├── LICENSE ├── README.md ├── convlstm.py ├── dataloader.py ├── model.py ├── requirements.txt ├── test.py ├── train.py └── utils.py /LICENSE: -------------------------------------------------------------------------------- 1 | Copyright (c) 2021-present, Sangmin Lee 2 | 3 | Attribution-NonCommercial 4.0 International 4 | 5 | ======================================================================= 6 | 7 | Creative Commons Corporation ("Creative Commons") is not a law firm and 8 | does not provide legal services or legal advice. Distribution of 9 | Creative Commons public licenses does not create a lawyer-client or 10 | other relationship. Creative Commons makes its licenses and related 11 | information available on an "as-is" basis. Creative Commons gives no 12 | warranties regarding its licenses, any material licensed under their 13 | terms and conditions, or any related information. Creative Commons 14 | disclaims all liability for damages resulting from their use to the 15 | fullest extent possible. 16 | 17 | Using Creative Commons Public Licenses 18 | 19 | Creative Commons public licenses provide a standard set of terms and 20 | conditions that creators and other rights holders may use to share 21 | original works of authorship and other material subject to copyright 22 | and certain other rights specified in the public license below. The 23 | following considerations are for informational purposes only, are not 24 | exhaustive, and do not form part of our licenses. 25 | 26 | Considerations for licensors: Our public licenses are 27 | intended for use by those authorized to give the public 28 | permission to use material in ways otherwise restricted by 29 | copyright and certain other rights. Our licenses are 30 | irrevocable. Licensors should read and understand the terms 31 | and conditions of the license they choose before applying it. 32 | Licensors should also secure all rights necessary before 33 | applying our licenses so that the public can reuse the 34 | material as expected. Licensors should clearly mark any 35 | material not subject to the license. This includes other CC- 36 | licensed material, or material used under an exception or 37 | limitation to copyright. More considerations for licensors: 38 | wiki.creativecommons.org/Considerations_for_licensors 39 | 40 | Considerations for the public: By using one of our public 41 | licenses, a licensor grants the public permission to use the 42 | licensed material under specified terms and conditions. If 43 | the licensor's permission is not necessary for any reason--for 44 | example, because of any applicable exception or limitation to 45 | copyright--then that use is not regulated by the license. Our 46 | licenses grant only permissions under copyright and certain 47 | other rights that a licensor has authority to grant. Use of 48 | the licensed material may still be restricted for other 49 | reasons, including because others have copyright or other 50 | rights in the material. A licensor may make special requests, 51 | such as asking that all changes be marked or described. 52 | Although not required by our licenses, you are encouraged to 53 | respect those requests where reasonable. More considerations 54 | for the public: 55 | wiki.creativecommons.org/Considerations_for_licensees 56 | 57 | ======================================================================= 58 | 59 | Creative Commons Attribution-NonCommercial 4.0 International Public 60 | License 61 | 62 | By exercising the Licensed Rights (defined below), You accept and agree 63 | to be bound by the terms and conditions of this Creative Commons 64 | Attribution-NonCommercial 4.0 International Public License ("Public 65 | License"). To the extent this Public License may be interpreted as a 66 | contract, You are granted the Licensed Rights in consideration of Your 67 | acceptance of these terms and conditions, and the Licensor grants You 68 | such rights in consideration of benefits the Licensor receives from 69 | making the Licensed Material available under these terms and 70 | conditions. 71 | 72 | 73 | Section 1 -- Definitions. 74 | 75 | a. Adapted Material means material subject to Copyright and Similar 76 | Rights that is derived from or based upon the Licensed Material 77 | and in which the Licensed Material is translated, altered, 78 | arranged, transformed, or otherwise modified in a manner requiring 79 | permission under the Copyright and Similar Rights held by the 80 | Licensor. For purposes of this Public License, where the Licensed 81 | Material is a musical work, performance, or sound recording, 82 | Adapted Material is always produced where the Licensed Material is 83 | synched in timed relation with a moving image. 84 | 85 | b. Adapter's License means the license You apply to Your Copyright 86 | and Similar Rights in Your contributions to Adapted Material in 87 | accordance with the terms and conditions of this Public License. 88 | 89 | c. Copyright and Similar Rights means copyright and/or similar rights 90 | closely related to copyright including, without limitation, 91 | performance, broadcast, sound recording, and Sui Generis Database 92 | Rights, without regard to how the rights are labeled or 93 | categorized. For purposes of this Public License, the rights 94 | specified in Section 2(b)(1)-(2) are not Copyright and Similar 95 | Rights. 96 | d. Effective Technological Measures means those measures that, in the 97 | absence of proper authority, may not be circumvented under laws 98 | fulfilling obligations under Article 11 of the WIPO Copyright 99 | Treaty adopted on December 20, 1996, and/or similar international 100 | agreements. 101 | 102 | e. Exceptions and Limitations means fair use, fair dealing, and/or 103 | any other exception or limitation to Copyright and Similar Rights 104 | that applies to Your use of the Licensed Material. 105 | 106 | f. Licensed Material means the artistic or literary work, database, 107 | or other material to which the Licensor applied this Public 108 | License. 109 | 110 | g. Licensed Rights means the rights granted to You subject to the 111 | terms and conditions of this Public License, which are limited to 112 | all Copyright and Similar Rights that apply to Your use of the 113 | Licensed Material and that the Licensor has authority to license. 114 | 115 | h. Licensor means the individual(s) or entity(ies) granting rights 116 | under this Public License. 117 | 118 | i. NonCommercial means not primarily intended for or directed towards 119 | commercial advantage or monetary compensation. For purposes of 120 | this Public License, the exchange of the Licensed Material for 121 | other material subject to Copyright and Similar Rights by digital 122 | file-sharing or similar means is NonCommercial provided there is 123 | no payment of monetary compensation in connection with the 124 | exchange. 125 | 126 | j. Share means to provide material to the public by any means or 127 | process that requires permission under the Licensed Rights, such 128 | as reproduction, public display, public performance, distribution, 129 | dissemination, communication, or importation, and to make material 130 | available to the public including in ways that members of the 131 | public may access the material from a place and at a time 132 | individually chosen by them. 133 | 134 | k. Sui Generis Database Rights means rights other than copyright 135 | resulting from Directive 96/9/EC of the European Parliament and of 136 | the Council of 11 March 1996 on the legal protection of databases, 137 | as amended and/or succeeded, as well as other essentially 138 | equivalent rights anywhere in the world. 139 | 140 | l. You means the individual or entity exercising the Licensed Rights 141 | under this Public License. Your has a corresponding meaning. 142 | 143 | 144 | Section 2 -- Scope. 145 | 146 | a. License grant. 147 | 148 | 1. Subject to the terms and conditions of this Public License, 149 | the Licensor hereby grants You a worldwide, royalty-free, 150 | non-sublicensable, non-exclusive, irrevocable license to 151 | exercise the Licensed Rights in the Licensed Material to: 152 | 153 | a. reproduce and Share the Licensed Material, in whole or 154 | in part, for NonCommercial purposes only; and 155 | 156 | b. produce, reproduce, and Share Adapted Material for 157 | NonCommercial purposes only. 158 | 159 | 2. Exceptions and Limitations. For the avoidance of doubt, where 160 | Exceptions and Limitations apply to Your use, this Public 161 | License does not apply, and You do not need to comply with 162 | its terms and conditions. 163 | 164 | 3. Term. The term of this Public License is specified in Section 165 | 6(a). 166 | 167 | 4. Media and formats; technical modifications allowed. The 168 | Licensor authorizes You to exercise the Licensed Rights in 169 | all media and formats whether now known or hereafter created, 170 | and to make technical modifications necessary to do so. The 171 | Licensor waives and/or agrees not to assert any right or 172 | authority to forbid You from making technical modifications 173 | necessary to exercise the Licensed Rights, including 174 | technical modifications necessary to circumvent Effective 175 | Technological Measures. For purposes of this Public License, 176 | simply making modifications authorized by this Section 2(a) 177 | (4) never produces Adapted Material. 178 | 179 | 5. Downstream recipients. 180 | 181 | a. Offer from the Licensor -- Licensed Material. Every 182 | recipient of the Licensed Material automatically 183 | receives an offer from the Licensor to exercise the 184 | Licensed Rights under the terms and conditions of this 185 | Public License. 186 | 187 | b. No downstream restrictions. You may not offer or impose 188 | any additional or different terms or conditions on, or 189 | apply any Effective Technological Measures to, the 190 | Licensed Material if doing so restricts exercise of the 191 | Licensed Rights by any recipient of the Licensed 192 | Material. 193 | 194 | 6. No endorsement. Nothing in this Public License constitutes or 195 | may be construed as permission to assert or imply that You 196 | are, or that Your use of the Licensed Material is, connected 197 | with, or sponsored, endorsed, or granted official status by, 198 | the Licensor or others designated to receive attribution as 199 | provided in Section 3(a)(1)(A)(i). 200 | 201 | b. Other rights. 202 | 203 | 1. Moral rights, such as the right of integrity, are not 204 | licensed under this Public License, nor are publicity, 205 | privacy, and/or other similar personality rights; however, to 206 | the extent possible, the Licensor waives and/or agrees not to 207 | assert any such rights held by the Licensor to the limited 208 | extent necessary to allow You to exercise the Licensed 209 | Rights, but not otherwise. 210 | 211 | 2. Patent and trademark rights are not licensed under this 212 | Public License. 213 | 214 | 3. To the extent possible, the Licensor waives any right to 215 | collect royalties from You for the exercise of the Licensed 216 | Rights, whether directly or through a collecting society 217 | under any voluntary or waivable statutory or compulsory 218 | licensing scheme. In all other cases the Licensor expressly 219 | reserves any right to collect such royalties, including when 220 | the Licensed Material is used other than for NonCommercial 221 | purposes. 222 | 223 | 224 | Section 3 -- License Conditions. 225 | 226 | Your exercise of the Licensed Rights is expressly made subject to the 227 | following conditions. 228 | 229 | a. Attribution. 230 | 231 | 1. If You Share the Licensed Material (including in modified 232 | form), You must: 233 | 234 | a. retain the following if it is supplied by the Licensor 235 | with the Licensed Material: 236 | 237 | i. identification of the creator(s) of the Licensed 238 | Material and any others designated to receive 239 | attribution, in any reasonable manner requested by 240 | the Licensor (including by pseudonym if 241 | designated); 242 | 243 | ii. a copyright notice; 244 | 245 | iii. a notice that refers to this Public License; 246 | 247 | iv. a notice that refers to the disclaimer of 248 | warranties; 249 | 250 | v. a URI or hyperlink to the Licensed Material to the 251 | extent reasonably practicable; 252 | 253 | b. indicate if You modified the Licensed Material and 254 | retain an indication of any previous modifications; and 255 | 256 | c. indicate the Licensed Material is licensed under this 257 | Public License, and include the text of, or the URI or 258 | hyperlink to, this Public License. 259 | 260 | 2. You may satisfy the conditions in Section 3(a)(1) in any 261 | reasonable manner based on the medium, means, and context in 262 | which You Share the Licensed Material. For example, it may be 263 | reasonable to satisfy the conditions by providing a URI or 264 | hyperlink to a resource that includes the required 265 | information. 266 | 267 | 3. If requested by the Licensor, You must remove any of the 268 | information required by Section 3(a)(1)(A) to the extent 269 | reasonably practicable. 270 | 271 | 4. If You Share Adapted Material You produce, the Adapter's 272 | License You apply must not prevent recipients of the Adapted 273 | Material from complying with this Public License. 274 | 275 | 276 | Section 4 -- Sui Generis Database Rights. 277 | 278 | Where the Licensed Rights include Sui Generis Database Rights that 279 | apply to Your use of the Licensed Material: 280 | 281 | a. for the avoidance of doubt, Section 2(a)(1) grants You the right 282 | to extract, reuse, reproduce, and Share all or a substantial 283 | portion of the contents of the database for NonCommercial purposes 284 | only; 285 | 286 | b. if You include all or a substantial portion of the database 287 | contents in a database in which You have Sui Generis Database 288 | Rights, then the database in which You have Sui Generis Database 289 | Rights (but not its individual contents) is Adapted Material; and 290 | 291 | c. You must comply with the conditions in Section 3(a) if You Share 292 | all or a substantial portion of the contents of the database. 293 | 294 | For the avoidance of doubt, this Section 4 supplements and does not 295 | replace Your obligations under this Public License where the Licensed 296 | Rights include other Copyright and Similar Rights. 297 | 298 | 299 | Section 5 -- Disclaimer of Warranties and Limitation of Liability. 300 | 301 | a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE 302 | EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS 303 | AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF 304 | ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS, 305 | IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION, 306 | WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR 307 | PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS, 308 | ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT 309 | KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT 310 | ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU. 311 | 312 | b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE 313 | TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION, 314 | NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT, 315 | INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES, 316 | COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR 317 | USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN 318 | ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR 319 | DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR 320 | IN PART, THIS LIMITATION MAY NOT APPLY TO YOU. 321 | 322 | c. The disclaimer of warranties and limitation of liability provided 323 | above shall be interpreted in a manner that, to the extent 324 | possible, most closely approximates an absolute disclaimer and 325 | waiver of all liability. 326 | 327 | 328 | Section 6 -- Term and Termination. 329 | 330 | a. This Public License applies for the term of the Copyright and 331 | Similar Rights licensed here. However, if You fail to comply with 332 | this Public License, then Your rights under this Public License 333 | terminate automatically. 334 | 335 | b. Where Your right to use the Licensed Material has terminated under 336 | Section 6(a), it reinstates: 337 | 338 | 1. automatically as of the date the violation is cured, provided 339 | it is cured within 30 days of Your discovery of the 340 | violation; or 341 | 342 | 2. upon express reinstatement by the Licensor. 343 | 344 | For the avoidance of doubt, this Section 6(b) does not affect any 345 | right the Licensor may have to seek remedies for Your violations 346 | of this Public License. 347 | 348 | c. For the avoidance of doubt, the Licensor may also offer the 349 | Licensed Material under separate terms or conditions or stop 350 | distributing the Licensed Material at any time; however, doing so 351 | will not terminate this Public License. 352 | 353 | d. Sections 1, 5, 6, 7, and 8 survive termination of this Public 354 | License. 355 | 356 | 357 | Section 7 -- Other Terms and Conditions. 358 | 359 | a. The Licensor shall not be bound by any additional or different 360 | terms or conditions communicated by You unless expressly agreed. 361 | 362 | b. Any arrangements, understandings, or agreements regarding the 363 | Licensed Material not stated herein are separate from and 364 | independent of the terms and conditions of this Public License. 365 | 366 | 367 | Section 8 -- Interpretation. 368 | 369 | a. For the avoidance of doubt, this Public License does not, and 370 | shall not be interpreted to, reduce, limit, restrict, or impose 371 | conditions on any use of the Licensed Material that could lawfully 372 | be made without permission under this Public License. 373 | 374 | b. To the extent possible, if any provision of this Public License is 375 | deemed unenforceable, it shall be automatically reformed to the 376 | minimum extent necessary to make it enforceable. If the provision 377 | cannot be reformed, it shall be severed from this Public License 378 | without affecting the enforceability of the remaining terms and 379 | conditions. 380 | 381 | c. No term or condition of this Public License will be waived and no 382 | failure to comply consented to unless expressly agreed to by the 383 | Licensor. 384 | 385 | d. Nothing in this Public License constitutes or may be interpreted 386 | as a limitation upon, or waiver of, any privileges and immunities 387 | that apply to the Licensor or You, including from the legal 388 | processes of any jurisdiction or authority. 389 | 390 | ======================================================================= 391 | 392 | Creative Commons is not a party to its public 393 | licenses. Notwithstanding, Creative Commons may elect to apply one of 394 | its public licenses to material it publishes and in those instances 395 | will be considered the “Licensor.” The text of the Creative Commons 396 | public licenses is dedicated to the public domain under the CC0 Public 397 | Domain Dedication. Except for the limited purpose of indicating that 398 | material is shared under a Creative Commons public license or as 399 | otherwise permitted by the Creative Commons policies published at 400 | creativecommons.org/policies, Creative Commons does not authorize the 401 | use of the trademark "Creative Commons" or any other trademark or logo 402 | of Creative Commons without its prior written consent including, 403 | without limitation, in connection with any unauthorized modifications 404 | to any of its public licenses or any other arrangements, 405 | understandings, or agreements concerning use of licensed material. For 406 | the avoidance of doubt, this paragraph does not form part of the 407 | public licenses. 408 | 409 | Creative Commons may be contacted at creativecommons.org. 410 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ## Video Prediction Recalling Long-term Motion Context via Memory Alignment Learning 2 | 3 | 4 |
5 | 6 | 7 | > 8 | This repository contains the official PyTorch implementation of the following paper: 9 | > **Video Prediction Recalling Long-term Motion Context via Memory Alignment Learning (CVPR 2021 Oral)**
10 | > Sangmin Lee, Hak Gu Kim, Dae Hwi Choi, Hyung-Il Kim, and Yong Man Ro
11 | > Paper: https://arxiv.org/abs/2104.00924
12 | > 13 | > **Abstract** *Our work addresses long-term motion context issues for predicting future frames. To predict the future precisely, it is required to capture which long-term motion context (e.g., walking or running) the input motion (e.g., leg movement) belongs to. The bottlenecks arising when dealing with the long-term motion context are: (i) how to predict the long-term motion context naturally matching input sequences with limited dynamics, (ii) how to predict the long-term motion context with high-dimensionality (e.g., complex motion). To address the issues, we propose novel motion context-aware video prediction. To solve the bottleneck (i), we introduce a long-term motion context memory (LMC-Memory) with memory alignment learning. The proposed memory alignment learning enables to store long-term motion contexts into the memory and to match them with sequences including limited dynamics. As a result, the long-term context can be recalled from the limited input sequence. In addition, to resolve the bottleneck (ii), we propose memory query decomposition to store local motion context (i.e., low-dimensional dynamics) and recall the suitable local context for each local part of the input individually. It enables to boost the alignment effects of the memory. Experimental results show that the proposed method outperforms other sophisticated RNN-based methods, especially in long-term condition. Further, we validate the effectiveness of the proposed network designs by conducting ablation studies and memory feature analysis. The source code of this work is available.* 14 | 15 | ## Preparation 16 | 17 | ### Requirements 18 | - python 3 19 | - pytorch 1.6+ 20 | - opencv-python 21 | - scikit-image 22 | - lpips 23 | - numpy 24 | 25 | ### Datasets 26 | This repository supports Moving-MNIST and KTH-Action datasets. 27 | - [Moving-MNIST](https://github.com/jthsieh/DDPAE-video-prediction/blob/master/data/moving_mnist.py) 28 | - [KTH-Action](https://www.csc.kth.se/cvap/actions/) 29 | 30 | After obtaining the datasets, preprocess the data as image files (refer to below). 31 | ```shell 32 | # Dataset preparation example: 33 | movingmnist 34 | ├── train 35 | │ ├── video_00000 36 | │ │ ├── frame_00000.jpg 37 | ... 38 | │ │ ├── frame_xxxxx.jpg 39 | ... 40 | │ ├── video_xxxxx 41 | ``` 42 | 43 | ## Training the Model 44 | `train.py` saves the weights in `--checkpoint_save_dir` and shows the training logs. 45 | 46 | To train the model, run following command: 47 | ```shell 48 | # Training example for Moving-MNIST 49 | python train.py \ 50 | --dataset 'movingmnist' \ 51 | --train_data_dir 'enter_the_path' --valid_data_dir 'enter_the_path' \ 52 | --checkpoint_save_dir './checkpoints' \ 53 | --img_size 64 --img_channel 1 --memory_size 100 \ 54 | --short_len 10 --long_len 30 --out_len 30 \ 55 | --batch_size 128 --lr 0.0002 --iterations 300000 56 | ``` 57 | ```shell 58 | # Training example for KTH-Action 59 | python train.py \ 60 | --dataset 'kth' \ 61 | --train_data_dir 'enter_the_path' --valid_data_dir 'enter_the_path' \ 62 | --checkpoint_save_dir './checkpoints' \ 63 | --img_size 128 --img_channel 1 --memory_size 100 \ 64 | --short_len 10 --long_len 40 --out_len 40 \ 65 | --batch_size 32 --lr 0.0002 --iterations 300000 66 | ``` 67 | Descriptions of training parameters are as follows: 68 | - `--dataset`: training dataset (movingmnist or kth) 69 | - `--train_data_dir`: directory of training set `--valid_data_dir`: directory of validation set 70 | - `--checkpoint_save_dir`: directory for saving checkpoints 71 | - `--img_size`: height and width of frame `--img_channel`: channel of frame `--memory_size`: memory slot size 72 | - `--short_len`: number of short frames `--long_len`: number of long frames `--out_len`: number of output frames 73 | - `--batch_size`: mini-batch size `--lr`: learning rate `--iterations`: number of total iterations 74 | - Refer to `train.py` for the other training parameters 75 | 76 | ## Testing the Model 77 | `test.py` saves the predicted frames in `--test_result_dir` or evalute the performances. 78 | 79 | To test the model, run following command: 80 | ```shell 81 | # Testing example for Moving-MNIST 82 | python test.py \ 83 | --dataset 'movingmnist' --make_frame True \ 84 | --test_data_dir 'enter_the_path' --test_result_dir 'enter_the_path' \ 85 | --checkpoint_load_file 'enter_the_path' \ 86 | --img_size 64 --img_channel 1 --memory_size 100 \ 87 | --short_len 10 --out_len 30 \ 88 | --batch_size 8 89 | ``` 90 | ```shell 91 | # Testing example for KTH-Action 92 | python test.py \ 93 | --dataset 'kth' --make_frame True \ 94 | --test_data_dir 'enter_the_path' --test_result_dir 'enter_the_path' \ 95 | --checkpoint_load_file 'enter_the_path' \ 96 | --img_size 128 --img_channel 1 --memory_size 100 \ 97 | --short_len 10 --out_len 40 \ 98 | --batch_size 8 99 | ``` 100 | Descriptions of testing parameters are as follows: 101 | - `--dataset`: test dataset (movingmnist or kth) `--make_frame`: whether to generate predicted frames 102 | - `--test_data_dir`: directory of test set `--test_result_dir`: directory for saving predicted frames 103 | - `--checkpoint_load_file`: file path for loading checkpoint 104 | - `--img_size`: height and width of frame `--img_channel`: channel of frame `--memory_size`: memory slot size 105 | - `--short_len`: number of short frames `--out_len`: number of output frames 106 | - `--batch_size`: mini-batch size 107 | - Refer to `test.py` for the other testing parameters 108 | 109 | ## Pretrained Models 110 | You can download the pretrained models. 111 | - [Pretrained model for Moving-MNIST](https://www.dropbox.com/s/c2yl2f7znzmj8mf/trained_file_movingmnist.pt?dl=0) 112 | - [Pretrained model for KTH-Action](https://www.dropbox.com/s/nt015y70moqgy76/trained_file_kth.pt?dl=0) 113 | 114 | ## Citation 115 | If you find this work useful in your research, please cite the paper: 116 | ``` 117 | @inproceedings{lee2021video, 118 | title={Video Prediction Recalling Long-term Motion Context via Memory Alignment Learning}, 119 | author={Lee, Sangmin and Kim, Hak Gu and Choi, Dae Hwi and Kim, Hyung-Il and Ro, Yong Man}, 120 | booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 121 | year={2021} 122 | } 123 | ``` 124 | -------------------------------------------------------------------------------- /convlstm.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | 5 | class NPUnit(nn.Module): 6 | def __init__(self, in_channels, out_channels, kernel_size): 7 | super(NPUnit, self).__init__() 8 | same_padding = int((kernel_size[0]-1)/2) 9 | self.conv2d_x = nn.Conv2d(in_channels=in_channels, out_channels=4*out_channels, 10 | kernel_size=kernel_size, stride=1, padding=same_padding, bias=True) 11 | self.conv2d_h = nn.Conv2d(in_channels=out_channels, out_channels=4*out_channels, 12 | kernel_size=kernel_size, stride=1, padding=same_padding, bias=True) 13 | 14 | def forward(self, x, h, c): 15 | x_after_conv = self.conv2d_x(x) 16 | h_after_conv = self.conv2d_h(h) 17 | xi, xc, xf, xo = torch.chunk(x_after_conv, 4, dim=1) 18 | hi, hc, hf, ho = torch.chunk(h_after_conv, 4, dim=1) 19 | 20 | it = torch.sigmoid(xi+hi) 21 | ft = torch.sigmoid(xf+hf) 22 | new_c = (ft*c)+(it*torch.tanh(xc+hc)) 23 | ot = torch.sigmoid(xo+ho) 24 | new_h = ot*torch.tanh(new_c) 25 | 26 | return new_h, new_c 27 | -------------------------------------------------------------------------------- /dataloader.py: -------------------------------------------------------------------------------- 1 | from torch.utils.data import Dataset 2 | 3 | import cv2 4 | import os 5 | import numpy as np 6 | 7 | 8 | def make_dataset(dataset_dir): 9 | frame_path = [] 10 | # Find and loop over all the clips in root `dir`. 11 | for index, folder in enumerate(sorted(os.listdir(dataset_dir))): 12 | clipsFolderPath = os.path.join(dataset_dir, folder) 13 | # Skip items which are not folders. 14 | if not (os.path.isdir(clipsFolderPath)): 15 | continue 16 | frame_path.append([]) 17 | # Find and loop over all the frames inside the clip. 18 | for image in sorted(os.listdir(clipsFolderPath)): 19 | # Add path to list. 20 | frame_path[index].append(os.path.join(clipsFolderPath, image)) 21 | return frame_path 22 | 23 | class MovingMNIST(Dataset): 24 | def __init__(self, dataset_dir, seq_len, train=True): 25 | self.frame_path = make_dataset(dataset_dir) 26 | self.seq_len = seq_len 27 | self.train = train 28 | 29 | self.clips = [] 30 | for video_i in range(len(self.frame_path)): 31 | video_frame_num = len(self.frame_path[video_i]) 32 | self.clips += [(video_i, t) for t in range(video_frame_num - seq_len + 1)] if train \ 33 | else [(video_i, t * seq_len) for t in range(video_frame_num // seq_len)] 34 | 35 | def __getitem__(self, idx): 36 | (video_idx, data_start) = self.clips[idx] 37 | sample = [] 38 | for frame_range_i in range(data_start, data_start+self.seq_len): 39 | frame = cv2.imread(self.frame_path[video_idx][frame_range_i], cv2.IMREAD_GRAYSCALE) 40 | frame = np.expand_dims(frame, axis=0) 41 | frame = frame.astype(np.float32) 42 | frame = frame/255.0 43 | sample.append(frame) 44 | return sample 45 | 46 | def __len__(self): 47 | return len(self.clips) 48 | 49 | class KTH(Dataset): 50 | def __init__(self, dataset_dir, seq_len, train=True): 51 | self.frame_path = make_dataset(dataset_dir) 52 | self.seq_len = seq_len 53 | self.train = train 54 | 55 | self.clips = [] 56 | for video_i in range(len(self.frame_path)): 57 | video_frame_num = len(self.frame_path[video_i]) 58 | self.clips += [(video_i, t) for t in range(video_frame_num - seq_len + 1)] if train \ 59 | else [(video_i, t * seq_len) for t in range(video_frame_num // seq_len)] 60 | 61 | def __getitem__(self, idx): 62 | (video_idx, data_start) = self.clips[idx] 63 | sample = [] 64 | for frame_range_i in range(data_start, data_start+self.seq_len): 65 | frame = cv2.imread(self.frame_path[video_idx][frame_range_i], cv2.IMREAD_GRAYSCALE) 66 | frame = np.expand_dims(frame, axis=0) 67 | frame = frame.astype(np.float32) 68 | frame = frame/255.0 69 | sample.append(frame) 70 | return sample 71 | 72 | def __len__(self): 73 | return len(self.clips) 74 | 75 | -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | import convlstm 2 | 3 | import torch 4 | import torch.nn as nn 5 | from torch.nn import functional as F 6 | 7 | import copy 8 | 9 | 10 | class Predictor(nn.Module): 11 | def __init__(self, args): 12 | super(Predictor, self).__init__() 13 | self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') 14 | self.encoder = nn.Sequential( 15 | nn.Conv2d(in_channels=1, out_channels=64, kernel_size=(3, 3), stride=2, padding=1), 16 | nn.ELU(), 17 | nn.Conv2d(in_channels=64, out_channels=64, kernel_size=(3, 3), stride=1, padding=1), 18 | nn.ELU(), 19 | nn.Conv2d(in_channels=64, out_channels=128, kernel_size=(3, 3), stride=2, padding=1), 20 | nn.ELU(), 21 | nn.Conv2d(in_channels=128, out_channels=128, kernel_size=(3, 3), stride=1, padding=1), 22 | nn.ELU()) 23 | self.decoder = nn.Sequential( 24 | nn.ConvTranspose2d(in_channels=256, out_channels=128, kernel_size=(3, 3), stride=1, padding=1, output_padding=0), 25 | nn.ELU(), 26 | nn.ConvTranspose2d(in_channels=128, out_channels=64, kernel_size=(3, 3), stride=2, padding=1, output_padding=1), 27 | nn.ELU(), 28 | nn.ConvTranspose2d(in_channels=64, out_channels=64, kernel_size=(3, 3), stride=1, padding=1, output_padding=0), 29 | nn.ELU(), 30 | nn.ConvTranspose2d(in_channels=64, out_channels=1, kernel_size=(3, 3), stride=2, padding=1, output_padding=1)) 31 | 32 | if args.dataset == 'kth': 33 | self.decoder.add_module("last_activation", nn.Sigmoid()) 34 | 35 | self.convlstm_num = 4 36 | self.convlstm_in_c = [128, 128, 128, 128] 37 | self.convlstm_out_c = [128, 128, 128, 128] 38 | self.convlstm_list = [] 39 | for layer_i in range(self.convlstm_num): 40 | self.convlstm_list.append(convlstm.NPUnit(in_channels=self.convlstm_in_c[layer_i], 41 | out_channels=self.convlstm_out_c[layer_i], 42 | kernel_size=[3, 3])) 43 | self.convlstm_list = nn.ModuleList(self.convlstm_list) 44 | 45 | self.memory = Memory(args.memory_size) 46 | 47 | self.attention_size = 128 48 | self.attention_func = nn.Sequential( 49 | nn.AdaptiveAvgPool2d([1, 1]), 50 | nn.Flatten(), 51 | nn.Linear(256, 16), 52 | nn.ReLU(), 53 | nn.Linear(16, self.attention_size), 54 | nn.Sigmoid()) 55 | 56 | def forward(self, short_x, long_x, out_len, phase): 57 | batch_size = short_x.size()[0] 58 | input_len= short_x.size()[1] 59 | 60 | # long-term motion context recall 61 | memory_x = long_x if phase == 1 else short_x 62 | memory_feature = self.memory(memory_x, phase) 63 | 64 | # motion context-aware video prediction 65 | h, c, out_pred = [], [], [] 66 | for layer_i in range(self.convlstm_num): 67 | zero_state = torch.zeros(batch_size, self.convlstm_in_c[layer_i], memory_feature.size()[2], memory_feature.size()[3]).to(self.device) 68 | h.append(zero_state) 69 | c.append(zero_state) 70 | for seq_i in range(input_len+out_len-1): 71 | if seq_i < input_len: 72 | input_x = short_x[:, seq_i, :, :, :] 73 | input_x = self.encoder(input_x) 74 | else: 75 | input_x = self.encoder(out_pred[-1]) 76 | 77 | for layer_i in range(self.convlstm_num): 78 | if layer_i == 0: 79 | h[layer_i], c[layer_i] = self.convlstm_list[layer_i](input_x, h[layer_i], c[layer_i]) 80 | else: 81 | h[layer_i], c[layer_i] = self.convlstm_list[layer_i](h[layer_i-1], h[layer_i], c[layer_i]) 82 | 83 | if seq_i >= input_len-1: 84 | attention = self.attention_func(torch.cat([c[-1], memory_feature], dim=1)) 85 | attention = torch.reshape(attention, (-1, self.attention_size, 1, 1)) 86 | memory_feature_att = memory_feature * attention 87 | out_pred.append(self.decoder(torch.cat([h[-1], memory_feature_att], dim=1))) 88 | 89 | out_pred = torch.stack(out_pred) 90 | out_pred = out_pred.transpose(dim0=0, dim1=1) 91 | out_pred = out_pred[:, -out_len:, :, :, :] 92 | 93 | return out_pred 94 | 95 | 96 | class Memory(nn.Module): 97 | def __init__(self, memory_size): 98 | super(Memory, self).__init__() 99 | self.motion_matching_encoder = nn.Sequential( 100 | nn.Conv3d(1, 64, kernel_size=(3, 3, 3), padding=(1, 1, 1)), 101 | nn.ReLU(), 102 | nn.MaxPool3d(kernel_size=(1, 2, 2), stride=(1, 2, 2)), 103 | nn.Conv3d(64, 128, kernel_size=(3, 3, 3), padding=(1, 1, 1)), 104 | nn.ReLU(), 105 | nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2)), 106 | nn.Conv3d(128, 256, kernel_size=(3, 3, 3), padding=(1, 1, 1)), 107 | nn.ReLU(), 108 | nn.Conv3d(256, 256, kernel_size=(3, 3, 3), padding=(1, 1, 1)), 109 | nn.ReLU(), 110 | nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2)), 111 | nn.Conv3d(256, 512, kernel_size=(3, 3, 3), padding=(1, 1, 1)), 112 | nn.ReLU(), 113 | nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1, 1, 1)), 114 | nn.ReLU(), 115 | nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2)), 116 | nn.AdaptiveAvgPool3d([1, None, None])) 117 | 118 | self.motion_context_encoder = copy.deepcopy(self.motion_matching_encoder) 119 | 120 | self.embedder = nn.Sequential( 121 | nn.ConvTranspose2d(in_channels=512, out_channels=256, kernel_size=(3, 3), stride=2, padding=1, output_padding=1), 122 | nn.ReLU(), 123 | nn.ConvTranspose2d(in_channels=256, out_channels=128, kernel_size=(3, 3), stride=2, padding=1, output_padding=1), 124 | nn.ReLU()) 125 | 126 | self.memory_shape = [memory_size, 512] 127 | self.memory_w = nn.init.normal_(torch.empty(self.memory_shape), mean=0.0, std=1.0) 128 | self.memory_w = nn.Parameter(self.memory_w, requires_grad=True) 129 | 130 | def forward(self, memory_x, phase): 131 | memory_x = memory_x[:, 1:, :, :, :] - memory_x[:, :-1, :, :, :] # make difference frames 132 | 133 | memory_x = memory_x.transpose(dim0=1, dim1=2) # make (N, C, T, H, W) for 3D Conv 134 | motion_encoder = self.motion_context_encoder if phase == 1 else self.motion_matching_encoder 135 | memory_query = torch.squeeze(motion_encoder(memory_x), dim=2) # make (N, C, H, W) 136 | 137 | query_c, query_h, query_w = memory_query.size()[1], memory_query.size()[2], memory_query.size()[3] 138 | memory_query = memory_query.permute(0, 2, 3, 1) # make (N, H, W, C) 139 | memory_query = torch.reshape(memory_query, (-1, query_c)) # make (N*H*W, C) 140 | 141 | # memory addressing 142 | query_norm = F.normalize(memory_query, dim=1) 143 | memory_norm = F.normalize(self.memory_w, dim=1) 144 | s = torch.mm(query_norm, memory_norm.transpose(dim0=0, dim1=1)) 145 | addressing_vec = F.softmax(s, dim=1) 146 | memory_feature = torch.mm(addressing_vec, self.memory_w) 147 | 148 | memory_feature = torch.reshape(memory_feature, (-1, query_h, query_w, query_c)) # make (N, H, W, C) 149 | memory_feature = memory_feature.permute(0, 3, 1, 2) # make (N, C, H, W) for 2D DeConv 150 | memory_feature = self.embedder(memory_feature) 151 | 152 | return memory_feature 153 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | torch==1.6.0 2 | opencv-python 3 | scikit-image 4 | lpips 5 | numpy 6 | -------------------------------------------------------------------------------- /test.py: -------------------------------------------------------------------------------- 1 | from model import Predictor 2 | from dataloader import MovingMNIST, KTH 3 | from utils import * 4 | 5 | import torch 6 | import torch.nn as nn 7 | from torch.utils.data import DataLoader 8 | 9 | import lpips 10 | import argparse 11 | import numpy as np 12 | import os 13 | import cv2 14 | 15 | 16 | parser = argparse.ArgumentParser() 17 | parser.add_argument('--dataset', type=str, default='movingmnist', 18 | help='testing dataset (movingmnist or kth)') 19 | parser.add_argument('--workers', type=int, default=4, 20 | help='number of data loading workers') 21 | parser.add_argument('--make_frame', type=bool, default=True, 22 | help='whether to generate predicted frames') 23 | parser.add_argument('--evaluate', type=bool, default=False, 24 | help='whether to evaluate performance') 25 | parser.add_argument('--test_data_dir', type=str, default='enter_the_path', 26 | help='directory of test set') 27 | parser.add_argument('--test_result_dir', type=str, default='./test_results', 28 | help='directory for saving predicted frames') 29 | parser.add_argument('--checkpoint_load_file', type=str, default='enter_the_path', 30 | help='file path for loading checkpoint') 31 | 32 | parser.add_argument('--img_size', type=int, default=64, 33 | help='height and width of video frame') 34 | parser.add_argument('--img_channel', type=int, default=1, 35 | help='channel of video frame') 36 | parser.add_argument('--memory_size', type=int, default=100, 37 | help='memory slot size') 38 | parser.add_argument('--short_len', type=int, default=10, 39 | help='number of input short-term frames') 40 | parser.add_argument('--out_len', type=int, default=30, 41 | help='number of output predicted frames') 42 | 43 | parser.add_argument('--batch_size', type=int, default=8, 44 | help='mini-batch size') 45 | args = parser.parse_args() 46 | 47 | if __name__ == '__main__': 48 | if not os.path.isdir(args.test_result_dir): 49 | os.makedirs(args.test_result_dir) 50 | 51 | # define the model 52 | device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') 53 | pred_model = Predictor(args).to(device) 54 | pred_model = nn.DataParallel(pred_model) 55 | 56 | # load checkpoint 57 | pred_model.load_state_dict(torch.load(args.checkpoint_load_file)) 58 | print('Checkpoint is loaded from ' + args.checkpoint_load_file) 59 | 60 | # prepare dataloader for selected dataset 61 | if args.dataset == 'movingmnist': 62 | test_dataset = MovingMNIST(args.test_data_dir, seq_len=args.short_len+args.out_len, train=False) 63 | testloader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, drop_last=False) 64 | elif args.dataset == 'kth': 65 | test_dataset = KTH(args.test_data_dir, seq_len=args.short_len+args.out_len, train=False) 66 | testloader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, drop_last=False) 67 | 68 | clips = testloader.sampler.data_source.clips 69 | lpips_dist = lpips.LPIPS(net = 'alex').to(device) 70 | valid_mse, valid_psnr, valid_ssim, valid_lpips = AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter() 71 | 72 | print('Start testing...') 73 | pred_model.eval() 74 | with torch.no_grad(): 75 | for test_i, test_data in enumerate(testloader): 76 | # define data indexes 77 | short_data_start, short_data_end = 0, args.short_len 78 | out_gt_start, out_gt_end = short_data_end, short_data_end+args.out_len 79 | 80 | # obtain input data and output gt 81 | test_data = torch.stack(test_data).to(device) 82 | test_data = test_data.transpose(dim0=0, dim1=1) 83 | short_data = test_data[:, short_data_start:short_data_end, :, :, :] 84 | out_gt = test_data[:, out_gt_start:out_gt_end, :, :, :] 85 | 86 | # frame prediction 87 | out_pred = pred_model(short_data, None, args.out_len, phase=2) 88 | out_pred = torch.clamp(out_pred, min = 0, max = 1) 89 | 90 | # calculate evaluation metrics 91 | batch_size_current = test_data.shape[0] 92 | if args.evaluate: 93 | mse, psnr, ssim, lpips = calculate_metrics(out_pred, out_gt, lpips_dist, args) 94 | valid_mse.update(np.mean(mse), batch_size_current) 95 | valid_psnr.update(np.mean(psnr), batch_size_current) 96 | valid_ssim.update(np.mean(ssim), batch_size_current) 97 | valid_lpips.update(np.mean(lpips), batch_size_current) 98 | 99 | # generate predicted frames 100 | if args.make_frame: 101 | for batch_i in range(batch_size_current): 102 | video_i, frame_start = clips[test_i*args.batch_size+batch_i] 103 | if not os.path.isdir(args.test_result_dir + '/video_'+ str(video_i)+'_' + str(frame_start)): 104 | os.makedirs(args.test_result_dir + '/video_'+ str(video_i)+'_' + str(frame_start)) 105 | for frame_i in range(args.short_len): 106 | cv2.imwrite(args.test_result_dir + '/video_'+ str(video_i)+'_' + str(frame_start)+ '/input_' 107 | +str(frame_i).zfill(5) + '.jpg', short_data[batch_i,frame_i,0,:,:].cpu().numpy()*255) 108 | for frame_i in range(args.out_len): 109 | cv2.imwrite(args.test_result_dir+'/video_'+str(video_i)+'_'+str(frame_start)+'/pred_'+ 110 | str(frame_i+args.short_len).zfill(5)+'.jpg', out_pred[batch_i,frame_i,0,:,:].cpu().numpy()*255) 111 | 112 | if args.evaluate: 113 | print('************** test_output_length [{}] **************' 114 | .format(args.out_len)) 115 | print('mse: {:.3f}, psnr: {:.3f}, ssim: {:.3f}, lpips: {:.3f}' 116 | .format(valid_mse.avg, valid_psnr.avg, valid_ssim.avg, valid_lpips.avg)) -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | from model import Predictor 2 | from dataloader import MovingMNIST, KTH 3 | from utils import * 4 | 5 | import torch 6 | import torch.nn as nn 7 | from torch.utils.data import DataLoader 8 | 9 | import lpips 10 | import argparse 11 | import numpy as np 12 | import time 13 | import os 14 | 15 | 16 | seed = 1234 17 | np.random.seed(seed) 18 | torch.manual_seed(seed) 19 | torch.cuda.manual_seed(seed) 20 | torch.cuda.manual_seed_all(seed) 21 | 22 | parser = argparse.ArgumentParser() 23 | parser.add_argument('--dataset', type=str, default='movingmnist', 24 | help='training dataset (movingmnist or kth)') 25 | parser.add_argument('--workers', type=int, default=4, 26 | help='number of data loading workers') 27 | parser.add_argument('--train_data_dir', type=str, default='enter_the_path', 28 | help='directory of training set') 29 | parser.add_argument('--valid_data_dir', type=str, default='enter_the_path', 30 | help='directory of validation set') 31 | parser.add_argument('--checkpoint_load', type=bool, default=False, 32 | help='whether to load checkpoint') 33 | parser.add_argument('--checkpoint_load_file', type=str, default='enter_the_path', 34 | help='file path for loading checkpoint') 35 | parser.add_argument('--checkpoint_save_dir', type=str, default='./checkpoints', 36 | help='directory for saving checkpoints') 37 | 38 | parser.add_argument('--img_size', type=int, default=64, 39 | help='height and width of video frame') 40 | parser.add_argument('--img_channel', type=int, default=1, 41 | help='channel of video frame') 42 | parser.add_argument('--memory_size', type=int, default=100, 43 | help='memory slot size') 44 | parser.add_argument('--short_len', type=int, default=10, 45 | help='number of input short-term frames') 46 | parser.add_argument('--long_len', type=int, default=30, 47 | help='number of input long-term frames') 48 | parser.add_argument('--out_len', type=int, default=30, 49 | help='number of output predicted frames') 50 | 51 | parser.add_argument('--batch_size', type=int, default=128, 52 | help='mini-batch size') 53 | parser.add_argument('--lr', type=float, default=0.0002, 54 | help='learning rate') 55 | parser.add_argument('--iterations', type=int, default=300000, 56 | help='number of total iterations') 57 | parser.add_argument('--iterations_warmup', type=int, default=5000, 58 | help='number of iterations for warming up model') 59 | parser.add_argument('--print_freq', type=int, default=1000, 60 | help='frequency of printing logs') 61 | args = parser.parse_args() 62 | 63 | 64 | if __name__ == '__main__': 65 | if not os.path.isdir(args.checkpoint_save_dir): 66 | os.makedirs(args.checkpoint_save_dir) 67 | 68 | # define the model 69 | device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') 70 | pred_model = Predictor(args).to(device) 71 | pred_model = nn.DataParallel(pred_model) 72 | 73 | # optionally load checkpoint 74 | if args.checkpoint_load: 75 | pred_model.load_state_dict(torch.load(args.checkpoint_load_file)) 76 | print('Checkpoint is loaded from ' + args.checkpoint_load_file) 77 | 78 | # prepare dataloader for selected dataset 79 | if args.dataset == 'movingmnist': 80 | train_dataset = MovingMNIST(args.train_data_dir, seq_len=args.short_len+args.out_len, train=True) 81 | trainloader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) 82 | valid_dataset = MovingMNIST(args.valid_data_dir, seq_len=args.short_len+args.out_len, train=False) 83 | validloader = DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, drop_last=False) 84 | elif args.dataset == 'kth': 85 | train_dataset = KTH(args.train_data_dir, seq_len=args.short_len+args.out_len, train=True) 86 | trainloader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) 87 | valid_dataset = KTH(args.valid_data_dir, seq_len=args.short_len+args.out_len, train=False) 88 | validloader = DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, drop_last=False) 89 | 90 | # define optimizer and loss function 91 | optimizer = torch.optim.Adam(pred_model.parameters(), lr=args.lr) 92 | l1_loss, l2_loss = nn.L1Loss().to(device), nn.MSELoss().to(device) 93 | lpips_dist = lpips.LPIPS(net = 'alex').to(device) 94 | 95 | mse_min, psnr_max, ssim_max, lpips_min = 99999, 0, 0, 99999 96 | train_loss = AverageMeter() 97 | valid_mse, valid_psnr, valid_ssim, valid_lpips = AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter() 98 | 99 | print('Start training...') 100 | start_time = time.time() 101 | data_iterator = iter(trainloader) 102 | for train_i in range(args.iterations): 103 | try: 104 | train_data = next(data_iterator) 105 | except: 106 | data_iterator = iter(trainloader) 107 | train_data = next(data_iterator) 108 | 109 | # define data indexes 110 | short_start, short_end = 0, args.short_len 111 | long_start = np.random.randint(0, args.short_len+args.out_len-args.long_len+1) 112 | long_end = long_start+args.long_len 113 | out_gt_start, out_gt_end = short_end, short_end+args.out_len 114 | 115 | # obtain input data and output gt 116 | train_data = torch.stack(train_data).to(device) 117 | train_data = train_data.transpose(dim0=0, dim1=1) # make (N, T, C, H, W) 118 | short_data = train_data[:, short_start:short_end, :, :, :] 119 | long_data = train_data[:, long_start:long_end, :, :, :] 120 | out_gt = train_data[:, out_gt_start:out_gt_end, :, :, :] 121 | 122 | # predict only 10 frames in the first few iterations to warm up the model 123 | if (not args.checkpoint_load) and (train_i < args.iterations_warmup): 124 | train_out_len = 10 125 | long_data = train_data[:, short_start:out_gt_start+train_out_len, :, :, :] 126 | out_gt = train_data[:, out_gt_start:out_gt_start+train_out_len, :, :, :] 127 | else: 128 | train_out_len = args.out_len 129 | 130 | pred_model.train() 131 | 132 | # training phase 1 with long-term sequence 133 | pred_model.module.memory.memory_w.requires_grad = True # train memory weights 134 | out_pred = pred_model(short_data, long_data, train_out_len, phase=1) 135 | loss_p1 = l1_loss(out_pred, out_gt) + l2_loss(out_pred, out_gt) 136 | optimizer.zero_grad() 137 | loss_p1.backward() 138 | optimizer.step() 139 | # training phase 2 without long-term sequence 140 | pred_model.module.memory.memory_w.requires_grad = False # do not train memory weights 141 | out_pred = pred_model(short_data, None, train_out_len, phase=2) 142 | loss_p2 = l1_loss(out_pred, out_gt) + l2_loss(out_pred, out_gt) 143 | optimizer.zero_grad() 144 | loss_p2.backward() 145 | optimizer.step() 146 | 147 | train_loss.update(float(loss_p1) +float(loss_p2)) 148 | 149 | if (train_i+1) % args.print_freq == 0: 150 | torch.save(pred_model.state_dict(), args.checkpoint_save_dir+'/trained_file_'+str(train_i+1).zfill(6)+'.pt') 151 | 152 | # validation phase 153 | pred_model.eval() 154 | with torch.no_grad(): 155 | for valid_data in validloader: 156 | # define data indexes 157 | short_start, short_end = 0, args.short_len 158 | out_gt_start, out_gt_end = short_end, short_end+args.out_len 159 | 160 | # obtain input data and output gt 161 | valid_data = torch.stack(valid_data).to(device) 162 | valid_data = valid_data.transpose(dim0=0, dim1=1) # make (N, T, C, H, W) 163 | short_data = valid_data[:, short_start:short_end, :, :, :] 164 | out_gt = valid_data[:, out_gt_start:out_gt_end, :, :, :] 165 | 166 | # frame prediction and calculate evaluation metrics 167 | out_pred = pred_model(short_data, None, args.out_len, phase=2) 168 | out_pred = torch.clamp(out_pred, min = 0, max = 1) 169 | mse, psnr, ssim, lpips = calculate_metrics(out_pred, out_gt, lpips_dist, args) 170 | 171 | batch_size_current = valid_data.shape[0] 172 | valid_mse.update(np.mean(mse), batch_size_current) 173 | valid_psnr.update(np.mean(psnr), batch_size_current) 174 | valid_ssim.update(np.mean(ssim), batch_size_current) 175 | valid_lpips.update(np.mean(lpips), batch_size_current) 176 | 177 | mse_min = valid_mse.avg if valid_mse.avg < mse_min else mse_min 178 | psnr_max = valid_psnr.avg if valid_psnr.avg > psnr_max else psnr_max 179 | ssim_max = valid_ssim.avg if valid_ssim.avg > ssim_max else ssim_max 180 | lpips_min = valid_lpips.avg if valid_lpips.avg < lpips_min else lpips_min 181 | 182 | elapsed_time = time.time() - start_time; start_time = time.time() 183 | print('******** iter [{}] / epoch [{:.4f}] / loss [{:.4f}] ********' 184 | .format(train_i+1, (train_i+1)/len(trainloader), train_loss.avg)) 185 | print('[current] mse: {:.3f}, psnr: {:.3f}, ssim: {:.3f}, lpips: {:.3f}' 186 | .format(valid_mse.avg, valid_psnr.avg, valid_ssim.avg, valid_lpips.avg)) 187 | print('[ best ] mse: {:.3f}, psnr: {:.3f}, ssim: {:.3f}, lpips: {:.3f}' 188 | .format(mse_min, psnr_max, ssim_max, lpips_min)) 189 | print('elapsed time: {:.0f} sec'.format(elapsed_time)) 190 | train_loss.reset(); valid_mse.reset(); valid_psnr.reset(); valid_ssim.reset(); valid_lpips.reset() 191 | -------------------------------------------------------------------------------- /utils.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | import numpy as np 4 | from skimage.metrics import peak_signal_noise_ratio 5 | from skimage.metrics import structural_similarity 6 | from skimage.metrics import mean_squared_error 7 | 8 | 9 | class AverageMeter(object): 10 | """Computes and stores the average and current value""" 11 | def __init__(self): 12 | self.reset() 13 | 14 | def reset(self): 15 | self.val = 0 16 | self.avg = 0 17 | self.sum = 0 18 | self.count = 0 19 | 20 | def update(self, val, n=1): 21 | self.val = val 22 | self.sum += val * n 23 | self.count += n 24 | self.avg = self.sum / self.count 25 | 26 | 27 | def calculate_metrics(pred, gt, lpips_dist, args): 28 | batch_size = pred.shape[0] 29 | multi_channel = True if args.img_channel > 1 else False 30 | 31 | pred = pred.cpu().numpy() 32 | pred = np.transpose(pred, [0,1,3,4,2]) 33 | gt = gt.cpu().numpy() 34 | gt = np.transpose(gt, [0,1,3,4,2]) 35 | 36 | mse_mean = np.zeros(args.out_len, dtype=pred.dtype) 37 | psnr_mean = np.zeros(args.out_len, dtype=pred.dtype) 38 | ssim_mean = np.zeros(args.out_len, dtype=pred.dtype) 39 | lpips_mean = np.zeros(args.out_len, dtype=pred.dtype) 40 | gt = gt.astype(dtype=pred.dtype) 41 | 42 | for frame_i in range(-args.out_len, 0): 43 | for batch_i in range(batch_size): 44 | gt_frame = gt[batch_i,frame_i,:,:,:] 45 | pred_frame = pred[batch_i,frame_i,:,:,:] 46 | if args.img_channel == 1: 47 | gt_frame = np.squeeze(gt_frame) 48 | pred_frame = np.squeeze(pred_frame) 49 | mse_mean[frame_i] += mean_squared_error(gt_frame, pred_frame)/batch_size 50 | psnr_mean[frame_i] += peak_signal_noise_ratio(gt_frame, pred_frame)/batch_size 51 | ssim_mean[frame_i] += structural_similarity(gt_frame, pred_frame, multichannel=multi_channel)/batch_size 52 | 53 | batch_gt = gt[:,frame_i,:,:,:] 54 | batch_pred = pred[:,frame_i,:,:,:] 55 | batch_gt = np.transpose(batch_gt, [0,3,1,2]) 56 | batch_pred = np.transpose(batch_pred, [0,3,1,2]) 57 | if args.img_channel == 1: 58 | batch_gt = np.repeat(batch_gt, 3, axis=1) 59 | batch_pred = np.repeat(batch_pred, 3, axis=1) 60 | batch_gt = torch.from_numpy(batch_gt).float().to('cuda') 61 | batch_pred = torch.from_numpy(batch_pred).float().to('cuda') 62 | lpips_mean[frame_i] += np.mean(lpips_dist(batch_gt, batch_pred).cpu().numpy()) 63 | 64 | mse_mean *= args.img_size**2 65 | 66 | return mse_mean, psnr_mean, ssim_mean, lpips_mean 67 | --------------------------------------------------------------------------------