | import os 9 | import imageio 10 | import numpy as np 11 | import warnings 12 | warnings.filterwarnings("ignore") 13 | 14 | def extract_bbox(frame, fa): 15 | if max(frame.shape[0], frame.shape[1]) > 640: 16 | scale_factor = max(frame.shape[0], frame.shape[1]) / 640.0 17 | frame = resize(frame, (int(frame.shape[0] / scale_factor), int(frame.shape[1] / scale_factor))) 18 | frame = img_as_ubyte(frame) 19 | else: 20 | scale_factor = 1 21 | frame = frame[..., :3] 22 | bboxes = fa.face_detector.detect_from_image(frame[..., ::-1]) 23 | if len(bboxes) == 0: 24 | return [] 25 | return np.array(bboxes)[:, :-1] * scale_factor 26 | 27 | 28 | 29 | def bb_intersection_over_union(boxA, boxB): 30 | xA = max(boxA[0], boxB[0]) 31 | yA = max(boxA[1], boxB[1]) 32 | xB = min(boxA[2], boxB[2]) 33 | yB = min(boxA[3], boxB[3]) 34 | interArea = max(0, xB - xA + 1) * max(0, yB - yA + 1) 35 | boxAArea = (boxA[2] - boxA[0] + 1) * (boxA[3] - boxA[1] + 1) 36 | boxBArea = (boxB[2] - boxB[0] + 1) * (boxB[3] - boxB[1] + 1) 37 | iou = interArea / float(boxAArea + boxBArea - interArea) 38 | return iou 39 | 40 | 41 | def join(tube_bbox, bbox): 42 | xA = min(tube_bbox[0], bbox[0]) 43 | yA = min(tube_bbox[1], bbox[1]) 44 | xB = max(tube_bbox[2], bbox[2]) 45 | yB = max(tube_bbox[3], bbox[3]) 46 | return (xA, yA, xB, yB) 47 | 48 | 49 | def compute_bbox(start, end, fps, tube_bbox, frame_shape, inp, increase_area=0.1): 50 | left, top, right, bot = tube_bbox 51 | width = right - left 52 | height = bot - top 53 | 54 | #Computing aspect preserving bbox 55 | width_increase = max(increase_area, ((1 + 2 * increase_area) * height - width) / (2 * width)) 56 | height_increase = max(increase_area, ((1 + 2 * increase_area) * width - height) / (2 * height)) 57 | 58 | left = int(left - width_increase * width) 59 | top = int(top - height_increase * height) 60 | right = int(right + width_increase * width) 61 | bot = int(bot + height_increase * height) 62 | 63 | top, bot, left, right = max(0, top), min(bot, frame_shape[0]), max(0, left), min(right, frame_shape[1]) 64 | h, w = bot - top, right - left 65 | 66 | start = start / fps 67 | end = end / fps 68 | time = end - start 69 | 70 | return f'ffmpeg -i {inp} -ss {start} -t {time} -filter:v "crop={w}:{h}:{left}:{top}, scale=256:256" crop.mp4' 71 | 72 | 73 | def compute_bbox_trajectories(trajectories, fps, frame_shape, args): 74 | commands = [] 75 | for i, (bbox, tube_bbox, start, end) in enumerate(trajectories): 76 | if (end - start) > args.min_frames: 77 | command = compute_bbox(start, end, fps, tube_bbox, frame_shape, inp=args.inp, increase_area=args.increase) 78 | commands.append(command) 79 | return commands 80 | 81 | 82 | def process_video(args): 83 | fa = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D, flip_input=False) 84 | video = imageio.get_reader(args.inp) 85 | 86 | trajectories = [] 87 | previous_frame = None 88 | fps = video.get_meta_data()['fps'] 89 | commands = [] 90 | try: 91 | for i, frame in tqdm(enumerate(video)): 92 | frame_shape = frame.shape 93 | bboxes = extract_bbox(frame, fa) 94 | ## For each trajectory check the criterion 95 | not_valid_trajectories = [] 96 | valid_trajectories = [] 97 | 98 | for trajectory in trajectories: 99 | tube_bbox = trajectory[0] 100 | intersection = 0 101 | for bbox in bboxes: 102 | intersection = max(intersection, bb_intersection_over_union(tube_bbox, bbox)) 103 | if intersection > args.iou_with_initial: 104 | valid_trajectories.append(trajectory) 105 | else: 106 | not_valid_trajectories.append(trajectory) 107 | 108 | commands += compute_bbox_trajectories(not_valid_trajectories, fps, frame_shape, args) 109 | trajectories = valid_trajectories 110 | 111 | ## Assign bbox to trajectories, create new trajectories 112 | for bbox in bboxes: 113 | intersection = 0 114 | current_trajectory = None 115 | for trajectory in trajectories: 116 | tube_bbox = trajectory[0] 117 | current_intersection = bb_intersection_over_union(tube_bbox, bbox) 118 | if intersection < current_intersection and current_intersection > args.iou_with_initial: 119 | intersection = bb_intersection_over_union(tube_bbox, bbox) 120 | current_trajectory = trajectory 121 | 122 | ## Create new trajectory 123 | if current_trajectory is None: 124 | trajectories.append([bbox, bbox, i, i]) 125 | else: 126 | current_trajectory[3] = i 127 | current_trajectory[1] = join(current_trajectory[1], bbox) 128 | 129 | 130 | except IndexError as e: 131 | raise (e) 132 | 133 | commands += compute_bbox_trajectories(trajectories, fps, frame_shape, args) 134 | return commands 135 | 136 | 137 | if __name__ == "__main__": 138 | parser = ArgumentParser() 139 | 140 | parser.add_argument("--image_shape", default=(256, 256), type=lambda x: tuple(map(int, x.split(','))), 141 | help="Image shape") 142 | parser.add_argument("--increase", default=0.1, type=float, help='Increase bbox by this amount') 143 | parser.add_argument("--iou_with_initial", type=float, default=0.25, help="The minimal allowed iou with inital bbox") 144 | parser.add_argument("--inp", required=True, help='Input image or video') 145 | parser.add_argument("--min_frames", type=int, default=150, help='Minimum number of frames') 146 | 147 | 148 | args = parser.parse_args() 149 | 150 | commands = process_video(args) 151 | for command in commands: 152 | print (command) 153 | 154 | -------------------------------------------------------------------------------- /data/bair256.csv: -------------------------------------------------------------------------------- 1 | distance,source,driving,frame 2 | 0,000054.mp4,000048.mp4,0 3 | 0,000050.mp4,000063.mp4,0 4 | 0,000073.mp4,000007.mp4,0 5 | 0,000021.mp4,000010.mp4,0 6 | 0,000084.mp4,000046.mp4,0 7 | 0,000031.mp4,000102.mp4,0 8 | 0,000029.mp4,000111.mp4,0 9 | 0,000090.mp4,000112.mp4,0 10 | 0,000039.mp4,000010.mp4,0 11 | 0,000008.mp4,000069.mp4,0 12 | 0,000068.mp4,000076.mp4,0 13 | 0,000051.mp4,000052.mp4,0 14 | 0,000022.mp4,000098.mp4,0 15 | 0,000096.mp4,000032.mp4,0 16 | 0,000032.mp4,000099.mp4,0 17 | 0,000006.mp4,000053.mp4,0 18 | 0,000098.mp4,000020.mp4,0 19 | 0,000029.mp4,000066.mp4,0 20 | 0,000022.mp4,000007.mp4,0 21 | 0,000027.mp4,000065.mp4,0 22 | 0,000026.mp4,000059.mp4,0 23 | 0,000015.mp4,000112.mp4,0 24 | 0,000086.mp4,000123.mp4,0 25 | 0,000103.mp4,000052.mp4,0 26 | 0,000123.mp4,000103.mp4,0 27 | 0,000051.mp4,000005.mp4,0 28 | 0,000062.mp4,000125.mp4,0 29 | 0,000126.mp4,000111.mp4,0 30 | 0,000066.mp4,000090.mp4,0 31 | 0,000075.mp4,000106.mp4,0 32 | 0,000020.mp4,000010.mp4,0 33 | 0,000076.mp4,000028.mp4,0 34 | 0,000062.mp4,000002.mp4,0 35 | 0,000095.mp4,000127.mp4,0 36 | 0,000113.mp4,000072.mp4,0 37 | 0,000027.mp4,000104.mp4,0 38 | 0,000054.mp4,000124.mp4,0 39 | 0,000019.mp4,000089.mp4,0 40 | 0,000052.mp4,000072.mp4,0 41 | 0,000108.mp4,000033.mp4,0 42 | 0,000044.mp4,000118.mp4,0 43 | 0,000029.mp4,000086.mp4,0 44 | 0,000068.mp4,000066.mp4,0 45 | 0,000014.mp4,000036.mp4,0 46 | 0,000053.mp4,000071.mp4,0 47 | 0,000022.mp4,000094.mp4,0 48 | 0,000000.mp4,000121.mp4,0 49 | 0,000071.mp4,000079.mp4,0 50 | 0,000127.mp4,000005.mp4,0 51 | 0,000085.mp4,000023.mp4,0 52 | -------------------------------------------------------------------------------- /data/taichi-loading/README.md: -------------------------------------------------------------------------------- 1 | # TaiChi dataset 2 | 3 | The scripst for loading the TaiChi dataset. 4 | 5 | We provide only the id of the corresponding video and the bounding box. Following script will download videos from youtube and crop them according to the provided bounding boxes. 6 | 7 | 1) Load youtube-dl: 8 | ``` 9 | wget https://yt-dl.org/downloads/latest/youtube-dl -O youtube-dl 10 | chmod a+rx youtube-dl 11 | ``` 12 | 13 | 2) Run script to download videos, there are 2 formats that can be used for storing videos one is .mp4 and another is folder with .png images. While .png images occupy significantly more space, the format is loss-less and have better i/o performance when training. 14 | 15 | ``` 16 | python load_videos.py --metadata taichi-metadata.csv --format .mp4 --out_folder taichi --workers 8 17 | ``` 18 | select number of workers based on number of cpu avaliable. Note .png format take aproximatly 80GB. 19 | -------------------------------------------------------------------------------- /data/taichi-loading/load_videos.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import pandas as pd 3 | import imageio 4 | import os 5 | import subprocess 6 | from multiprocessing import Pool 7 | from itertools import cycle 8 | import warnings 9 | import glob 10 | import time 11 | from tqdm import tqdm 12 | from argparse import ArgumentParser 13 | from skimage import img_as_ubyte 14 | from skimage.transform import resize 15 | warnings.filterwarnings("ignore") 16 | 17 | DEVNULL = open(os.devnull, 'wb') 18 | 19 | 20 | def save(path, frames, format): 21 | if format == '.mp4': 22 | imageio.mimsave(path, frames) 23 | elif format == '.png': 24 | if os.path.exists(path): 25 | print ("Warning: skiping video %s" % os.path.basename(path)) 26 | return 27 | else: 28 | os.makedirs(path) 29 | for j, frame in enumerate(frames): 30 | imageio.imsave(os.path.join(path, str(j).zfill(7) + '.png'), frames[j]) 31 | else: 32 | print ("Unknown format %s" % format) 33 | exit() 34 | 35 | 36 | def download(video_id, args): 37 | video_path = os.path.join(args.video_folder, video_id + ".mp4") 38 | subprocess.call([args.youtube, '-f', "''best/mp4''", '--write-auto-sub', '--write-sub', 39 | '--sub-lang', 'en', '--skip-unavailable-fragments', 40 | "https://www.youtube.com/watch?v=" + video_id, "--output", 41 | video_path], stdout=DEVNULL, stderr=DEVNULL) 42 | return video_path 43 | 44 | 45 | def run(data): 46 | video_id, args = data 47 | if not os.path.exists(os.path.join(args.video_folder, video_id.split('#')[0] + '.mp4')): 48 | download(video_id.split('#')[0], args) 49 | 50 | if not os.path.exists(os.path.join(args.video_folder, video_id.split('#')[0] + '.mp4')): 51 | print ('Can not load video %s, broken link' % video_id.split('#')[0]) 52 | return 53 | reader = imageio.get_reader(os.path.join(args.video_folder, video_id.split('#')[0] + '.mp4')) 54 | fps = reader.get_meta_data()['fps'] 55 | 56 | df = pd.read_csv(args.metadata) 57 | df = df[df['video_id'] == video_id] 58 | 59 | all_chunks_dict = [{'start': df['start'].iloc[j], 'end': df['end'].iloc[j], 60 | 'bbox': list(map(int, df['bbox'].iloc[j].split('-'))), 'frames':[]} for j in range(df.shape[0])] 61 | ref_fps = df['fps'].iloc[0] 62 | ref_height = df['height'].iloc[0] 63 | ref_width = df['width'].iloc[0] 64 | partition = df['partition'].iloc[0] 65 | try: 66 | for i, frame in enumerate(reader): 67 | for entry in all_chunks_dict: 68 | if (i * ref_fps >= entry['start'] * fps) and (i * ref_fps < entry['end'] * fps): 69 | left, top, right, bot = entry['bbox'] 70 | left = int(left / (ref_width / frame.shape[1])) 71 | top = int(top / (ref_height / frame.shape[0])) 72 | right = int(right / (ref_width / frame.shape[1])) 73 | bot = int(bot / (ref_height / frame.shape[0])) 74 | crop = frame[top:bot, left:right] 75 | if args.image_shape is not None: 76 | crop = img_as_ubyte(resize(crop, args.image_shape, anti_aliasing=True)) 77 | entry['frames'].append(crop) 78 | except imageio.core.format.CannotReadFrameError: 79 | None 80 | 81 | for entry in all_chunks_dict: 82 | first_part = '#'.join(video_id.split('#')[::-1]) 83 | path = first_part + '#' + str(entry['start']).zfill(6) + '#' + str(entry['end']).zfill(6) + '.mp4' 84 | save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format) 85 | 86 | 87 | if __name__ == "__main__": 88 | parser = ArgumentParser() 89 | parser.add_argument("--video_folder", default='youtube-taichi', help='Path to youtube videos') 90 | parser.add_argument("--metadata", default='taichi-metadata-new.csv', help='Path to metadata') 91 | parser.add_argument("--out_folder", default='taichi-png', help='Path to output') 92 | parser.add_argument("--format", default='.png', help='Storing format') 93 | parser.add_argument("--workers", default=1, type=int, help='Number of workers') 94 | parser.add_argument("--youtube", default='./youtube-dl', help='Path to youtube-dl') 95 | 96 | parser.add_argument("--image_shape", default=(256, 256), type=lambda x: tuple(map(int, x.split(','))), 97 | help="Image shape, None for no resize") 98 | 99 | args = parser.parse_args() 100 | if not os.path.exists(args.video_folder): 101 | os.makedirs(args.video_folder) 102 | if not os.path.exists(args.out_folder): 103 | os.makedirs(args.out_folder) 104 | for partition in ['test', 'train']: 105 | if not os.path.exists(os.path.join(args.out_folder, partition)): 106 | os.makedirs(os.path.join(args.out_folder, partition)) 107 | 108 | df = pd.read_csv(args.metadata) 109 | video_ids = set(df['video_id']) 110 | pool = Pool(processes=args.workers) 111 | args_list = cycle([args]) 112 | for chunks_data in tqdm(pool.imap_unordered(run, zip(video_ids, args_list))): 113 | None 114 | -------------------------------------------------------------------------------- /data/taichi256.csv: -------------------------------------------------------------------------------- 1 | distance,source,driving,frame 2 | 3.54437869822485,ab28GAufK8o#000261#000596.mp4,aDyyTMUBoLE#000164#000351.mp4,0 3 | 2.8639053254437887,DMEaUoA8EPE#000028#000354.mp4,0Q914by5A98#010440#010764.mp4,0 4 | 2.153846153846153,L82WHgYRq6I#000021#000479.mp4,0Q914by5A98#010440#010764.mp4,0 5 | 2.8994082840236666,oNkBx4CZuEg#000000#001024.mp4,DMEaUoA8EPE#000028#000354.mp4,0 6 | 3.3905325443786998,ab28GAufK8o#000261#000596.mp4,uEqWZ9S_-Lw#000089#000581.mp4,0 7 | 3.266272189349112,0Q914by5A98#010440#010764.mp4,ab28GAufK8o#000261#000596.mp4,0 8 | 2.7514792899408294,WlDYrq8K6nk#008186#008512.mp4,OiblkvkAHWM#014331#014459.mp4,0 9 | 3.0177514792899407,oNkBx4CZuEg#001024#002048.mp4,aDyyTMUBoLE#000375#000518.mp4,0 10 | 3.4792899408284064,aDyyTMUBoLE#000164#000351.mp4,w2awOCDRtrc#001729#002009.mp4,0 11 | 2.769230769230769,oNkBx4CZuEg#000000#001024.mp4,L82WHgYRq6I#000021#000479.mp4,0 12 | 3.8047337278106514,ab28GAufK8o#000261#000596.mp4,w2awOCDRtrc#001729#002009.mp4,0 13 | 3.4260355029585763,w2awOCDRtrc#001729#002009.mp4,oNkBx4CZuEg#000000#001024.mp4,0 14 | 3.313609467455621,DMEaUoA8EPE#000028#000354.mp4,WlDYrq8K6nk#005943#006135.mp4,0 15 | 3.8402366863905333,oNkBx4CZuEg#001024#002048.mp4,ab28GAufK8o#000261#000596.mp4,0 16 | 3.3254437869822504,aDyyTMUBoLE#000164#000351.mp4,oNkBx4CZuEg#000000#001024.mp4,0 17 | 1.2485207100591724,0Q914by5A98#010440#010764.mp4,aDyyTMUBoLE#000164#000351.mp4,0 18 | 3.804733727810652,OiblkvkAHWM#006251#006533.mp4,aDyyTMUBoLE#000375#000518.mp4,0 19 | 3.662721893491124,uEqWZ9S_-Lw#000089#000581.mp4,DMEaUoA8EPE#000028#000354.mp4,0 20 | 3.230769230769233,A3ZmT97hAWU#000095#000678.mp4,ab28GAufK8o#000261#000596.mp4,0 21 | 3.3668639053254434,w81Tr0Dp1K8#015329#015485.mp4,WlDYrq8K6nk#008186#008512.mp4,0 22 | 3.313609467455621,WlDYrq8K6nk#005943#006135.mp4,DMEaUoA8EPE#000028#000354.mp4,0 23 | 2.7514792899408294,OiblkvkAHWM#014331#014459.mp4,WlDYrq8K6nk#008186#008512.mp4,0 24 | 1.964497041420118,L82WHgYRq6I#000021#000479.mp4,DMEaUoA8EPE#000028#000354.mp4,0 25 | 3.78698224852071,FBuF0xOal9M#046824#047542.mp4,lCb5w6n8kPs#011879#012014.mp4,0 26 | 3.92307692307692,ab28GAufK8o#000261#000596.mp4,L82WHgYRq6I#000021#000479.mp4,0 27 | 3.8402366863905333,ab28GAufK8o#000261#000596.mp4,oNkBx4CZuEg#001024#002048.mp4,0 28 | 3.828402366863905,ab28GAufK8o#000261#000596.mp4,OiblkvkAHWM#006251#006533.mp4,0 29 | 2.041420118343196,L82WHgYRq6I#000021#000479.mp4,aDyyTMUBoLE#000164#000351.mp4,0 30 | 3.2485207100591724,0Q914by5A98#010440#010764.mp4,w2awOCDRtrc#001729#002009.mp4,0 31 | 3.2485207100591746,oNkBx4CZuEg#000000#001024.mp4,0Q914by5A98#010440#010764.mp4,0 32 | 1.964497041420118,DMEaUoA8EPE#000028#000354.mp4,L82WHgYRq6I#000021#000479.mp4,0 33 | 3.5266272189349115,kgvcI9oe3NI#001578#001763.mp4,lCb5w6n8kPs#004451#004631.mp4,0 34 | 3.005917159763317,A3ZmT97hAWU#000095#000678.mp4,0Q914by5A98#010440#010764.mp4,0 35 | 3.230769230769233,ab28GAufK8o#000261#000596.mp4,A3ZmT97hAWU#000095#000678.mp4,0 36 | 3.5266272189349115,lCb5w6n8kPs#004451#004631.mp4,kgvcI9oe3NI#001578#001763.mp4,0 37 | 2.769230769230769,L82WHgYRq6I#000021#000479.mp4,oNkBx4CZuEg#000000#001024.mp4,0 38 | 3.165680473372782,WlDYrq8K6nk#005943#006135.mp4,w81Tr0Dp1K8#001375#001516.mp4,0 39 | 2.8994082840236666,DMEaUoA8EPE#000028#000354.mp4,oNkBx4CZuEg#000000#001024.mp4,0 40 | 2.4556213017751523,0Q914by5A98#010440#010764.mp4,mndSqTrxpts#000000#000175.mp4,0 41 | 2.201183431952659,A3ZmT97hAWU#000095#000678.mp4,VMSqvTE90hk#007168#007312.mp4,0 42 | 3.8047337278106514,w2awOCDRtrc#001729#002009.mp4,ab28GAufK8o#000261#000596.mp4,0 43 | 3.769230769230769,uEqWZ9S_-Lw#000089#000581.mp4,0Q914by5A98#010440#010764.mp4,0 44 | 3.6568047337278102,A3ZmT97hAWU#000095#000678.mp4,aDyyTMUBoLE#000164#000351.mp4,0 45 | 3.7869822485207107,uEqWZ9S_-Lw#000089#000581.mp4,L82WHgYRq6I#000021#000479.mp4,0 46 | 3.78698224852071,lCb5w6n8kPs#011879#012014.mp4,FBuF0xOal9M#046824#047542.mp4,0 47 | 3.591715976331361,nAQEOC1Z10M#020177#020600.mp4,w81Tr0Dp1K8#004036#004218.mp4,0 48 | 3.8757396449704156,uEqWZ9S_-Lw#000089#000581.mp4,aDyyTMUBoLE#000164#000351.mp4,0 49 | 2.45562130177515,aDyyTMUBoLE#000164#000351.mp4,DMEaUoA8EPE#000028#000354.mp4,0 50 | 3.5502958579881647,uEqWZ9S_-Lw#000089#000581.mp4,OiblkvkAHWM#006251#006533.mp4,0 51 | 3.7928994082840224,aDyyTMUBoLE#000375#000518.mp4,ab28GAufK8o#000261#000596.mp4,0 52 | -------------------------------------------------------------------------------- /demo.py: -------------------------------------------------------------------------------- 1 | import matplotlib 2 | matplotlib.use('Agg') 3 | import os 4 | import yaml 5 | from argparse import ArgumentParser 6 | from tqdm import tqdm 7 | 8 | import imageio 9 | import numpy as np 10 | from skimage.transform import resize 11 | 12 | import torch 13 | from sync_batchnorm import DataParallelWithCallback 14 | 15 | from modules.generator import OcclusionAwareGenerator 16 | from modules.keypoint_detector import KPDetector 17 | from animate import normalize_kp 18 | from scipy.spatial import ConvexHull 19 | 20 | 21 | 22 | def load_checkpoints(config_path, checkpoint_path): 23 | 24 | with open(config_path) as f: 25 | config = yaml.load(f) 26 | 27 | generator = OcclusionAwareGenerator(**config['model_params']['generator_params'], 28 | **config['model_params']['common_params']) 29 | generator.cuda() 30 | 31 | kp_detector = KPDetector(**config['model_params']['kp_detector_params'], 32 | **config['model_params']['common_params']) 33 | kp_detector.cuda() 34 | 35 | checkpoint = torch.load(checkpoint_path) 36 | generator.load_state_dict(checkpoint['generator']) 37 | kp_detector.load_state_dict(checkpoint['kp_detector']) 38 | 39 | generator = DataParallelWithCallback(generator) 40 | kp_detector = DataParallelWithCallback(kp_detector) 41 | 42 | generator.eval() 43 | kp_detector.eval() 44 | 45 | return generator, kp_detector 46 | 47 | 48 | def make_animation(source_image, driving_video, generator, kp_detector, relative=True, adapt_movement_scale=True): 49 | with torch.no_grad(): 50 | predictions = [] 51 | source = torch.tensor(source_image[np.newaxis].astype(np.float32)).permute(0, 3, 1, 2).cuda() 52 | driving = torch.tensor(np.array(driving_video)[np.newaxis].astype(np.float32)).permute(0, 4, 1, 2, 3).cuda() 53 | kp_source = kp_detector(source) 54 | kp_driving_initial = kp_detector(driving[:, :, 0]) 55 | 56 | for frame_idx in tqdm(range(driving.shape[2])): 57 | driving_frame = driving[:, :, frame_idx] 58 | kp_driving = kp_detector(driving_frame) 59 | kp_norm = normalize_kp(kp_source=kp_source, kp_driving=kp_driving, 60 | kp_driving_initial=kp_driving_initial, use_relative_movement=relative, 61 | use_relative_jacobian=relative, adapt_movement_scale=adapt_movement_scale) 62 | out = generator(source, kp_source=kp_source, kp_driving=kp_norm) 63 | 64 | predictions.append(np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0]) 65 | return predictions 66 | 67 | def find_best_frame(source, driving): 68 | import face_alignment 69 | 70 | def normalize_kp(kp): 71 | kp = kp - kp.mean(axis=0, keepdims=True) 72 | area = ConvexHull(kp[:, :2]).volume 73 | area = np.sqrt(area) 74 | kp[:, :2] = kp[:, :2] / area 75 | return kp 76 | 77 | fa = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D, flip_input=True) 78 | kp_source = fa.get_landmarks(255 * source)[0] 79 | kp_source = normalize_kp(kp_source) 80 | norm = float('inf') 81 | frame_num = 0 82 | for i, image in tqdm(enumerate(driving)): 83 | kp_driving = fa.get_landmarks(255 * image)[0] 84 | kp_driving = normalize_kp(kp_driving) 85 | new_norm = (np.abs(kp_source - kp_driving) ** 2).sum() 86 | if new_norm < norm: 87 | norm = new_norm 88 | frame_num = i 89 | return frame_num 90 | 91 | if __name__ == "__main__": 92 | parser = ArgumentParser() 93 | parser.add_argument("--config", required=True, help="path to config") 94 | parser.add_argument("--checkpoint", default='vox-cpk.pth.tar', help="path to checkpoint to restore") 95 | 96 | parser.add_argument("--source_image", default='sup-mat/source.png', help="path to source image") 97 | parser.add_argument("--driving_video", default='sup-mat/source.png', help="path to driving video") 98 | parser.add_argument("--result_video", default='result.mp4', help="path to output") 99 | 100 | parser.add_argument("--relative", dest="relative", action="store_true", help="use relative or absolute keypoint coordinates") 101 | parser.add_argument("--adapt_scale", dest="adapt_scale", action="store_true", help="adapt movement scale based on convex hull of keypoints") 102 | 103 | parser.add_argument("--find_best_frame", dest="find_best_frame", action="store_true", 104 | help="Generate from the frame that is the most alligned with source. (Only for faces, requires face_aligment lib)") 105 | 106 | parser.add_argument("--best_frame", dest="best_frame", type=int, default=None, 107 | help="Set frame to start from.") 108 | 109 | 110 | parser.set_defaults(relative=False) 111 | parser.set_defaults(adapt_scale=False) 112 | 113 | opt = parser.parse_args() 114 | 115 | source_image = imageio.imread(opt.source_image) 116 | reader = imageio.get_reader(opt.driving_video) 117 | fps = reader.get_meta_data()['fps'] 118 | reader.close() 119 | driving_video = imageio.mimread(opt.driving_video, memtest=False) 120 | 121 | source_image = resize(source_image, (256, 256))[..., :3] 122 | driving_video = [resize(frame, (256, 256))[..., :3] for frame in driving_video] 123 | generator, kp_detector = load_checkpoints(config_path=opt.config, checkpoint_path=opt.checkpoint) 124 | 125 | if opt.find_best_frame or opt.best_frame is not None: 126 | i = opt.best_frame if opt.best_frame is not None else find_best_frame(source_image, driving_video) 127 | print (i) 128 | driving_forward = driving_video[i:] 129 | driving_backward = driving_video[:(i+1)][::-1] 130 | predictions_forward = make_animation(source_image, driving_forward, generator, kp_detector, relative=opt.relative, adapt_movement_scale=opt.adapt_scale) 131 | predictions_backward = make_animation(source_image, driving_backward, generator, kp_detector, relative=opt.relative, adapt_movement_scale=opt.adapt_scale) 132 | predictions = predictions_backward[::-1] + predictions_forward[1:] 133 | else: 134 | predictions = make_animation(source_image, driving_video, generator, kp_detector, relative=opt.relative, adapt_movement_scale=opt.adapt_scale) 135 | imageio.mimsave(opt.result_video, predictions, fps=fps) 136 | 137 | -------------------------------------------------------------------------------- /frames_dataset.py: -------------------------------------------------------------------------------- 1 | import os 2 | from skimage import io, img_as_float32 3 | from skimage.color import gray2rgb 4 | from sklearn.model_selection import train_test_split 5 | from imageio import mimread 6 | 7 | import numpy as np 8 | from torch.utils.data import Dataset 9 | import pandas as pd 10 | from augmentation import AllAugmentationTransform 11 | import glob 12 | 13 | 14 | def read_video(name, frame_shape): 15 | """ 16 | Read video which can be: 17 | - an image of concatenated frames 18 | - '.mp4' and'.gif' 19 | - folder with videos 20 | """ 21 | 22 | if os.path.isdir(name): 23 | frames = sorted(os.listdir(name)) 24 | num_frames = len(frames) 25 | video_array = np.array( 26 | [img_as_float32(io.imread(os.path.join(name, frames[idx]))) for idx in range(num_frames)]) 27 | elif name.lower().endswith('.png') or name.lower().endswith('.jpg'): 28 | image = io.imread(name) 29 | 30 | if len(image.shape) == 2 or image.shape[2] == 1: 31 | image = gray2rgb(image) 32 | 33 | if image.shape[2] == 4: 34 | image = image[..., :3] 35 | 36 | image = img_as_float32(image) 37 | 38 | video_array = np.moveaxis(image, 1, 0) 39 | 40 | video_array = video_array.reshape((-1,) + frame_shape) 41 | video_array = np.moveaxis(video_array, 1, 2) 42 | elif name.lower().endswith('.gif') or name.lower().endswith('.mp4') or name.lower().endswith('.mov'): 43 | video = np.array(mimread(name)) 44 | if len(video.shape) == 3: 45 | video = np.array([gray2rgb(frame) for frame in video]) 46 | if video.shape[-1] == 4: 47 | video = video[..., :3] 48 | video_array = img_as_float32(video) 49 | else: 50 | raise Exception("Unknown file extensions %s" % name) 51 | 52 | return video_array 53 | 54 | 55 | class FramesDataset(Dataset): 56 | """ 57 | Dataset of videos, each video can be represented as: 58 | - an image of concatenated frames 59 | - '.mp4' or '.gif' 60 | - folder with all frames 61 | """ 62 | 63 | def __init__(self, root_dir, frame_shape=(256, 256, 3), id_sampling=False, is_train=True, 64 | random_seed=0, pairs_list=None, augmentation_params=None): 65 | self.root_dir = root_dir 66 | self.videos = os.listdir(root_dir) 67 | self.frame_shape = tuple(frame_shape) 68 | self.pairs_list = pairs_list 69 | self.id_sampling = id_sampling 70 | if os.path.exists(os.path.join(root_dir, 'train')): 71 | assert os.path.exists(os.path.join(root_dir, 'test')) 72 | print("Use predefined train-test split.") 73 | if id_sampling: 74 | train_videos = {os.path.basename(video).split('#')[0] for video in 75 | os.listdir(os.path.join(root_dir, 'train'))} 76 | train_videos = list(train_videos) 77 | else: 78 | train_videos = os.listdir(os.path.join(root_dir, 'train')) 79 | test_videos = os.listdir(os.path.join(root_dir, 'test')) 80 | self.root_dir = os.path.join(self.root_dir, 'train' if is_train else 'test') 81 | else: 82 | print("Use random train-test split.") 83 | train_videos, test_videos = train_test_split(self.videos, random_state=random_seed, test_size=0.2) 84 | 85 | if is_train: 86 | self.videos = train_videos 87 | else: 88 | self.videos = test_videos 89 | 90 | self.is_train = is_train 91 | 92 | if self.is_train: 93 | self.transform = AllAugmentationTransform(**augmentation_params) 94 | else: 95 | self.transform = None 96 | 97 | def __len__(self): 98 | return len(self.videos) 99 | 100 | def __getitem__(self, idx): 101 | if self.is_train and self.id_sampling: 102 | name = self.videos[idx] 103 | path = np.random.choice(glob.glob(os.path.join(self.root_dir, name + '*.mp4'))) 104 | else: 105 | name = self.videos[idx] 106 | path = os.path.join(self.root_dir, name) 107 | 108 | video_name = os.path.basename(path) 109 | 110 | if self.is_train and os.path.isdir(path): 111 | frames = os.listdir(path) 112 | num_frames = len(frames) 113 | frame_idx = np.sort(np.random.choice(num_frames, replace=True, size=2)) 114 | video_array = [img_as_float32(io.imread(os.path.join(path, frames[idx]))) for idx in frame_idx] 115 | else: 116 | video_array = read_video(path, frame_shape=self.frame_shape) 117 | num_frames = len(video_array) 118 | frame_idx = np.sort(np.random.choice(num_frames, replace=True, size=2)) if self.is_train else range( 119 | num_frames) 120 | video_array = video_array[frame_idx] 121 | 122 | if self.transform is not None: 123 | video_array = self.transform(video_array) 124 | 125 | out = {} 126 | if self.is_train: 127 | source = np.array(video_array[0], dtype='float32') 128 | driving = np.array(video_array[1], dtype='float32') 129 | 130 | out['driving'] = driving.transpose((2, 0, 1)) 131 | out['source'] = source.transpose((2, 0, 1)) 132 | else: 133 | video = np.array(video_array, dtype='float32') 134 | out['video'] = video.transpose((3, 0, 1, 2)) 135 | 136 | out['name'] = video_name 137 | 138 | return out 139 | 140 | 141 | class DatasetRepeater(Dataset): 142 | """ 143 | Pass several times over the same dataset for better i/o performance 144 | """ 145 | 146 | def __init__(self, dataset, num_repeats=100): 147 | self.dataset = dataset 148 | self.num_repeats = num_repeats 149 | 150 | def __len__(self): 151 | return self.num_repeats * self.dataset.__len__() 152 | 153 | def __getitem__(self, idx): 154 | return self.dataset[idx % self.dataset.__len__()] 155 | 156 | 157 | class PairedDataset(Dataset): 158 | """ 159 | Dataset of pairs for animation. 160 | """ 161 | 162 | def __init__(self, initial_dataset, number_of_pairs, seed=0): 163 | self.initial_dataset = initial_dataset 164 | pairs_list = self.initial_dataset.pairs_list 165 | 166 | np.random.seed(seed) 167 | 168 | if pairs_list is None: 169 | max_idx = min(number_of_pairs, len(initial_dataset)) 170 | nx, ny = max_idx, max_idx 171 | xy = np.mgrid[:nx, :ny].reshape(2, -1).T 172 | number_of_pairs = min(xy.shape[0], number_of_pairs) 173 | self.pairs = xy.take(np.random.choice(xy.shape[0], number_of_pairs, replace=False), axis=0) 174 | else: 175 | videos = self.initial_dataset.videos 176 | name_to_index = {name: index for index, name in enumerate(videos)} 177 | pairs = pd.read_csv(pairs_list) 178 | pairs = pairs[np.logical_and(pairs['source'].isin(videos), pairs['driving'].isin(videos))] 179 | 180 | number_of_pairs = min(pairs.shape[0], number_of_pairs) 181 | self.pairs = [] 182 | self.start_frames = [] 183 | for ind in range(number_of_pairs): 184 | self.pairs.append( 185 | (name_to_index[pairs['driving'].iloc[ind]], name_to_index[pairs['source'].iloc[ind]])) 186 | 187 | def __len__(self): 188 | return len(self.pairs) 189 | 190 | def __getitem__(self, idx): 191 | pair = self.pairs[idx] 192 | first = self.initial_dataset[pair[0]] 193 | second = self.initial_dataset[pair[1]] 194 | first = {'driving_' + key: value for key, value in first.items()} 195 | second = {'source_' + key: value for key, value in second.items()} 196 | 197 | return {**first, **second} 198 | -------------------------------------------------------------------------------- /logger.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | import torch.nn.functional as F 4 | import imageio 5 | 6 | import os 7 | from skimage.draw import circle 8 | 9 | import matplotlib.pyplot as plt 10 | import collections 11 | 12 | 13 | class Logger: 14 | def __init__(self, log_dir, checkpoint_freq=100, visualizer_params=None, zfill_num=8, log_file_name='log.txt'): 15 | 16 | self.loss_list = [] 17 | self.cpk_dir = log_dir 18 | self.visualizations_dir = os.path.join(log_dir, 'train-vis') 19 | if not os.path.exists(self.visualizations_dir): 20 | os.makedirs(self.visualizations_dir) 21 | self.log_file = open(os.path.join(log_dir, log_file_name), 'a') 22 | self.zfill_num = zfill_num 23 | self.visualizer = Visualizer(**visualizer_params) 24 | self.checkpoint_freq = checkpoint_freq 25 | self.epoch = 0 26 | self.best_loss = float('inf') 27 | self.names = None 28 | 29 | def log_scores(self, loss_names): 30 | loss_mean = np.array(self.loss_list).mean(axis=0) 31 | 32 | loss_string = "; ".join(["%s - %.5f" % (name, value) for name, value in zip(loss_names, loss_mean)]) 33 | loss_string = str(self.epoch).zfill(self.zfill_num) + ") " + loss_string 34 | 35 | print(loss_string, file=self.log_file) 36 | self.loss_list = [] 37 | self.log_file.flush() 38 | 39 | def visualize_rec(self, inp, out): 40 | image = self.visualizer.visualize(inp['driving'], inp['source'], out) 41 | imageio.imsave(os.path.join(self.visualizations_dir, "%s-rec.png" % str(self.epoch).zfill(self.zfill_num)), image) 42 | 43 | def save_cpk(self, emergent=False): 44 | cpk = {k: v.state_dict() for k, v in self.models.items()} 45 | cpk['epoch'] = self.epoch 46 | cpk_path = os.path.join(self.cpk_dir, '%s-checkpoint.pth.tar' % str(self.epoch).zfill(self.zfill_num)) 47 | if not (os.path.exists(cpk_path) and emergent): 48 | torch.save(cpk, cpk_path) 49 | 50 | @staticmethod 51 | def load_cpk(checkpoint_path, generator=None, discriminator=None, kp_detector=None, 52 | optimizer_generator=None, optimizer_discriminator=None, optimizer_kp_detector=None): 53 | checkpoint = torch.load(checkpoint_path) 54 | if generator is not None: 55 | generator.load_state_dict(checkpoint['generator']) 56 | if kp_detector is not None: 57 | kp_detector.load_state_dict(checkpoint['kp_detector']) 58 | if discriminator is not None: 59 | try: 60 | discriminator.load_state_dict(checkpoint['discriminator']) 61 | except: 62 | print ('No discriminator in the state-dict. Dicriminator will be randomly initialized') 63 | if optimizer_generator is not None: 64 | optimizer_generator.load_state_dict(checkpoint['optimizer_generator']) 65 | if optimizer_discriminator is not None: 66 | try: 67 | optimizer_discriminator.load_state_dict(checkpoint['optimizer_discriminator']) 68 | except RuntimeError as e: 69 | print ('No discriminator optimizer in the state-dict. Optimizer will be not initialized') 70 | if optimizer_kp_detector is not None: 71 | optimizer_kp_detector.load_state_dict(checkpoint['optimizer_kp_detector']) 72 | 73 | return checkpoint['epoch'] 74 | 75 | def __enter__(self): 76 | return self 77 | 78 | def __exit__(self, exc_type, exc_val, exc_tb): 79 | if 'models' in self.__dict__: 80 | self.save_cpk() 81 | self.log_file.close() 82 | 83 | def log_iter(self, losses): 84 | losses = collections.OrderedDict(losses.items()) 85 | if self.names is None: 86 | self.names = list(losses.keys()) 87 | self.loss_list.append(list(losses.values())) 88 | 89 | def log_epoch(self, epoch, models, inp, out): 90 | self.epoch = epoch 91 | self.models = models 92 | if (self.epoch + 1) % self.checkpoint_freq == 0: 93 | self.save_cpk() 94 | self.log_scores(self.names) 95 | self.visualize_rec(inp, out) 96 | 97 | 98 | class Visualizer: 99 | def __init__(self, kp_size=5, draw_border=False, colormap='gist_rainbow'): 100 | self.kp_size = kp_size 101 | self.draw_border = draw_border 102 | self.colormap = plt.get_cmap(colormap) 103 | 104 | def draw_image_with_kp(self, image, kp_array): 105 | image = np.copy(image) 106 | spatial_size = np.array(image.shape[:2][::-1])[np.newaxis] 107 | kp_array = spatial_size * (kp_array + 1) / 2 108 | num_kp = kp_array.shape[0] 109 | for kp_ind, kp in enumerate(kp_array): 110 | rr, cc = circle(kp[1], kp[0], self.kp_size, shape=image.shape[:2]) 111 | image[rr, cc] = np.array(self.colormap(kp_ind / num_kp))[:3] 112 | return image 113 | 114 | def create_image_column_with_kp(self, images, kp): 115 | image_array = np.array([self.draw_image_with_kp(v, k) for v, k in zip(images, kp)]) 116 | return self.create_image_column(image_array) 117 | 118 | def create_image_column(self, images): 119 | if self.draw_border: 120 | images = np.copy(images) 121 | images[:, :, [0, -1]] = (1, 1, 1) 122 | images[:, :, [0, -1]] = (1, 1, 1) 123 | return np.concatenate(list(images), axis=0) 124 | 125 | def create_image_grid(self, *args): 126 | out = [] 127 | for arg in args: 128 | if type(arg) == tuple: 129 | out.append(self.create_image_column_with_kp(arg[0], arg[1])) 130 | else: 131 | out.append(self.create_image_column(arg)) 132 | return np.concatenate(out, axis=1) 133 | 134 | def visualize(self, driving, source, out): 135 | images = [] 136 | 137 | # Source image with keypoints 138 | source = source.data.cpu() 139 | kp_source = out['kp_source']['value'].data.cpu().numpy() 140 | source = np.transpose(source, [0, 2, 3, 1]) 141 | images.append((source, kp_source)) 142 | 143 | # Equivariance visualization 144 | if 'transformed_frame' in out: 145 | transformed = out['transformed_frame'].data.cpu().numpy() 146 | transformed = np.transpose(transformed, [0, 2, 3, 1]) 147 | transformed_kp = out['transformed_kp']['value'].data.cpu().numpy() 148 | images.append((transformed, transformed_kp)) 149 | 150 | # Driving image with keypoints 151 | kp_driving = out['kp_driving']['value'].data.cpu().numpy() 152 | driving = driving.data.cpu().numpy() 153 | driving = np.transpose(driving, [0, 2, 3, 1]) 154 | images.append((driving, kp_driving)) 155 | 156 | # Deformed image 157 | if 'deformed' in out: 158 | deformed = out['deformed'].data.cpu().numpy() 159 | deformed = np.transpose(deformed, [0, 2, 3, 1]) 160 | images.append(deformed) 161 | 162 | # Result with and without keypoints 163 | prediction = out['prediction'].data.cpu().numpy() 164 | prediction = np.transpose(prediction, [0, 2, 3, 1]) 165 | if 'kp_norm' in out: 166 | kp_norm = out['kp_norm']['value'].data.cpu().numpy() 167 | images.append((prediction, kp_norm)) 168 | images.append(prediction) 169 | 170 | 171 | ## Occlusion map 172 | if 'occlusion_map' in out: 173 | occlusion_map = out['occlusion_map'].data.cpu().repeat(1, 3, 1, 1) 174 | occlusion_map = F.interpolate(occlusion_map, size=source.shape[1:3]).numpy() 175 | occlusion_map = np.transpose(occlusion_map, [0, 2, 3, 1]) 176 | images.append(occlusion_map) 177 | 178 | # Deformed images according to each individual transform 179 | if 'sparse_deformed' in out: 180 | full_mask = [] 181 | for i in range(out['sparse_deformed'].shape[1]): 182 | image = out['sparse_deformed'][:, i].data.cpu() 183 | image = F.interpolate(image, size=source.shape[1:3]) 184 | mask = out['mask'][:, i:(i+1)].data.cpu().repeat(1, 3, 1, 1) 185 | mask = F.interpolate(mask, size=source.shape[1:3]) 186 | image = np.transpose(image.numpy(), (0, 2, 3, 1)) 187 | mask = np.transpose(mask.numpy(), (0, 2, 3, 1)) 188 | 189 | if i != 0: 190 | color = np.array(self.colormap((i - 1) / (out['sparse_deformed'].shape[1] - 1)))[:3] 191 | else: 192 | color = np.array((0, 0, 0)) 193 | 194 | color = color.reshape((1, 1, 1, 3)) 195 | 196 | images.append(image) 197 | if i != 0: 198 | images.append(mask * color) 199 | else: 200 | images.append(mask) 201 | 202 | full_mask.append(mask * color) 203 | 204 | images.append(sum(full_mask)) 205 | 206 | image = self.create_image_grid(*images) 207 | image = (255 * image).astype(np.uint8) 208 | return image 209 | -------------------------------------------------------------------------------- /modules/dense_motion.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | import torch.nn.functional as F 3 | import torch 4 | from modules.util import Hourglass, AntiAliasInterpolation2d, make_coordinate_grid, kp2gaussian 5 | 6 | 7 | class DenseMotionNetwork(nn.Module): 8 | """ 9 | Module that predicting a dense motion from sparse motion representation given by kp_source and kp_driving 10 | """ 11 | 12 | def __init__(self, block_expansion, num_blocks, max_features, num_kp, num_channels, estimate_occlusion_map=False, 13 | scale_factor=1, kp_variance=0.01): 14 | super(DenseMotionNetwork, self).__init__() 15 | self.hourglass = Hourglass(block_expansion=block_expansion, in_features=(num_kp + 1) * (num_channels + 1), 16 | max_features=max_features, num_blocks=num_blocks) 17 | 18 | self.mask = nn.Conv2d(self.hourglass.out_filters, num_kp + 1, kernel_size=(7, 7), padding=(3, 3)) 19 | 20 | if estimate_occlusion_map: 21 | self.occlusion = nn.Conv2d(self.hourglass.out_filters, 1, kernel_size=(7, 7), padding=(3, 3)) 22 | else: 23 | self.occlusion = None 24 | 25 | self.num_kp = num_kp 26 | self.scale_factor = scale_factor 27 | self.kp_variance = kp_variance 28 | 29 | if self.scale_factor != 1: 30 | self.down = AntiAliasInterpolation2d(num_channels, self.scale_factor) 31 | 32 | def create_heatmap_representations(self, source_image, kp_driving, kp_source): 33 | """ 34 | Eq 6. in the paper H_k(z) 35 | """ 36 | spatial_size = source_image.shape[2:] 37 | gaussian_driving = kp2gaussian(kp_driving, spatial_size=spatial_size, kp_variance=self.kp_variance) 38 | gaussian_source = kp2gaussian(kp_source, spatial_size=spatial_size, kp_variance=self.kp_variance) 39 | heatmap = gaussian_driving - gaussian_source 40 | 41 | #adding background feature 42 | zeros = torch.zeros(heatmap.shape[0], 1, spatial_size[0], spatial_size[1]).type(heatmap.type()) 43 | heatmap = torch.cat([zeros, heatmap], dim=1) 44 | heatmap = heatmap.unsqueeze(2) 45 | return heatmap 46 | 47 | def create_sparse_motions(self, source_image, kp_driving, kp_source): 48 | """ 49 | Eq 4. in the paper T_{s<-d}(z) 50 | """ 51 | bs, _, h, w = source_image.shape 52 | identity_grid = make_coordinate_grid((h, w), type=kp_source['value'].type()) 53 | identity_grid = identity_grid.view(1, 1, h, w, 2) 54 | coordinate_grid = identity_grid - kp_driving['value'].view(bs, self.num_kp, 1, 1, 2) 55 | if 'jacobian' in kp_driving: 56 | jacobian = torch.matmul(kp_source['jacobian'], torch.inverse(kp_driving['jacobian'])) 57 | jacobian = jacobian.unsqueeze(-3).unsqueeze(-3) 58 | jacobian = jacobian.repeat(1, 1, h, w, 1, 1) 59 | coordinate_grid = torch.matmul(jacobian, coordinate_grid.unsqueeze(-1)) 60 | coordinate_grid = coordinate_grid.squeeze(-1) 61 | 62 | driving_to_source = coordinate_grid + kp_source['value'].view(bs, self.num_kp, 1, 1, 2) 63 | 64 | #adding background feature 65 | identity_grid = identity_grid.repeat(bs, 1, 1, 1, 1) 66 | sparse_motions = torch.cat([identity_grid, driving_to_source], dim=1) 67 | return sparse_motions 68 | 69 | def create_deformed_source_image(self, source_image, sparse_motions): 70 | """ 71 | Eq 7. in the paper \hat{T}_{s<-d}(z) 72 | """ 73 | bs, _, h, w = source_image.shape 74 | source_repeat = source_image.unsqueeze(1).unsqueeze(1).repeat(1, self.num_kp + 1, 1, 1, 1, 1) 75 | source_repeat = source_repeat.view(bs * (self.num_kp + 1), -1, h, w) 76 | sparse_motions = sparse_motions.view((bs * (self.num_kp + 1), h, w, -1)) 77 | sparse_deformed = F.grid_sample(source_repeat, sparse_motions) 78 | sparse_deformed = sparse_deformed.view((bs, self.num_kp + 1, -1, h, w)) 79 | return sparse_deformed 80 | 81 | def forward(self, source_image, kp_driving, kp_source): 82 | if self.scale_factor != 1: 83 | source_image = self.down(source_image) 84 | 85 | bs, _, h, w = source_image.shape 86 | 87 | out_dict = dict() 88 | heatmap_representation = self.create_heatmap_representations(source_image, kp_driving, kp_source) 89 | sparse_motion = self.create_sparse_motions(source_image, kp_driving, kp_source) 90 | deformed_source = self.create_deformed_source_image(source_image, sparse_motion) 91 | out_dict['sparse_deformed'] = deformed_source 92 | 93 | input = torch.cat([heatmap_representation, deformed_source], dim=2) 94 | input = input.view(bs, -1, h, w) 95 | 96 | prediction = self.hourglass(input) 97 | 98 | mask = self.mask(prediction) 99 | mask = F.softmax(mask, dim=1) 100 | out_dict['mask'] = mask 101 | mask = mask.unsqueeze(2) 102 | sparse_motion = sparse_motion.permute(0, 1, 4, 2, 3) 103 | deformation = (sparse_motion * mask).sum(dim=1) 104 | deformation = deformation.permute(0, 2, 3, 1) 105 | 106 | out_dict['deformation'] = deformation 107 | 108 | # Sec. 3.2 in the paper 109 | if self.occlusion: 110 | occlusion_map = torch.sigmoid(self.occlusion(prediction)) 111 | out_dict['occlusion_map'] = occlusion_map 112 | 113 | return out_dict 114 | -------------------------------------------------------------------------------- /modules/discriminator.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | import torch.nn.functional as F 3 | from modules.util import kp2gaussian 4 | import torch 5 | 6 | 7 | class DownBlock2d(nn.Module): 8 | """ 9 | Simple block for processing video (encoder). 10 | """ 11 | 12 | def __init__(self, in_features, out_features, norm=False, kernel_size=4, pool=False, sn=False): 13 | super(DownBlock2d, self).__init__() 14 | self.conv = nn.Conv2d(in_channels=in_features, out_channels=out_features, kernel_size=kernel_size) 15 | 16 | if sn: 17 | self.conv = nn.utils.spectral_norm(self.conv) 18 | 19 | if norm: 20 | self.norm = nn.InstanceNorm2d(out_features, affine=True) 21 | else: 22 | self.norm = None 23 | self.pool = pool 24 | 25 | def forward(self, x): 26 | out = x 27 | out = self.conv(out) 28 | if self.norm: 29 | out = self.norm(out) 30 | out = F.leaky_relu(out, 0.2) 31 | if self.pool: 32 | out = F.avg_pool2d(out, (2, 2)) 33 | return out 34 | 35 | 36 | class Discriminator(nn.Module): 37 | """ 38 | Discriminator similar to Pix2Pix 39 | """ 40 | 41 | def __init__(self, num_channels=3, block_expansion=64, num_blocks=4, max_features=512, 42 | sn=False, use_kp=False, num_kp=10, kp_variance=0.01, **kwargs): 43 | super(Discriminator, self).__init__() 44 | 45 | down_blocks = [] 46 | for i in range(num_blocks): 47 | down_blocks.append( 48 | DownBlock2d(num_channels + num_kp * use_kp if i == 0 else min(max_features, block_expansion * (2 ** i)), 49 | min(max_features, block_expansion * (2 ** (i + 1))), 50 | norm=(i != 0), kernel_size=4, pool=(i != num_blocks - 1), sn=sn)) 51 | 52 | self.down_blocks = nn.ModuleList(down_blocks) 53 | self.conv = nn.Conv2d(self.down_blocks[-1].conv.out_channels, out_channels=1, kernel_size=1) 54 | if sn: 55 | self.conv = nn.utils.spectral_norm(self.conv) 56 | self.use_kp = use_kp 57 | self.kp_variance = kp_variance 58 | 59 | def forward(self, x, kp=None): 60 | feature_maps = [] 61 | out = x 62 | if self.use_kp: 63 | heatmap = kp2gaussian(kp, x.shape[2:], self.kp_variance) 64 | out = torch.cat([out, heatmap], dim=1) 65 | 66 | for down_block in self.down_blocks: 67 | feature_maps.append(down_block(out)) 68 | out = feature_maps[-1] 69 | prediction_map = self.conv(out) 70 | 71 | return feature_maps, prediction_map 72 | 73 | 74 | class MultiScaleDiscriminator(nn.Module): 75 | """ 76 | Multi-scale (scale) discriminator 77 | """ 78 | 79 | def __init__(self, scales=(), **kwargs): 80 | super(MultiScaleDiscriminator, self).__init__() 81 | self.scales = scales 82 | discs = {} 83 | for scale in scales: 84 | discs[str(scale).replace('.', '-')] = Discriminator(**kwargs) 85 | self.discs = nn.ModuleDict(discs) 86 | 87 | def forward(self, x, kp=None): 88 | out_dict = {} 89 | for scale, disc in self.discs.items(): 90 | scale = str(scale).replace('-', '.') 91 | key = 'prediction_' + scale 92 | feature_maps, prediction_map = disc(x[key], kp) 93 | out_dict['feature_maps_' + scale] = feature_maps 94 | out_dict['prediction_map_' + scale] = prediction_map 95 | return out_dict 96 | -------------------------------------------------------------------------------- /modules/generator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | import torch.nn.functional as F 4 | from modules.util import ResBlock2d, SameBlock2d, UpBlock2d, DownBlock2d 5 | from modules.dense_motion import DenseMotionNetwork 6 | 7 | 8 | class OcclusionAwareGenerator(nn.Module): 9 | """ 10 | Generator that given source image and and keypoints try to transform image according to movement trajectories 11 | induced by keypoints. Generator follows Johnson architecture. 12 | """ 13 | 14 | def __init__(self, num_channels, num_kp, block_expansion, max_features, num_down_blocks, 15 | num_bottleneck_blocks, estimate_occlusion_map=False, dense_motion_params=None, estimate_jacobian=False): 16 | super(OcclusionAwareGenerator, self).__init__() 17 | 18 | if dense_motion_params is not None: 19 | self.dense_motion_network = DenseMotionNetwork(num_kp=num_kp, num_channels=num_channels, 20 | estimate_occlusion_map=estimate_occlusion_map, 21 | **dense_motion_params) 22 | else: 23 | self.dense_motion_network = None 24 | 25 | self.first = SameBlock2d(num_channels, block_expansion, kernel_size=(7, 7), padding=(3, 3)) 26 | 27 | down_blocks = [] 28 | for i in range(num_down_blocks): 29 | in_features = min(max_features, block_expansion * (2 ** i)) 30 | out_features = min(max_features, block_expansion * (2 ** (i + 1))) 31 | down_blocks.append(DownBlock2d(in_features, out_features, kernel_size=(3, 3), padding=(1, 1))) 32 | self.down_blocks = nn.ModuleList(down_blocks) 33 | 34 | up_blocks = [] 35 | for i in range(num_down_blocks): 36 | in_features = min(max_features, block_expansion * (2 ** (num_down_blocks - i))) 37 | out_features = min(max_features, block_expansion * (2 ** (num_down_blocks - i - 1))) 38 | up_blocks.append(UpBlock2d(in_features, out_features, kernel_size=(3, 3), padding=(1, 1))) 39 | self.up_blocks = nn.ModuleList(up_blocks) 40 | 41 | self.bottleneck = torch.nn.Sequential() 42 | in_features = min(max_features, block_expansion * (2 ** num_down_blocks)) 43 | for i in range(num_bottleneck_blocks): 44 | self.bottleneck.add_module('r' + str(i), ResBlock2d(in_features, kernel_size=(3, 3), padding=(1, 1))) 45 | 46 | self.final = nn.Conv2d(block_expansion, num_channels, kernel_size=(7, 7), padding=(3, 3)) 47 | self.estimate_occlusion_map = estimate_occlusion_map 48 | self.num_channels = num_channels 49 | 50 | def deform_input(self, inp, deformation): 51 | _, h_old, w_old, _ = deformation.shape 52 | _, _, h, w = inp.shape 53 | if h_old != h or w_old != w: 54 | deformation = deformation.permute(0, 3, 1, 2) 55 | deformation = F.interpolate(deformation, size=(h, w), mode='bilinear') 56 | deformation = deformation.permute(0, 2, 3, 1) 57 | return F.grid_sample(inp, deformation) 58 | 59 | def forward(self, source_image, kp_driving, kp_source): 60 | # Encoding (downsampling) part 61 | out = self.first(source_image) 62 | for i in range(len(self.down_blocks)): 63 | out = self.down_blocks[i](out) 64 | 65 | # Transforming feature representation according to deformation and occlusion 66 | output_dict = {} 67 | if self.dense_motion_network is not None: 68 | dense_motion = self.dense_motion_network(source_image=source_image, kp_driving=kp_driving, 69 | kp_source=kp_source) 70 | output_dict['mask'] = dense_motion['mask'] 71 | output_dict['sparse_deformed'] = dense_motion['sparse_deformed'] 72 | 73 | if 'occlusion_map' in dense_motion: 74 | occlusion_map = dense_motion['occlusion_map'] 75 | output_dict['occlusion_map'] = occlusion_map 76 | else: 77 | occlusion_map = None 78 | deformation = dense_motion['deformation'] 79 | out = self.deform_input(out, deformation) 80 | 81 | if occlusion_map is not None: 82 | if out.shape[2] != occlusion_map.shape[2] or out.shape[3] != occlusion_map.shape[3]: 83 | occlusion_map = F.interpolate(occlusion_map, size=out.shape[2:], mode='bilinear') 84 | out = out * occlusion_map 85 | 86 | output_dict["deformed"] = self.deform_input(source_image, deformation) 87 | 88 | # Decoding part 89 | out = self.bottleneck(out) 90 | for i in range(len(self.up_blocks)): 91 | out = self.up_blocks[i](out) 92 | out = self.final(out) 93 | out = F.sigmoid(out) 94 | 95 | output_dict["prediction"] = out 96 | 97 | return output_dict 98 | -------------------------------------------------------------------------------- /modules/keypoint_detector.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | import torch 3 | import torch.nn.functional as F 4 | from modules.util import Hourglass, make_coordinate_grid, AntiAliasInterpolation2d 5 | 6 | 7 | class KPDetector(nn.Module): 8 | """ 9 | Detecting a keypoints. Return keypoint position and jacobian near each keypoint. 10 | """ 11 | 12 | def __init__(self, block_expansion, num_kp, num_channels, max_features, 13 | num_blocks, temperature, estimate_jacobian=False, scale_factor=1, 14 | single_jacobian_map=False, pad=0): 15 | super(KPDetector, self).__init__() 16 | 17 | self.predictor = Hourglass(block_expansion, in_features=num_channels, 18 | max_features=max_features, num_blocks=num_blocks) 19 | 20 | self.kp = nn.Conv2d(in_channels=self.predictor.out_filters, out_channels=num_kp, kernel_size=(7, 7), 21 | padding=pad) 22 | 23 | if estimate_jacobian: 24 | self.num_jacobian_maps = 1 if single_jacobian_map else num_kp 25 | self.jacobian = nn.Conv2d(in_channels=self.predictor.out_filters, 26 | out_channels=4 * self.num_jacobian_maps, kernel_size=(7, 7), padding=pad) 27 | self.jacobian.weight.data.zero_() 28 | self.jacobian.bias.data.copy_(torch.tensor([1, 0, 0, 1] * self.num_jacobian_maps, dtype=torch.float)) 29 | else: 30 | self.jacobian = None 31 | 32 | self.temperature = temperature 33 | self.scale_factor = scale_factor 34 | if self.scale_factor != 1: 35 | self.down = AntiAliasInterpolation2d(num_channels, self.scale_factor) 36 | 37 | def gaussian2kp(self, heatmap): 38 | """ 39 | Extract the mean and from a heatmap 40 | """ 41 | shape = heatmap.shape 42 | heatmap = heatmap.unsqueeze(-1) 43 | grid = make_coordinate_grid(shape[2:], heatmap.type()).unsqueeze_(0).unsqueeze_(0) 44 | value = (heatmap * grid).sum(dim=(2, 3)) 45 | kp = {'value': value} 46 | 47 | return kp 48 | 49 | def forward(self, x): 50 | if self.scale_factor != 1: 51 | x = self.down(x) 52 | 53 | feature_map = self.predictor(x) 54 | prediction = self.kp(feature_map) 55 | 56 | final_shape = prediction.shape 57 | heatmap = prediction.view(final_shape[0], final_shape[1], -1) 58 | heatmap = F.softmax(heatmap / self.temperature, dim=2) 59 | heatmap = heatmap.view(*final_shape) 60 | 61 | out = self.gaussian2kp(heatmap) 62 | 63 | if self.jacobian is not None: 64 | jacobian_map = self.jacobian(feature_map) 65 | jacobian_map = jacobian_map.reshape(final_shape[0], self.num_jacobian_maps, 4, final_shape[2], 66 | final_shape[3]) 67 | heatmap = heatmap.unsqueeze(2) 68 | 69 | jacobian = heatmap * jacobian_map 70 | jacobian = jacobian.view(final_shape[0], final_shape[1], 4, -1) 71 | jacobian = jacobian.sum(dim=-1) 72 | jacobian = jacobian.view(jacobian.shape[0], jacobian.shape[1], 2, 2) 73 | out['jacobian'] = jacobian 74 | 75 | return out 76 | -------------------------------------------------------------------------------- /modules/model.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | import torch 3 | import torch.nn.functional as F 4 | from modules.util import AntiAliasInterpolation2d, make_coordinate_grid 5 | from torchvision import models 6 | import numpy as np 7 | from torch.autograd import grad 8 | 9 | 10 | class Vgg19(torch.nn.Module): 11 | """ 12 | Vgg19 network for perceptual loss. See Sec 3.3. 13 | """ 14 | def __init__(self, requires_grad=False): 15 | super(Vgg19, self).__init__() 16 | vgg_pretrained_features = models.vgg19(pretrained=True).features 17 | self.slice1 = torch.nn.Sequential() 18 | self.slice2 = torch.nn.Sequential() 19 | self.slice3 = torch.nn.Sequential() 20 | self.slice4 = torch.nn.Sequential() 21 | self.slice5 = torch.nn.Sequential() 22 | for x in range(2): 23 | self.slice1.add_module(str(x), vgg_pretrained_features[x]) 24 | for x in range(2, 7): 25 | self.slice2.add_module(str(x), vgg_pretrained_features[x]) 26 | for x in range(7, 12): 27 | self.slice3.add_module(str(x), vgg_pretrained_features[x]) 28 | for x in range(12, 21): 29 | self.slice4.add_module(str(x), vgg_pretrained_features[x]) 30 | for x in range(21, 30): 31 | self.slice5.add_module(str(x), vgg_pretrained_features[x]) 32 | 33 | self.mean = torch.nn.Parameter(data=torch.Tensor(np.array([0.485, 0.456, 0.406]).reshape((1, 3, 1, 1))), 34 | requires_grad=False) 35 | self.std = torch.nn.Parameter(data=torch.Tensor(np.array([0.229, 0.224, 0.225]).reshape((1, 3, 1, 1))), 36 | requires_grad=False) 37 | 38 | if not requires_grad: 39 | for param in self.parameters(): 40 | param.requires_grad = False 41 | 42 | def forward(self, X): 43 | X = (X - self.mean) / self.std 44 | h_relu1 = self.slice1(X) 45 | h_relu2 = self.slice2(h_relu1) 46 | h_relu3 = self.slice3(h_relu2) 47 | h_relu4 = self.slice4(h_relu3) 48 | h_relu5 = self.slice5(h_relu4) 49 | out = [h_relu1, h_relu2, h_relu3, h_relu4, h_relu5] 50 | return out 51 | 52 | 53 | class ImagePyramide(torch.nn.Module): 54 | """ 55 | Create image pyramide for computing pyramide perceptual loss. See Sec 3.3 56 | """ 57 | def __init__(self, scales, num_channels): 58 | super(ImagePyramide, self).__init__() 59 | downs = {} 60 | for scale in scales: 61 | downs[str(scale).replace('.', '-')] = AntiAliasInterpolation2d(num_channels, scale) 62 | self.downs = nn.ModuleDict(downs) 63 | 64 | def forward(self, x): 65 | out_dict = {} 66 | for scale, down_module in self.downs.items(): 67 | out_dict['prediction_' + str(scale).replace('-', '.')] = down_module(x) 68 | return out_dict 69 | 70 | 71 | class Transform: 72 | """ 73 | Random tps transformation for equivariance constraints. See Sec 3.3 74 | """ 75 | def __init__(self, bs, **kwargs): 76 | noise = torch.normal(mean=0, std=kwargs['sigma_affine'] * torch.ones([bs, 2, 3])) 77 | self.theta = noise + torch.eye(2, 3).view(1, 2, 3) 78 | self.bs = bs 79 | 80 | if ('sigma_tps' in kwargs) and ('points_tps' in kwargs): 81 | self.tps = True 82 | self.control_points = make_coordinate_grid((kwargs['points_tps'], kwargs['points_tps']), type=noise.type()) 83 | self.control_points = self.control_points.unsqueeze(0) 84 | self.control_params = torch.normal(mean=0, 85 | std=kwargs['sigma_tps'] * torch.ones([bs, 1, kwargs['points_tps'] ** 2])) 86 | else: 87 | self.tps = False 88 | 89 | def transform_frame(self, frame): 90 | grid = make_coordinate_grid(frame.shape[2:], type=frame.type()).unsqueeze(0) 91 | grid = grid.view(1, frame.shape[2] * frame.shape[3], 2) 92 | grid = self.warp_coordinates(grid).view(self.bs, frame.shape[2], frame.shape[3], 2) 93 | return F.grid_sample(frame, grid, padding_mode="reflection") 94 | 95 | def warp_coordinates(self, coordinates): 96 | theta = self.theta.type(coordinates.type()) 97 | theta = theta.unsqueeze(1) 98 | transformed = torch.matmul(theta[:, :, :, :2], coordinates.unsqueeze(-1)) + theta[:, :, :, 2:] 99 | transformed = transformed.squeeze(-1) 100 | 101 | if self.tps: 102 | control_points = self.control_points.type(coordinates.type()) 103 | control_params = self.control_params.type(coordinates.type()) 104 | distances = coordinates.view(coordinates.shape[0], -1, 1, 2) - control_points.view(1, 1, -1, 2) 105 | distances = torch.abs(distances).sum(-1) 106 | 107 | result = distances ** 2 108 | result = result * torch.log(distances + 1e-6) 109 | result = result * control_params 110 | result = result.sum(dim=2).view(self.bs, coordinates.shape[1], 1) 111 | transformed = transformed + result 112 | 113 | return transformed 114 | 115 | def jacobian(self, coordinates): 116 | new_coordinates = self.warp_coordinates(coordinates) 117 | grad_x = grad(new_coordinates[..., 0].sum(), coordinates, create_graph=True) 118 | grad_y = grad(new_coordinates[..., 1].sum(), coordinates, create_graph=True) 119 | jacobian = torch.cat([grad_x[0].unsqueeze(-2), grad_y[0].unsqueeze(-2)], dim=-2) 120 | return jacobian 121 | 122 | 123 | def detach_kp(kp): 124 | return {key: value.detach() for key, value in kp.items()} 125 | 126 | 127 | class GeneratorFullModel(torch.nn.Module): 128 | """ 129 | Merge all generator related updates into single model for better multi-gpu usage 130 | """ 131 | 132 | def __init__(self, kp_extractor, generator, discriminator, train_params): 133 | super(GeneratorFullModel, self).__init__() 134 | self.kp_extractor = kp_extractor 135 | self.generator = generator 136 | self.discriminator = discriminator 137 | self.train_params = train_params 138 | self.scales = train_params['scales'] 139 | self.disc_scales = self.discriminator.scales 140 | self.pyramid = ImagePyramide(self.scales, generator.num_channels) 141 | if torch.cuda.is_available(): 142 | self.pyramid = self.pyramid.cuda() 143 | 144 | self.loss_weights = train_params['loss_weights'] 145 | 146 | if sum(self.loss_weights['perceptual']) != 0: 147 | self.vgg = Vgg19() 148 | if torch.cuda.is_available(): 149 | self.vgg = self.vgg.cuda() 150 | 151 | def forward(self, x): 152 | kp_source = self.kp_extractor(x['source']) 153 | kp_driving = self.kp_extractor(x['driving']) 154 | 155 | generated = self.generator(x['source'], kp_source=kp_source, kp_driving=kp_driving) 156 | generated.update({'kp_source': kp_source, 'kp_driving': kp_driving}) 157 | 158 | loss_values = {} 159 | 160 | pyramide_real = self.pyramid(x['driving']) 161 | pyramide_generated = self.pyramid(generated['prediction']) 162 | 163 | if sum(self.loss_weights['perceptual']) != 0: 164 | value_total = 0 165 | for scale in self.scales: 166 | x_vgg = self.vgg(pyramide_generated['prediction_' + str(scale)]) 167 | y_vgg = self.vgg(pyramide_real['prediction_' + str(scale)]) 168 | 169 | for i, weight in enumerate(self.loss_weights['perceptual']): 170 | value = torch.abs(x_vgg[i] - y_vgg[i].detach()).mean() 171 | value_total += self.loss_weights['perceptual'][i] * value 172 | loss_values['perceptual'] = value_total 173 | 174 | if self.loss_weights['generator_gan'] != 0: 175 | discriminator_maps_generated = self.discriminator(pyramide_generated, kp=detach_kp(kp_driving)) 176 | discriminator_maps_real = self.discriminator(pyramide_real, kp=detach_kp(kp_driving)) 177 | value_total = 0 178 | for scale in self.disc_scales: 179 | key = 'prediction_map_%s' % scale 180 | value = ((1 - discriminator_maps_generated[key]) ** 2).mean() 181 | value_total += self.loss_weights['generator_gan'] * value 182 | loss_values['gen_gan'] = value_total 183 | 184 | if sum(self.loss_weights['feature_matching']) != 0: 185 | value_total = 0 186 | for scale in self.disc_scales: 187 | key = 'feature_maps_%s' % scale 188 | for i, (a, b) in enumerate(zip(discriminator_maps_real[key], discriminator_maps_generated[key])): 189 | if self.loss_weights['feature_matching'][i] == 0: 190 | continue 191 | value = torch.abs(a - b).mean() 192 | value_total += self.loss_weights['feature_matching'][i] * value 193 | loss_values['feature_matching'] = value_total 194 | 195 | if (self.loss_weights['equivariance_value'] + self.loss_weights['equivariance_jacobian']) != 0: 196 | transform = Transform(x['driving'].shape[0], **self.train_params['transform_params']) 197 | transformed_frame = transform.transform_frame(x['driving']) 198 | transformed_kp = self.kp_extractor(transformed_frame) 199 | 200 | generated['transformed_frame'] = transformed_frame 201 | generated['transformed_kp'] = transformed_kp 202 | 203 | ## Value loss part 204 | if self.loss_weights['equivariance_value'] != 0: 205 | value = torch.abs(kp_driving['value'] - transform.warp_coordinates(transformed_kp['value'])).mean() 206 | loss_values['equivariance_value'] = self.loss_weights['equivariance_value'] * value 207 | 208 | ## jacobian loss part 209 | if self.loss_weights['equivariance_jacobian'] != 0: 210 | jacobian_transformed = torch.matmul(transform.jacobian(transformed_kp['value']), 211 | transformed_kp['jacobian']) 212 | 213 | normed_driving = torch.inverse(kp_driving['jacobian']) 214 | normed_transformed = jacobian_transformed 215 | value = torch.matmul(normed_driving, normed_transformed) 216 | 217 | eye = torch.eye(2).view(1, 1, 2, 2).type(value.type()) 218 | 219 | value = torch.abs(eye - value).mean() 220 | loss_values['equivariance_jacobian'] = self.loss_weights['equivariance_jacobian'] * value 221 | 222 | return loss_values, generated 223 | 224 | 225 | class DiscriminatorFullModel(torch.nn.Module): 226 | """ 227 | Merge all discriminator related updates into single model for better multi-gpu usage 228 | """ 229 | 230 | def __init__(self, kp_extractor, generator, discriminator, train_params): 231 | super(DiscriminatorFullModel, self).__init__() 232 | self.kp_extractor = kp_extractor 233 | self.generator = generator 234 | self.discriminator = discriminator 235 | self.train_params = train_params 236 | self.scales = self.discriminator.scales 237 | self.pyramid = ImagePyramide(self.scales, generator.num_channels) 238 | if torch.cuda.is_available(): 239 | self.pyramid = self.pyramid.cuda() 240 | 241 | self.loss_weights = train_params['loss_weights'] 242 | 243 | def forward(self, x, generated): 244 | pyramide_real = self.pyramid(x['driving']) 245 | pyramide_generated = self.pyramid(generated['prediction'].detach()) 246 | 247 | kp_driving = generated['kp_driving'] 248 | discriminator_maps_generated = self.discriminator(pyramide_generated, kp=detach_kp(kp_driving)) 249 | discriminator_maps_real = self.discriminator(pyramide_real, kp=detach_kp(kp_driving)) 250 | 251 | loss_values = {} 252 | value_total = 0 253 | for scale in self.scales: 254 | key = 'prediction_map_%s' % scale 255 | value = (1 - discriminator_maps_real[key]) ** 2 + discriminator_maps_generated[key] ** 2 256 | value_total += self.loss_weights['discriminator_gan'] * value.mean() 257 | loss_values['disc_gan'] = value_total 258 | 259 | return loss_values 260 | -------------------------------------------------------------------------------- /modules/util.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | import torch.nn.functional as F 4 | import torch 5 | 6 | from sync_batchnorm import SynchronizedBatchNorm2d as BatchNorm2d 7 | 8 | 9 | def kp2gaussian(kp, spatial_size, kp_variance): 10 | """ 11 | Transform a keypoint into gaussian like representation 12 | """ 13 | mean = kp['value'] 14 | 15 | coordinate_grid = make_coordinate_grid(spatial_size, mean.type()) 16 | number_of_leading_dimensions = len(mean.shape) - 1 17 | shape = (1,) * number_of_leading_dimensions + coordinate_grid.shape 18 | coordinate_grid = coordinate_grid.view(*shape) 19 | repeats = mean.shape[:number_of_leading_dimensions] + (1, 1, 1) 20 | coordinate_grid = coordinate_grid.repeat(*repeats) 21 | 22 | # Preprocess kp shape 23 | shape = mean.shape[:number_of_leading_dimensions] + (1, 1, 2) 24 | mean = mean.view(*shape) 25 | 26 | mean_sub = (coordinate_grid - mean) 27 | 28 | out = torch.exp(-0.5 * (mean_sub ** 2).sum(-1) / kp_variance) 29 | 30 | return out 31 | 32 | 33 | def make_coordinate_grid(spatial_size, type): 34 | """ 35 | Create a meshgrid [-1,1] x [-1,1] of given spatial_size. 36 | """ 37 | h, w = spatial_size 38 | x = torch.arange(w).type(type) 39 | y = torch.arange(h).type(type) 40 | 41 | x = (2 * (x / (w - 1)) - 1) 42 | y = (2 * (y / (h - 1)) - 1) 43 | 44 | yy = y.view(-1, 1).repeat(1, w) 45 | xx = x.view(1, -1).repeat(h, 1) 46 | 47 | meshed = torch.cat([xx.unsqueeze_(2), yy.unsqueeze_(2)], 2) 48 | 49 | return meshed 50 | 51 | 52 | class ResBlock2d(nn.Module): 53 | """ 54 | Res block, preserve spatial resolution. 55 | """ 56 | 57 | def __init__(self, in_features, kernel_size, padding): 58 | super(ResBlock2d, self).__init__() 59 | self.conv1 = nn.Conv2d(in_channels=in_features, out_channels=in_features, kernel_size=kernel_size, 60 | padding=padding) 61 | self.conv2 = nn.Conv2d(in_channels=in_features, out_channels=in_features, kernel_size=kernel_size, 62 | padding=padding) 63 | self.norm1 = BatchNorm2d(in_features, affine=True) 64 | self.norm2 = BatchNorm2d(in_features, affine=True) 65 | 66 | def forward(self, x): 67 | out = self.norm1(x) 68 | out = F.relu(out) 69 | out = self.conv1(out) 70 | out = self.norm2(out) 71 | out = F.relu(out) 72 | out = self.conv2(out) 73 | out += x 74 | return out 75 | 76 | 77 | class UpBlock2d(nn.Module): 78 | """ 79 | Upsampling block for use in decoder. 80 | """ 81 | 82 | def __init__(self, in_features, out_features, kernel_size=3, padding=1, groups=1): 83 | super(UpBlock2d, self).__init__() 84 | 85 | self.conv = nn.Conv2d(in_channels=in_features, out_channels=out_features, kernel_size=kernel_size, 86 | padding=padding, groups=groups) 87 | self.norm = BatchNorm2d(out_features, affine=True) 88 | 89 | def forward(self, x): 90 | out = F.interpolate(x, scale_factor=2) 91 | out = self.conv(out) 92 | out = self.norm(out) 93 | out = F.relu(out) 94 | return out 95 | 96 | 97 | class DownBlock2d(nn.Module): 98 | """ 99 | Downsampling block for use in encoder. 100 | """ 101 | 102 | def __init__(self, in_features, out_features, kernel_size=3, padding=1, groups=1): 103 | super(DownBlock2d, self).__init__() 104 | self.conv = nn.Conv2d(in_channels=in_features, out_channels=out_features, kernel_size=kernel_size, 105 | padding=padding, groups=groups) 106 | self.norm = BatchNorm2d(out_features, affine=True) 107 | self.pool = nn.AvgPool2d(kernel_size=(2, 2)) 108 | 109 | def forward(self, x): 110 | out = self.conv(x) 111 | out = self.norm(out) 112 | out = F.relu(out) 113 | out = self.pool(out) 114 | return out 115 | 116 | 117 | class SameBlock2d(nn.Module): 118 | """ 119 | Simple block, preserve spatial resolution. 120 | """ 121 | 122 | def __init__(self, in_features, out_features, groups=1, kernel_size=3, padding=1): 123 | super(SameBlock2d, self).__init__() 124 | self.conv = nn.Conv2d(in_channels=in_features, out_channels=out_features, 125 | kernel_size=kernel_size, padding=padding, groups=groups) 126 | self.norm = BatchNorm2d(out_features, affine=True) 127 | 128 | def forward(self, x): 129 | out = self.conv(x) 130 | out = self.norm(out) 131 | out = F.relu(out) 132 | return out 133 | 134 | 135 | class Encoder(nn.Module): 136 | """ 137 | Hourglass Encoder 138 | """ 139 | 140 | def __init__(self, block_expansion, in_features, num_blocks=3, max_features=256): 141 | super(Encoder, self).__init__() 142 | 143 | down_blocks = [] 144 | for i in range(num_blocks): 145 | down_blocks.append(DownBlock2d(in_features if i == 0 else min(max_features, block_expansion * (2 ** i)), 146 | min(max_features, block_expansion * (2 ** (i + 1))), 147 | kernel_size=3, padding=1)) 148 | self.down_blocks = nn.ModuleList(down_blocks) 149 | 150 | def forward(self, x): 151 | outs = [x] 152 | for down_block in self.down_blocks: 153 | outs.append(down_block(outs[-1])) 154 | return outs 155 | 156 | 157 | class Decoder(nn.Module): 158 | """ 159 | Hourglass Decoder 160 | """ 161 | 162 | def __init__(self, block_expansion, in_features, num_blocks=3, max_features=256): 163 | super(Decoder, self).__init__() 164 | 165 | up_blocks = [] 166 | 167 | for i in range(num_blocks)[::-1]: 168 | in_filters = (1 if i == num_blocks - 1 else 2) * min(max_features, block_expansion * (2 ** (i + 1))) 169 | out_filters = min(max_features, block_expansion * (2 ** i)) 170 | up_blocks.append(UpBlock2d(in_filters, out_filters, kernel_size=3, padding=1)) 171 | 172 | self.up_blocks = nn.ModuleList(up_blocks) 173 | self.out_filters = block_expansion + in_features 174 | 175 | def forward(self, x): 176 | out = x.pop() 177 | for up_block in self.up_blocks: 178 | out = up_block(out) 179 | skip = x.pop() 180 | out = torch.cat([out, skip], dim=1) 181 | return out 182 | 183 | 184 | class Hourglass(nn.Module): 185 | """ 186 | Hourglass architecture. 187 | """ 188 | 189 | def __init__(self, block_expansion, in_features, num_blocks=3, max_features=256): 190 | super(Hourglass, self).__init__() 191 | self.encoder = Encoder(block_expansion, in_features, num_blocks, max_features) 192 | self.decoder = Decoder(block_expansion, in_features, num_blocks, max_features) 193 | self.out_filters = self.decoder.out_filters 194 | 195 | def forward(self, x): 196 | return self.decoder(self.encoder(x)) 197 | 198 | 199 | class AntiAliasInterpolation2d(nn.Module): 200 | """ 201 | Band-limited downsampling, for better preservation of the input signal. 202 | """ 203 | def __init__(self, channels, scale): 204 | super(AntiAliasInterpolation2d, self).__init__() 205 | sigma = (1 / scale - 1) / 2 206 | kernel_size = 2 * round(sigma * 4) + 1 207 | self.ka = kernel_size // 2 208 | self.kb = self.ka - 1 if kernel_size % 2 == 0 else self.ka 209 | 210 | kernel_size = [kernel_size, kernel_size] 211 | sigma = [sigma, sigma] 212 | # The gaussian kernel is the product of the 213 | # gaussian function of each dimension. 214 | kernel = 1 215 | meshgrids = torch.meshgrid( 216 | [ 217 | torch.arange(size, dtype=torch.float32) 218 | for size in kernel_size 219 | ] 220 | ) 221 | for size, std, mgrid in zip(kernel_size, sigma, meshgrids): 222 | mean = (size - 1) / 2 223 | kernel *= torch.exp(-(mgrid - mean) ** 2 / (2 * std ** 2)) 224 | 225 | # Make sure sum of values in gaussian kernel equals 1. 226 | kernel = kernel / torch.sum(kernel) 227 | # Reshape to depthwise convolutional weight 228 | kernel = kernel.view(1, 1, *kernel.size()) 229 | kernel = kernel.repeat(channels, *[1] * (kernel.dim() - 1)) 230 | 231 | self.register_buffer('weight', kernel) 232 | self.groups = channels 233 | self.scale = scale 234 | 235 | def forward(self, input): 236 | if self.scale == 1.0: 237 | return input 238 | 239 | out = F.pad(input, (self.ka, self.kb, self.ka, self.kb)) 240 | out = F.conv2d(out, weight=self.weight, groups=self.groups) 241 | out = F.interpolate(out, scale_factor=(self.scale, self.scale)) 242 | 243 | return out 244 | -------------------------------------------------------------------------------- /reconstruction.py: -------------------------------------------------------------------------------- 1 | import os 2 | from tqdm import tqdm 3 | import torch 4 | from torch.utils.data import DataLoader 5 | from logger import Logger, Visualizer 6 | import numpy as np 7 | import imageio 8 | from sync_batchnorm import DataParallelWithCallback 9 | 10 | 11 | def reconstruction(config, generator, kp_detector, checkpoint, log_dir, dataset): 12 | png_dir = os.path.join(log_dir, 'reconstruction/png') 13 | log_dir = os.path.join(log_dir, 'reconstruction') 14 | 15 | if checkpoint is not None: 16 | Logger.load_cpk(checkpoint, generator=generator, kp_detector=kp_detector) 17 | else: 18 | raise AttributeError("Checkpoint should be specified for mode='reconstruction'.") 19 | dataloader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=1) 20 | 21 | if not os.path.exists(log_dir): 22 | os.makedirs(log_dir) 23 | 24 | if not os.path.exists(png_dir): 25 | os.makedirs(png_dir) 26 | 27 | loss_list = [] 28 | if torch.cuda.is_available(): 29 | generator = DataParallelWithCallback(generator) 30 | kp_detector = DataParallelWithCallback(kp_detector) 31 | 32 | generator.eval() 33 | kp_detector.eval() 34 | 35 | for it, x in tqdm(enumerate(dataloader)): 36 | if config['reconstruction_params']['num_videos'] is not None: 37 | if it > config['reconstruction_params']['num_videos']: 38 | break 39 | with torch.no_grad(): 40 | predictions = [] 41 | visualizations = [] 42 | if torch.cuda.is_available(): 43 | x['video'] = x['video'].cuda() 44 | kp_source = kp_detector(x['video'][:, :, 0]) 45 | for frame_idx in range(x['video'].shape[2]): 46 | source = x['video'][:, :, 0] 47 | driving = x['video'][:, :, frame_idx] 48 | kp_driving = kp_detector(driving) 49 | out = generator(source, kp_source=kp_source, kp_driving=kp_driving) 50 | out['kp_source'] = kp_source 51 | out['kp_driving'] = kp_driving 52 | del out['sparse_deformed'] 53 | predictions.append(np.transpose(out['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0]) 54 | 55 | visualization = Visualizer(**config['visualizer_params']).visualize(source=source, 56 | driving=driving, out=out) 57 | visualizations.append(visualization) 58 | 59 | loss_list.append(torch.abs(out['prediction'] - driving).mean().cpu().numpy()) 60 | 61 | predictions = np.concatenate(predictions, axis=1) 62 | imageio.imsave(os.path.join(png_dir, x['name'][0] + '.png'), (255 * predictions).astype(np.uint8)) 63 | 64 | image_name = x['name'][0] + config['reconstruction_params']['format'] 65 | imageio.mimsave(os.path.join(log_dir, image_name), visualizations) 66 | 67 | print("Reconstruction loss: %s" % np.mean(loss_list)) 68 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | cffi==1.11.5 2 | cloudpickle==0.5.3 3 | cycler==0.10.0 4 | dask==2021.10.0 5 | decorator==4.3.0 6 | imageio==2.3.0 7 | kiwisolver==1.0.1 8 | matplotlib==2.2.2 9 | networkx==2.1 10 | numpy==1.22.0 11 | pandas==0.23.4 12 | Pillow>=6.2.2 13 | pycparser==2.18 14 | pygit==0.1 15 | pyparsing==2.2.0 16 | python-dateutil==2.7.3 17 | pytz==2018.5 18 | PyWavelets==0.5.2 19 | PyYAML>=5.4 20 | scikit-image==0.14.0 21 | scikit-learn==1.5.0 22 | scipy==1.1.0 23 | six==1.11.0 24 | toolz==0.9.0 25 | torch==2.2.0 26 | torchvision==0.2.1 27 | tqdm==4.24.0 28 | -------------------------------------------------------------------------------- /run.py: -------------------------------------------------------------------------------- 1 | import matplotlib 2 | 3 | matplotlib.use('Agg') 4 | 5 | import os 6 | import yaml 7 | from argparse import ArgumentParser 8 | from time import gmtime, strftime 9 | from shutil import copy 10 | 11 | from frames_dataset import FramesDataset 12 | 13 | from modules.generator import OcclusionAwareGenerator 14 | from modules.discriminator import MultiScaleDiscriminator 15 | from modules.keypoint_detector import KPDetector 16 | 17 | import torch 18 | 19 | from train import train 20 | from reconstruction import reconstruction 21 | from animate import animate 22 | 23 | if __name__ == "__main__": 24 | parser = ArgumentParser() 25 | parser.add_argument("--config", required=True, help="path to config") 26 | parser.add_argument("--mode", default="train", choices=["train", "reconstruction", "animate"]) 27 | parser.add_argument("--log_dir", default='log', help="path to log into") 28 | parser.add_argument("--checkpoint", default=None, help="path to checkpoint to restore") 29 | parser.add_argument("--device_ids", default="0", type=lambda x: list(map(int, x.split(','))), 30 | help="Names of the devices comma separated.") 31 | parser.add_argument("--verbose", dest="verbose", action="store_true", help="Print model architecture") 32 | parser.set_defaults(verbose=False) 33 | 34 | opt = parser.parse_args() 35 | with open(opt.config) as f: 36 | config = yaml.load(f, Loader=yaml.FullLoader) 37 | 38 | if opt.checkpoint is not None: 39 | log_dir = os.path.join(*os.path.split(opt.checkpoint)[:-1]) 40 | else: 41 | log_dir = os.path.join(opt.log_dir, os.path.basename(opt.config).split('.')[0]) 42 | log_dir += ' ' + strftime("%d_%m_%y_%H.%M.%S", gmtime()) 43 | 44 | generator = OcclusionAwareGenerator(**config['model_params']['generator_params'], 45 | **config['model_params']['common_params']) 46 | 47 | if torch.cuda.is_available(): 48 | generator.to(opt.device_ids[0]) 49 | if opt.verbose: 50 | print(generator) 51 | 52 | discriminator = MultiScaleDiscriminator(**config['model_params']['discriminator_params'], 53 | **config['model_params']['common_params']) 54 | if torch.cuda.is_available(): 55 | discriminator.to(opt.device_ids[0]) 56 | if opt.verbose: 57 | print(discriminator) 58 | 59 | kp_detector = KPDetector(**config['model_params']['kp_detector_params'], 60 | **config['model_params']['common_params']) 61 | 62 | if torch.cuda.is_available(): 63 | kp_detector.to(opt.device_ids[0]) 64 | 65 | if opt.verbose: 66 | print(kp_detector) 67 | 68 | dataset = FramesDataset(is_train=(opt.mode == 'train'), **config['dataset_params']) 69 | 70 | if not os.path.exists(log_dir): 71 | os.makedirs(log_dir) 72 | if not os.path.exists(os.path.join(log_dir, os.path.basename(opt.config))): 73 | copy(opt.config, log_dir) 74 | 75 | if opt.mode == 'train': 76 | print("Training...") 77 | train(config, generator, discriminator, kp_detector, opt.checkpoint, log_dir, dataset, opt.device_ids) 78 | elif opt.mode == 'reconstruction': 79 | print("Reconstruction...") 80 | reconstruction(config, generator, kp_detector, opt.checkpoint, log_dir, dataset) 81 | elif opt.mode == 'animate': 82 | print("Animate...") 83 | animate(config, generator, kp_detector, opt.checkpoint, log_dir, dataset) 84 | -------------------------------------------------------------------------------- /sup-mat/absolute-demo.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/snehitvaddi/Deep-Fake_First_Order_Model/da342d15fe15e4ef3ca9c9e4cbed0bbb87e8e58a/sup-mat/absolute-demo.gif -------------------------------------------------------------------------------- /sup-mat/download.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/snehitvaddi/Deep-Fake_First_Order_Model/da342d15fe15e4ef3ca9c9e4cbed0bbb87e8e58a/sup-mat/download.gif -------------------------------------------------------------------------------- /sup-mat/fashion-teaser.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/snehitvaddi/Deep-Fake_First_Order_Model/da342d15fe15e4ef3ca9c9e4cbed0bbb87e8e58a/sup-mat/fashion-teaser.gif -------------------------------------------------------------------------------- /sup-mat/mgif-teaser.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/snehitvaddi/Deep-Fake_First_Order_Model/da342d15fe15e4ef3ca9c9e4cbed0bbb87e8e58a/sup-mat/mgif-teaser.gif -------------------------------------------------------------------------------- /sup-mat/relative-demo.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/snehitvaddi/Deep-Fake_First_Order_Model/da342d15fe15e4ef3ca9c9e4cbed0bbb87e8e58a/sup-mat/relative-demo.gif -------------------------------------------------------------------------------- /sup-mat/vox-teaser.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/snehitvaddi/Deep-Fake_First_Order_Model/da342d15fe15e4ef3ca9c9e4cbed0bbb87e8e58a/sup-mat/vox-teaser.gif -------------------------------------------------------------------------------- /sync_batchnorm/__init__.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # File : __init__.py 3 | # Author : Jiayuan Mao 4 | # Email : maojiayuan@gmail.com 5 | # Date : 27/01/2018 6 | # 7 | # This file is part of Synchronized-BatchNorm-PyTorch. 8 | # https://github.com/vacancy/Synchronized-BatchNorm-PyTorch 9 | # Distributed under MIT License. 10 | 11 | from .batchnorm import SynchronizedBatchNorm1d, SynchronizedBatchNorm2d, SynchronizedBatchNorm3d 12 | from .replicate import DataParallelWithCallback, patch_replication_callback 13 | -------------------------------------------------------------------------------- /sync_batchnorm/batchnorm.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # File : batchnorm.py 3 | # Author : Jiayuan Mao 4 | # Email : maojiayuan@gmail.com 5 | # Date : 27/01/2018 6 | # 7 | # This file is part of Synchronized-BatchNorm-PyTorch. 8 | # https://github.com/vacancy/Synchronized-BatchNorm-PyTorch 9 | # Distributed under MIT License. 10 | 11 | import collections 12 | 13 | import torch 14 | import torch.nn.functional as F 15 | 16 | from torch.nn.modules.batchnorm import _BatchNorm 17 | from torch.nn.parallel._functions import ReduceAddCoalesced, Broadcast 18 | 19 | from .comm import SyncMaster 20 | 21 | __all__ = ['SynchronizedBatchNorm1d', 'SynchronizedBatchNorm2d', 'SynchronizedBatchNorm3d'] 22 | 23 | 24 | def _sum_ft(tensor): 25 | """sum over the first and last dimention""" 26 | return tensor.sum(dim=0).sum(dim=-1) 27 | 28 | 29 | def _unsqueeze_ft(tensor): 30 | """add new dementions at the front and the tail""" 31 | return tensor.unsqueeze(0).unsqueeze(-1) 32 | 33 | 34 | _ChildMessage = collections.namedtuple('_ChildMessage', ['sum', 'ssum', 'sum_size']) 35 | _MasterMessage = collections.namedtuple('_MasterMessage', ['sum', 'inv_std']) 36 | 37 | 38 | class _SynchronizedBatchNorm(_BatchNorm): 39 | def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True): 40 | super(_SynchronizedBatchNorm, self).__init__(num_features, eps=eps, momentum=momentum, affine=affine) 41 | 42 | self._sync_master = SyncMaster(self._data_parallel_master) 43 | 44 | self._is_parallel = False 45 | self._parallel_id = None 46 | self._slave_pipe = None 47 | 48 | def forward(self, input): 49 | # If it is not parallel computation or is in evaluation mode, use PyTorch's implementation. 50 | if not (self._is_parallel and self.training): 51 | return F.batch_norm( 52 | input, self.running_mean, self.running_var, self.weight, self.bias, 53 | self.training, self.momentum, self.eps) 54 | 55 | # Resize the input to (B, C, -1). 56 | input_shape = input.size() 57 | input = input.view(input.size(0), self.num_features, -1) 58 | 59 | # Compute the sum and square-sum. 60 | sum_size = input.size(0) * input.size(2) 61 | input_sum = _sum_ft(input) 62 | input_ssum = _sum_ft(input ** 2) 63 | 64 | # Reduce-and-broadcast the statistics. 65 | if self._parallel_id == 0: 66 | mean, inv_std = self._sync_master.run_master(_ChildMessage(input_sum, input_ssum, sum_size)) 67 | else: 68 | mean, inv_std = self._slave_pipe.run_slave(_ChildMessage(input_sum, input_ssum, sum_size)) 69 | 70 | # Compute the output. 71 | if self.affine: 72 | # MJY:: Fuse the multiplication for speed. 73 | output = (input - _unsqueeze_ft(mean)) * _unsqueeze_ft(inv_std * self.weight) + _unsqueeze_ft(self.bias) 74 | else: 75 | output = (input - _unsqueeze_ft(mean)) * _unsqueeze_ft(inv_std) 76 | 77 | # Reshape it. 78 | return output.view(input_shape) 79 | 80 | def __data_parallel_replicate__(self, ctx, copy_id): 81 | self._is_parallel = True 82 | self._parallel_id = copy_id 83 | 84 | # parallel_id == 0 means master device. 85 | if self._parallel_id == 0: 86 | ctx.sync_master = self._sync_master 87 | else: 88 | self._slave_pipe = ctx.sync_master.register_slave(copy_id) 89 | 90 | def _data_parallel_master(self, intermediates): 91 | """Reduce the sum and square-sum, compute the statistics, and broadcast it.""" 92 | 93 | # Always using same "device order" makes the ReduceAdd operation faster. 94 | # Thanks to:: Tete Xiao (http://tetexiao.com/) 95 | intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device()) 96 | 97 | to_reduce = [i[1][:2] for i in intermediates] 98 | to_reduce = [j for i in to_reduce for j in i] # flatten 99 | target_gpus = [i[1].sum.get_device() for i in intermediates] 100 | 101 | sum_size = sum([i[1].sum_size for i in intermediates]) 102 | sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce) 103 | mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size) 104 | 105 | broadcasted = Broadcast.apply(target_gpus, mean, inv_std) 106 | 107 | outputs = [] 108 | for i, rec in enumerate(intermediates): 109 | outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2]))) 110 | 111 | return outputs 112 | 113 | def _compute_mean_std(self, sum_, ssum, size): 114 | """Compute the mean and standard-deviation with sum and square-sum. This method 115 | also maintains the moving average on the master device.""" 116 | assert size > 1, 'BatchNorm computes unbiased standard-deviation, which requires size > 1.' 117 | mean = sum_ / size 118 | sumvar = ssum - sum_ * mean 119 | unbias_var = sumvar / (size - 1) 120 | bias_var = sumvar / size 121 | 122 | self.running_mean = (1 - self.momentum) * self.running_mean + self.momentum * mean.data 123 | self.running_var = (1 - self.momentum) * self.running_var + self.momentum * unbias_var.data 124 | 125 | return mean, bias_var.clamp(self.eps) ** -0.5 126 | 127 | 128 | class SynchronizedBatchNorm1d(_SynchronizedBatchNorm): 129 | r"""Applies Synchronized Batch Normalization over a 2d or 3d input that is seen as a 130 | mini-batch. 131 | 132 | .. math:: 133 | 134 | y = \frac{x - mean[x]}{ \sqrt{Var[x] + \epsilon}} * gamma + beta 135 | 136 | This module differs from the built-in PyTorch BatchNorm1d as the mean and 137 | standard-deviation are reduced across all devices during training. 138 | 139 | For example, when one uses `nn.DataParallel` to wrap the network during 140 | training, PyTorch's implementation normalize the tensor on each device using 141 | the statistics only on that device, which accelerated the computation and 142 | is also easy to implement, but the statistics might be inaccurate. 143 | Instead, in this synchronized version, the statistics will be computed 144 | over all training samples distributed on multiple devices. 145 | 146 | Note that, for one-GPU or CPU-only case, this module behaves exactly same 147 | as the built-in PyTorch implementation. 148 | 149 | The mean and standard-deviation are calculated per-dimension over 150 | the mini-batches and gamma and beta are learnable parameter vectors 151 | of size C (where C is the input size). 152 | 153 | During training, this layer keeps a running estimate of its computed mean 154 | and variance. The running sum is kept with a default momentum of 0.1. 155 | 156 | During evaluation, this running mean/variance is used for normalization. 157 | 158 | Because the BatchNorm is done over the `C` dimension, computing statistics 159 | on `(N, L)` slices, it's common terminology to call this Temporal BatchNorm 160 | 161 | Args: 162 | num_features: num_features from an expected input of size 163 | `batch_size x num_features [x width]` 164 | eps: a value added to the denominator for numerical stability. 165 | Default: 1e-5 166 | momentum: the value used for the running_mean and running_var 167 | computation. Default: 0.1 168 | affine: a boolean value that when set to ``True``, gives the layer learnable 169 | affine parameters. Default: ``True`` 170 | 171 | Shape: 172 | - Input: :math:`(N, C)` or :math:`(N, C, L)` 173 | - Output: :math:`(N, C)` or :math:`(N, C, L)` (same shape as input) 174 | 175 | Examples: 176 | >>> # With Learnable Parameters 177 | >>> m = SynchronizedBatchNorm1d(100) 178 | >>> # Without Learnable Parameters 179 | >>> m = SynchronizedBatchNorm1d(100, affine=False) 180 | >>> input = torch.autograd.Variable(torch.randn(20, 100)) 181 | >>> output = m(input) 182 | """ 183 | 184 | def _check_input_dim(self, input): 185 | if input.dim() != 2 and input.dim() != 3: 186 | raise ValueError('expected 2D or 3D input (got {}D input)' 187 | .format(input.dim())) 188 | super(SynchronizedBatchNorm1d, self)._check_input_dim(input) 189 | 190 | 191 | class SynchronizedBatchNorm2d(_SynchronizedBatchNorm): 192 | r"""Applies Batch Normalization over a 4d input that is seen as a mini-batch 193 | of 3d inputs 194 | 195 | .. math:: 196 | 197 | y = \frac{x - mean[x]}{ \sqrt{Var[x] + \epsilon}} * gamma + beta 198 | 199 | This module differs from the built-in PyTorch BatchNorm2d as the mean and 200 | standard-deviation are reduced across all devices during training. 201 | 202 | For example, when one uses `nn.DataParallel` to wrap the network during 203 | training, PyTorch's implementation normalize the tensor on each device using 204 | the statistics only on that device, which accelerated the computation and 205 | is also easy to implement, but the statistics might be inaccurate. 206 | Instead, in this synchronized version, the statistics will be computed 207 | over all training samples distributed on multiple devices. 208 | 209 | Note that, for one-GPU or CPU-only case, this module behaves exactly same 210 | as the built-in PyTorch implementation. 211 | 212 | The mean and standard-deviation are calculated per-dimension over 213 | the mini-batches and gamma and beta are learnable parameter vectors 214 | of size C (where C is the input size). 215 | 216 | During training, this layer keeps a running estimate of its computed mean 217 | and variance. The running sum is kept with a default momentum of 0.1. 218 | 219 | During evaluation, this running mean/variance is used for normalization. 220 | 221 | Because the BatchNorm is done over the `C` dimension, computing statistics 222 | on `(N, H, W)` slices, it's common terminology to call this Spatial BatchNorm 223 | 224 | Args: 225 | num_features: num_features from an expected input of 226 | size batch_size x num_features x height x width 227 | eps: a value added to the denominator for numerical stability. 228 | Default: 1e-5 229 | momentum: the value used for the running_mean and running_var 230 | computation. Default: 0.1 231 | affine: a boolean value that when set to ``True``, gives the layer learnable 232 | affine parameters. Default: ``True`` 233 | 234 | Shape: 235 | - Input: :math:`(N, C, H, W)` 236 | - Output: :math:`(N, C, H, W)` (same shape as input) 237 | 238 | Examples: 239 | >>> # With Learnable Parameters 240 | >>> m = SynchronizedBatchNorm2d(100) 241 | >>> # Without Learnable Parameters 242 | >>> m = SynchronizedBatchNorm2d(100, affine=False) 243 | >>> input = torch.autograd.Variable(torch.randn(20, 100, 35, 45)) 244 | >>> output = m(input) 245 | """ 246 | 247 | def _check_input_dim(self, input): 248 | if input.dim() != 4: 249 | raise ValueError('expected 4D input (got {}D input)' 250 | .format(input.dim())) 251 | super(SynchronizedBatchNorm2d, self)._check_input_dim(input) 252 | 253 | 254 | class SynchronizedBatchNorm3d(_SynchronizedBatchNorm): 255 | r"""Applies Batch Normalization over a 5d input that is seen as a mini-batch 256 | of 4d inputs 257 | 258 | .. math:: 259 | 260 | y = \frac{x - mean[x]}{ \sqrt{Var[x] + \epsilon}} * gamma + beta 261 | 262 | This module differs from the built-in PyTorch BatchNorm3d as the mean and 263 | standard-deviation are reduced across all devices during training. 264 | 265 | For example, when one uses `nn.DataParallel` to wrap the network during 266 | training, PyTorch's implementation normalize the tensor on each device using 267 | the statistics only on that device, which accelerated the computation and 268 | is also easy to implement, but the statistics might be inaccurate. 269 | Instead, in this synchronized version, the statistics will be computed 270 | over all training samples distributed on multiple devices. 271 | 272 | Note that, for one-GPU or CPU-only case, this module behaves exactly same 273 | as the built-in PyTorch implementation. 274 | 275 | The mean and standard-deviation are calculated per-dimension over 276 | the mini-batches and gamma and beta are learnable parameter vectors 277 | of size C (where C is the input size). 278 | 279 | During training, this layer keeps a running estimate of its computed mean 280 | and variance. The running sum is kept with a default momentum of 0.1. 281 | 282 | During evaluation, this running mean/variance is used for normalization. 283 | 284 | Because the BatchNorm is done over the `C` dimension, computing statistics 285 | on `(N, D, H, W)` slices, it's common terminology to call this Volumetric BatchNorm 286 | or Spatio-temporal BatchNorm 287 | 288 | Args: 289 | num_features: num_features from an expected input of 290 | size batch_size x num_features x depth x height x width 291 | eps: a value added to the denominator for numerical stability. 292 | Default: 1e-5 293 | momentum: the value used for the running_mean and running_var 294 | computation. Default: 0.1 295 | affine: a boolean value that when set to ``True``, gives the layer learnable 296 | affine parameters. Default: ``True`` 297 | 298 | Shape: 299 | - Input: :math:`(N, C, D, H, W)` 300 | - Output: :math:`(N, C, D, H, W)` (same shape as input) 301 | 302 | Examples: 303 | >>> # With Learnable Parameters 304 | >>> m = SynchronizedBatchNorm3d(100) 305 | >>> # Without Learnable Parameters 306 | >>> m = SynchronizedBatchNorm3d(100, affine=False) 307 | >>> input = torch.autograd.Variable(torch.randn(20, 100, 35, 45, 10)) 308 | >>> output = m(input) 309 | """ 310 | 311 | def _check_input_dim(self, input): 312 | if input.dim() != 5: 313 | raise ValueError('expected 5D input (got {}D input)' 314 | .format(input.dim())) 315 | super(SynchronizedBatchNorm3d, self)._check_input_dim(input) 316 | -------------------------------------------------------------------------------- /sync_batchnorm/comm.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # File : comm.py 3 | # Author : Jiayuan Mao 4 | # Email : maojiayuan@gmail.com 5 | # Date : 27/01/2018 6 | # 7 | # This file is part of Synchronized-BatchNorm-PyTorch. 8 | # https://github.com/vacancy/Synchronized-BatchNorm-PyTorch 9 | # Distributed under MIT License. 10 | 11 | import queue 12 | import collections 13 | import threading 14 | 15 | __all__ = ['FutureResult', 'SlavePipe', 'SyncMaster'] 16 | 17 | 18 | class FutureResult(object): 19 | """A thread-safe future implementation. Used only as one-to-one pipe.""" 20 | 21 | def __init__(self): 22 | self._result = None 23 | self._lock = threading.Lock() 24 | self._cond = threading.Condition(self._lock) 25 | 26 | def put(self, result): 27 | with self._lock: 28 | assert self._result is None, 'Previous result has\'t been fetched.' 29 | self._result = result 30 | self._cond.notify() 31 | 32 | def get(self): 33 | with self._lock: 34 | if self._result is None: 35 | self._cond.wait() 36 | 37 | res = self._result 38 | self._result = None 39 | return res 40 | 41 | 42 | _MasterRegistry = collections.namedtuple('MasterRegistry', ['result']) 43 | _SlavePipeBase = collections.namedtuple('_SlavePipeBase', ['identifier', 'queue', 'result']) 44 | 45 | 46 | class SlavePipe(_SlavePipeBase): 47 | """Pipe for master-slave communication.""" 48 | 49 | def run_slave(self, msg): 50 | self.queue.put((self.identifier, msg)) 51 | ret = self.result.get() 52 | self.queue.put(True) 53 | return ret 54 | 55 | 56 | class SyncMaster(object): 57 | """An abstract `SyncMaster` object. 58 | 59 | - During the replication, as the data parallel will trigger an callback of each module, all slave devices should 60 | call `register(id)` and obtain an `SlavePipe` to communicate with the master. 61 | - During the forward pass, master device invokes `run_master`, all messages from slave devices will be collected, 62 | and passed to a registered callback. 63 | - After receiving the messages, the master device should gather the information and determine to message passed 64 | back to each slave devices. 65 | """ 66 | 67 | def __init__(self, master_callback): 68 | """ 69 | 70 | Args: 71 | master_callback: a callback to be invoked after having collected messages from slave devices. 72 | """ 73 | self._master_callback = master_callback 74 | self._queue = queue.Queue() 75 | self._registry = collections.OrderedDict() 76 | self._activated = False 77 | 78 | def __getstate__(self): 79 | return {'master_callback': self._master_callback} 80 | 81 | def __setstate__(self, state): 82 | self.__init__(state['master_callback']) 83 | 84 | def register_slave(self, identifier): 85 | """ 86 | Register an slave device. 87 | 88 | Args: 89 | identifier: an identifier, usually is the device id. 90 | 91 | Returns: a `SlavePipe` object which can be used to communicate with the master device. 92 | 93 | """ 94 | if self._activated: 95 | assert self._queue.empty(), 'Queue is not clean before next initialization.' 96 | self._activated = False 97 | self._registry.clear() 98 | future = FutureResult() 99 | self._registry[identifier] = _MasterRegistry(future) 100 | return SlavePipe(identifier, self._queue, future) 101 | 102 | def run_master(self, master_msg): 103 | """ 104 | Main entry for the master device in each forward pass. 105 | The messages were first collected from each devices (including the master device), and then 106 | an callback will be invoked to compute the message to be sent back to each devices 107 | (including the master device). 108 | 109 | Args: 110 | master_msg: the message that the master want to send to itself. This will be placed as the first 111 | message when calling `master_callback`. For detailed usage, see `_SynchronizedBatchNorm` for an example. 112 | 113 | Returns: the message to be sent back to the master device. 114 | 115 | """ 116 | self._activated = True 117 | 118 | intermediates = [(0, master_msg)] 119 | for i in range(self.nr_slaves): 120 | intermediates.append(self._queue.get()) 121 | 122 | results = self._master_callback(intermediates) 123 | assert results[0][0] == 0, 'The first result should belongs to the master.' 124 | 125 | for i, res in results: 126 | if i == 0: 127 | continue 128 | self._registry[i].result.put(res) 129 | 130 | for i in range(self.nr_slaves): 131 | assert self._queue.get() is True 132 | 133 | return results[0][1] 134 | 135 | @property 136 | def nr_slaves(self): 137 | return len(self._registry) 138 | -------------------------------------------------------------------------------- /sync_batchnorm/replicate.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # File : replicate.py 3 | # Author : Jiayuan Mao 4 | # Email : maojiayuan@gmail.com 5 | # Date : 27/01/2018 6 | # 7 | # This file is part of Synchronized-BatchNorm-PyTorch. 8 | # https://github.com/vacancy/Synchronized-BatchNorm-PyTorch 9 | # Distributed under MIT License. 10 | 11 | import functools 12 | 13 | from torch.nn.parallel.data_parallel import DataParallel 14 | 15 | __all__ = [ 16 | 'CallbackContext', 17 | 'execute_replication_callbacks', 18 | 'DataParallelWithCallback', 19 | 'patch_replication_callback' 20 | ] 21 | 22 | 23 | class CallbackContext(object): 24 | pass 25 | 26 | 27 | def execute_replication_callbacks(modules): 28 | """ 29 | Execute an replication callback `__data_parallel_replicate__` on each module created by original replication. 30 | 31 | The callback will be invoked with arguments `__data_parallel_replicate__(ctx, copy_id)` 32 | 33 | Note that, as all modules are isomorphism, we assign each sub-module with a context 34 | (shared among multiple copies of this module on different devices). 35 | Through this context, different copies can share some information. 36 | 37 | We guarantee that the callback on the master copy (the first copy) will be called ahead of calling the callback 38 | of any slave copies. 39 | """ 40 | master_copy = modules[0] 41 | nr_modules = len(list(master_copy.modules())) 42 | ctxs = [CallbackContext() for _ in range(nr_modules)] 43 | 44 | for i, module in enumerate(modules): 45 | for j, m in enumerate(module.modules()): 46 | if hasattr(m, '__data_parallel_replicate__'): 47 | m.__data_parallel_replicate__(ctxs[j], i) 48 | 49 | 50 | class DataParallelWithCallback(DataParallel): 51 | """ 52 | Data Parallel with a replication callback. 53 | 54 | An replication callback `__data_parallel_replicate__` of each module will be invoked after being created by 55 | original `replicate` function. 56 | The callback will be invoked with arguments `__data_parallel_replicate__(ctx, copy_id)` 57 | 58 | Examples: 59 | > sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False) 60 | > sync_bn = DataParallelWithCallback(sync_bn, device_ids=[0, 1]) 61 | # sync_bn.__data_parallel_replicate__ will be invoked. 62 | """ 63 | 64 | def replicate(self, module, device_ids): 65 | modules = super(DataParallelWithCallback, self).replicate(module, device_ids) 66 | execute_replication_callbacks(modules) 67 | return modules 68 | 69 | 70 | def patch_replication_callback(data_parallel): 71 | """ 72 | Monkey-patch an existing `DataParallel` object. Add the replication callback. 73 | Useful when you have customized `DataParallel` implementation. 74 | 75 | Examples: 76 | > sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False) 77 | > sync_bn = DataParallel(sync_bn, device_ids=[0, 1]) 78 | > patch_replication_callback(sync_bn) 79 | # this is equivalent to 80 | > sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False) 81 | > sync_bn = DataParallelWithCallback(sync_bn, device_ids=[0, 1]) 82 | """ 83 | 84 | assert isinstance(data_parallel, DataParallel) 85 | 86 | old_replicate = data_parallel.replicate 87 | 88 | @functools.wraps(old_replicate) 89 | def new_replicate(module, device_ids): 90 | modules = old_replicate(module, device_ids) 91 | execute_replication_callbacks(modules) 92 | return modules 93 | 94 | data_parallel.replicate = new_replicate 95 | -------------------------------------------------------------------------------- /sync_batchnorm/unittest.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # File : unittest.py 3 | # Author : Jiayuan Mao 4 | # Email : maojiayuan@gmail.com 5 | # Date : 27/01/2018 6 | # 7 | # This file is part of Synchronized-BatchNorm-PyTorch. 8 | # https://github.com/vacancy/Synchronized-BatchNorm-PyTorch 9 | # Distributed under MIT License. 10 | 11 | import unittest 12 | 13 | import numpy as np 14 | from torch.autograd import Variable 15 | 16 | 17 | def as_numpy(v): 18 | if isinstance(v, Variable): 19 | v = v.data 20 | return v.cpu().numpy() 21 | 22 | 23 | class TorchTestCase(unittest.TestCase): 24 | def assertTensorClose(self, a, b, atol=1e-3, rtol=1e-3): 25 | npa, npb = as_numpy(a), as_numpy(b) 26 | self.assertTrue( 27 | np.allclose(npa, npb, atol=atol), 28 | 'Tensor close check failed\n{}\n{}\nadiff={}, rdiff={}'.format(a, b, np.abs(npa - npb).max(), np.abs((npa - npb) / np.fmax(npa, 1e-5)).max()) 29 | ) 30 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | from tqdm import trange 2 | import torch 3 | 4 | from torch.utils.data import DataLoader 5 | 6 | from logger import Logger 7 | from modules.model import GeneratorFullModel, DiscriminatorFullModel 8 | 9 | from torch.optim.lr_scheduler import MultiStepLR 10 | 11 | from sync_batchnorm import DataParallelWithCallback 12 | 13 | from frames_dataset import DatasetRepeater 14 | 15 | 16 | def train(config, generator, discriminator, kp_detector, checkpoint, log_dir, dataset, device_ids): 17 | train_params = config['train_params'] 18 | 19 | optimizer_generator = torch.optim.Adam(generator.parameters(), lr=train_params['lr_generator'], betas=(0.5, 0.999)) 20 | optimizer_discriminator = torch.optim.Adam(discriminator.parameters(), lr=train_params['lr_discriminator'], betas=(0.5, 0.999)) 21 | optimizer_kp_detector = torch.optim.Adam(kp_detector.parameters(), lr=train_params['lr_kp_detector'], betas=(0.5, 0.999)) 22 | 23 | if checkpoint is not None: 24 | start_epoch = Logger.load_cpk(checkpoint, generator, discriminator, kp_detector, 25 | optimizer_generator, optimizer_discriminator, 26 | None if train_params['lr_kp_detector'] == 0 else optimizer_kp_detector) 27 | else: 28 | start_epoch = 0 29 | 30 | scheduler_generator = MultiStepLR(optimizer_generator, train_params['epoch_milestones'], gamma=0.1, 31 | last_epoch=start_epoch - 1) 32 | scheduler_discriminator = MultiStepLR(optimizer_discriminator, train_params['epoch_milestones'], gamma=0.1, 33 | last_epoch=start_epoch - 1) 34 | scheduler_kp_detector = MultiStepLR(optimizer_kp_detector, train_params['epoch_milestones'], gamma=0.1, 35 | last_epoch=-1 + start_epoch * (train_params['lr_kp_detector'] != 0)) 36 | 37 | if 'num_repeats' in train_params or train_params['num_repeats'] != 1: 38 | dataset = DatasetRepeater(dataset, train_params['num_repeats']) 39 | dataloader = DataLoader(dataset, batch_size=train_params['batch_size'], shuffle=True, num_workers=6, drop_last=True) 40 | 41 | generator_full = GeneratorFullModel(kp_detector, generator, discriminator, train_params) 42 | discriminator_full = DiscriminatorFullModel(kp_detector, generator, discriminator, train_params) 43 | 44 | if torch.cuda.is_available(): 45 | generator_full = DataParallelWithCallback(generator_full, device_ids=device_ids) 46 | discriminator_full = DataParallelWithCallback(discriminator_full, device_ids=device_ids) 47 | 48 | with Logger(log_dir=log_dir, visualizer_params=config['visualizer_params'], checkpoint_freq=train_params['checkpoint_freq']) as logger: 49 | for epoch in trange(start_epoch, train_params['num_epochs']): 50 | for x in dataloader: 51 | losses_generator, generated = generator_full(x) 52 | 53 | loss_values = [val.mean() for val in losses_generator.values()] 54 | loss = sum(loss_values) 55 | 56 | loss.backward() 57 | optimizer_generator.step() 58 | optimizer_generator.zero_grad() 59 | optimizer_kp_detector.step() 60 | optimizer_kp_detector.zero_grad() 61 | 62 | if train_params['loss_weights']['generator_gan'] != 0: 63 | optimizer_discriminator.zero_grad() 64 | losses_discriminator = discriminator_full(x, generated) 65 | loss_values = [val.mean() for val in losses_discriminator.values()] 66 | loss = sum(loss_values) 67 | 68 | loss.backward() 69 | optimizer_discriminator.step() 70 | optimizer_discriminator.zero_grad() 71 | else: 72 | losses_discriminator = {} 73 | 74 | losses_generator.update(losses_discriminator) 75 | losses = {key: value.mean().detach().data.cpu().numpy() for key, value in losses_generator.items()} 76 | logger.log_iter(losses=losses) 77 | 78 | scheduler_generator.step() 79 | scheduler_discriminator.step() 80 | scheduler_kp_detector.step() 81 | 82 | logger.log_epoch(epoch, {'generator': generator, 83 | 'discriminator': discriminator, 84 | 'kp_detector': kp_detector, 85 | 'optimizer_generator': optimizer_generator, 86 | 'optimizer_discriminator': optimizer_discriminator, 87 | 'optimizer_kp_detector': optimizer_kp_detector}, inp=x, out=generated) 88 | --------------------------------------------------------------------------------