├── INPLACE_ABN_TIPS.md ├── LICENSE ├── MODEL_ZOO.md ├── README.md ├── figures ├── herbarium_2020.png ├── main_pic.png ├── ms_coco_scores.png ├── sotabench.png ├── table_1.png ├── table_3.png ├── table_4.png ├── table_5.png └── table_6.png ├── infer.py ├── requirements.txt ├── src ├── helper_functions │ └── helper_functions.py └── models │ ├── __init__.py │ ├── tresnet │ ├── __init__.py │ ├── layers │ │ ├── anti_aliasing.py │ │ ├── avg_pool.py │ │ ├── space_to_depth.py │ │ └── squeeze_and_excite.py │ └── tresnet.py │ ├── tresnet_v2 │ ├── __init__.py │ ├── layers │ │ ├── anti_aliasing.py │ │ ├── avg_pool.py │ │ ├── space_to_depth.py │ │ └── squeeze_and_excite.py │ └── tresnet_v2.py │ └── utils │ ├── __init__.py │ └── factory.py └── tests └── test_TResNetV2 /INPLACE_ABN_TIPS.md: -------------------------------------------------------------------------------- 1 | ## Some Tips For Working With Inplace-ABN 2 | 3 | Inplace batch-norm 4 | ([Inplace-ABN](https://github.com/mapillary/inplace_abn)) module has 5 | exactly the same fields as 6 | [regular BatchNorm](https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/batchnorm.py): 7 | * module.weight 8 | * module.bias 9 | * module.running_mean 10 | * module.running_var 11 | * module.num_batches_tracked 12 | 13 | Therefore, any function that operates on BatchNorm can run on 14 | Inplace-ABN. 15 | 16 | However, problems can arise when a logic condition seeks explicitly for 17 | BatchNorm layers only: 18 | ``` 19 | if isinstance(module, nn.BatchNorm2d): 20 | do_something(module) 21 | ``` 22 | 23 | Anywhere you see a code segement like this, it needs to be replaced with 24 | a condition that includes Inplace-ABN: 25 | ``` 26 | if isinstance(module, nn.BatchNorm2d) or isinstance(module, inplace_abn.InPlaceABN): 27 | do_something(module) 28 | ``` 29 | 30 | ##### NVIDIA Apex mixed precision 31 | [NVIDIA Apex](https://github.com/NVIDIA/apex) O0, O1 and O3 32 | mixed-precision options work seemlesly on Inplace-ABN. 33 | 34 | For O2 mixed precision, we need to convert manually Inplace-ABN to fp32, 35 | since NVIDIA Apex inner code has explicit 'if' condition for BatchNorm 36 | only. 37 | 38 | Conversion can be done easily with the helper function 39 | 'IABN2float': 40 | ``` 41 | if args.use_apex: 42 | model, optimizer = apex.amp.initialize(model, optimizer, opt_level=args.opt_level) 43 | if args.opt_level == 'O2': # IABN needs adjustment for O2 44 | from src.models.tresnet import IABN2Float 45 | model = IABN2Float(model) 46 | ``` -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /MODEL_ZOO.md: -------------------------------------------------------------------------------- 1 | # TResNet pre-trained models 2 | 3 | We provide a collection of TResNet models pre-trained on [ImageNet dataset](http://www.image-net.org/). 4 | 5 | | Model name | Input Size 6 | | ------------ | :--------------: | 7 | | [tresnet_m.pth](https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/tresnet/tresnet_m.pth) | 224 | 8 | | [tresnet_m_448.pth](https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/tresnet/tresnet_m_448.pth) | 448 | 9 | | [tresnet_l.pth](https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/tresnet/tresnet_l.pth) | 224 | 10 | | [tresnet_l_448.pth](https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/tresnet/tresnet_l_448.pth) | 448 | 11 | | [tresnet_xl.pth](https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/tresnet/tresnet_xl.pth) | 224 | 12 | | [tresnet_xl_448.pth](https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/tresnet/tresnet_xl_448.pth) | 448 | 13 | 14 | We also provide The TResNet model that achieved SOTA results on 15 | [stanford-cars](https://paperswithcode.com/sota/fine-grained-image-classification-on-stanford) 16 | dataset: 17 | 18 | | Model name | Input Size 19 | | ------------ | :--------------: | 20 | | [tresnet_l-v2.pth](https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/tresnet/stanford_cars_tresnet-l-v2_96_27.pth) | 368 | 21 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # TResNet: High Performance GPU-Dedicated Architecture 2 | 3 | [paperV2](https://arxiv.org/pdf/2003.13630.pdf) | 4 | [pretrained models](MODEL_ZOO.md) 5 | 6 | Official PyTorch Implementation 7 | 8 | > Tal Ridnik, Hussam Lawen, Asaf Noy, Itamar Friedman, Emanuel Ben Baruch, Gilad Sharir
9 | > DAMO Academy, Alibaba Group 10 | 11 | 12 | 13 | **Abstract** 14 | 15 | > Many deep learning models, developed in recent years, reach higher 16 | > ImageNet accuracy than ResNet50, with fewer or comparable FLOPS count. 17 | > While FLOPs are often seen as a proxy for network efficiency, when 18 | > measuring actual GPU training and inference throughput, vanilla 19 | > ResNet50 is usually significantly faster than its recent competitors, 20 | > offering better throughput-accuracy trade-off. In this work, we 21 | > introduce a series of architecture modifications that aim to boost 22 | > neural networks' accuracy, while retaining their GPU training and 23 | > inference efficiency. We first demonstrate and discuss the bottlenecks 24 | > induced by FLOPs-optimizations. We then suggest alternative designs 25 | > that better utilize GPU structure and assets. Finally, we introduce a 26 | > new family of GPU-dedicated models, called TResNet, which achieve 27 | > better accuracy and efficiency than previous ConvNets. Using a TResNet 28 | > model, with similar GPU throughput to ResNet50, we reach 80.7\% 29 | > top-1 accuracy on ImageNet. Our TResNet models also transfer well and 30 | > achieve state-of-the-art accuracy on competitive datasets such as 31 | > Stanford cars (96.0\%), CIFAR-10 (99.0\%), CIFAR-100 (91.5\%) and 32 | > Oxford-Flowers (99.1\%). They also perform well on multi-label classification and object detection tasks. 33 | 34 | ## 29/11/2021 Update - New article released, offering new classification head with state-of-the-art results 35 | Checkout our new project, [Ml-Decoder](https://github.com/Alibaba-MIIL/ML_Decoder), which presents a unified classification head for multi-label, single-label and 36 | zero-shot tasks. Backbones with ML-Decoder reach SOTA results, while also improving speed-accuracy tradeoff. 37 | 38 |

39 | 40 | 41 | 42 | 43 | 44 | 45 |
46 |

47 | 48 | ## 11/1/2023 Update 49 | Added [tests](https://github.com/Alibaba-MIIL/TResNet/blob/master/tests/test_TResNetV2) auto-generated by [CodiumAI](https://www.codium.ai/) tool 50 | 51 | 52 | ## 23/4/2021 Update - ImageNet21K Pretraining 53 | In a new [article](https://github.com/Alibaba-MIIL/ImageNet21K) we released, we share pretrain weights for TResNet models from ImageNet21K training, that dramatically outperfrom standard pretraining. 54 | TResNet-M model, for example, improves its ImageNet-1K score, from 80.7% to 83.1% ! 55 | This kind of improvement is consistently achieved on all downstream tasks. 56 | 57 | 58 | ## 28/8/2020: V2 of TResNet Article Released 59 | 60 | ## Sotabench Comparisons 61 | Comparative results from 62 | [sotabench benchamrk](https://sotabench.com/benchmarks/image-classification-on-imagenet#code), 63 | demonstartaing that TReNset models give excellent speed-accuracy tradoff: 64 |

65 | 66 | 67 | 68 | 69 |
70 |

71 | 72 | ## 11/6/2020: V1 of TResNet Article Released 73 | The main change - In addition to single label SOTA results, we also 74 | added top results for multi-label classification and object detection 75 | tasks, using TResNet. For example, we set a new SOTA record for MS-COCO 76 | multi-label dataset, surpassing the previous top results by more than 77 | 2.5% mAP ! 78 |

79 | 80 | 81 | 82 | 83 | 84 | 85 | 86 | 87 | 88 | 89 | 90 | 91 | 92 | 93 |
BacbkonemAP
KSSNet (previous SOTA)83.7
TResNet-L86.4
94 |

95 | 96 | 97 | ## 2/6/2020: CVPR-Kaggle competitions 98 | We participated and won top places in two 99 | major CVPR-Kaggle competitions: 100 | * [2nd place](https://www.kaggle.com/c/herbarium-2020-fgvc7/discussion/154186) 101 | in Herbarium 2020 competition, out of 153 teams. 102 | * [7th place](https://www.kaggle.com/c/plant-pathology-2020-fgvc7/discussion/154086) 103 | in Plant-Pathology 2020 competition, out of 1317 teams. 104 |
*TResNet* was a vital part of our solution for both competitions, 105 | allowing us to work on high resolutions and reach top scores while 106 | doing fast and efficient experiments. 107 |
108 | 109 | 110 | 111 | 112 |
113 |
114 | 115 | 116 | ## Main Article Results 117 | #### TResNet Models 118 | TResNet models accuracy and GPU throughput on ImageNet, compared to ResNet50. All measurements were done on Nvidia V100 GPU, with mixed precision. All models are trained on input resolution of 224. 119 |

120 | 121 | 122 | 123 | 124 | 125 | 126 | 127 | 128 | 129 | 130 | 131 | 132 | 133 | 134 | 135 | 136 | 137 | 138 | 139 | 140 | 141 | 142 | 143 | 144 | 145 | 146 | 147 | 148 | 149 | 150 | 151 | 152 | 153 | 154 | 155 | 156 | 157 | 158 | 159 | 160 | 161 | 162 | 163 |
ModelsTop Training Speed
(img/sec)		Top Inference Speed
(img/sec)		Max Train Batch SizeTop-1 Acc.
ResNet50805283028879.0
EfficientNetB1440274019679.2
TResNet-M730293051280.8
TResNet-L345139031681.5
TResNet-XL250106024082.0
164 |

165 | 166 | #### Comparison To Other Networks 167 | 168 | Comparison of ResNet50 to top modern networks, with similar top-1 ImageNet accuracy. 169 | All measurements were done on Nvidia V100 GPU with mixed precision. For gaining optimal speeds, training and inference were measured on 90\% of maximal possible batch size. 170 | Except TResNet-M, all the models' ImageNet scores were taken from the [public repository](https://github.com/rwightman/pytorch-image-models), which specialized in providing top implementations for modern networks. Except EfficientNet-B1, which has input resolution of 240, all other models have input resolution of 224. 171 |

172 | 173 | 174 | 175 | 176 | 177 | 178 | 179 | 180 | 181 | 182 | 183 | 184 | 185 | 186 | 187 | 188 | 189 | 190 | 191 | 192 | 193 | 194 | 195 | 196 | 197 | 198 | 199 | 200 | 201 | 202 | 203 | 204 | 205 | 206 | 207 | 208 | 209 | 210 | 211 | 212 | 213 | 214 | 215 | 216 | 217 | 218 | 219 | 220 | 221 | 222 | 223 | 224 | 225 | 226 | 227 | 228 | 229 |
ModelTop Training Speed
(img/sec)		Top Inference Speed
(img/sec)		Top-1 Acc.Flops[G]
ResNet50805283079.04.1
ResNet50-D600267079.34.4
ResNeXt50490194079.44.3
EfficientNetB1440274079.20.6
SEResNeXt50400177079.94.3
MixNet-L400140079.00.5
TResNet-M730293080.85.5
230 |

231 | 232 |
233 |

234 | 235 | 236 | 237 | 238 | 239 |
240 |

241 | 242 | 243 |

244 | 245 | #### Transfer Learning SotA Results 246 | Comparison of TResNet to state-of-the-art models on transfer learning datasets (only ImageNet-based transfer learning results). Models inference speed is measured on a mixed precision V100 GPU. Since no official implementation of Gpipe was provided, its inference speed is unknown 247 | 248 |

249 | 250 | 251 | 254 | 257 | 262 | 267 | 270 | 271 | 272 | 275 | 278 | 281 | 284 | 287 | 288 | 289 | 292 | 295 | 298 | 301 | 302 | 303 | 306 | 309 | 312 | 315 | 318 | 319 | 320 | 323 | 326 | 329 | 332 | 333 | 334 | 337 | 340 | 343 | 346 | 349 | 350 | 351 | 354 | 357 | 360 | 363 | 364 | 365 | 368 | 371 | 374 | 377 | 380 | 381 | 382 | 385 | 388 | 391 | 394 | 395 |
252 | Dataset 253 | 255 | Model 256 | 258 | Top-1 259 |
260 | Acc. 261 | 		263 | Speed 264 |
265 | img/sec 266 | 		268 | Input 269 |
273 | CIFAR-10 274 | 276 | Gpipe 277 | 279 | 99.0 280 | 282 | - 283 | 285 | 480 286 |
290 | TResNet-XL 291 | 293 | 99.0 294 | 296 | 1060 297 | 299 | 224 300 |
304 | CIFAR-100 305 | 307 | EfficientNet-B7 308 | 310 | 91.7 311 | 313 | 70 314 | 316 | 600 317 |
321 | TResNet-XL 322 | 324 | 91.5 325 | 327 | 1060 328 | 330 | 224 331 |
335 | Stanford Cars 336 | 338 | EfficientNet-B7 339 | 341 | 94.7 342 | 344 | 70 345 | 347 | 600 348 |
352 | TResNet-L 353 | 355 | 96.0 356 | 358 | 500 359 | 361 | 368 362 |
366 | Oxford-Flowers 367 | 369 | EfficientNet-B7 370 | 372 | 98.8 373 | 375 | 70 376 | 378 | 600 379 |
383 | TResNet-L 384 | 386 | 99.1 387 | 389 | 500 390 | 392 | 368 393 |
396 |

397 | 398 | 399 | ## Reproduce Article Scores 400 | We provide code for reproducing the validation top-1 score of TResNet 401 | models on ImageNet. First, download pretrained models from 402 | [here](MODEL_ZOO.md). 403 | 404 | Then, run the infer.py script. For example, for tresnet_m (input size 224) 405 | run: 406 | ```bash 407 | python -m infer.py \ 408 | --val_dir=/path/to/imagenet_val_folder \ 409 | --model_path=/model/path/to/tresnet_m.pth \ 410 | --model_name=tresnet_m 411 | --input_size=224 412 | ``` 413 | ## TResNet Training 414 | Due to IP limitations, we do not provide the exact training code that 415 | was used to obtain the article results. 416 | 417 | However, TResNet is now an integral part of the popular [rwightman / 418 | pytorch-image-models](https://github.com/rwightman/pytorch-image-models) 419 | repo. Using that repo, you can reach very similar results to the one 420 | stated in the article. 421 | 422 | For example, training tresnet_m on [rwightman / 423 | pytorch-image-models](https://github.com/rwightman/pytorch-image-models) with 424 | the command line: 425 | ```bash 426 | python -u -m torch.distributed.launch --nproc_per_node=8 \ 427 | --nnodes=1 --node_rank=0 ./train.py /data/imagenet/ \ 428 | -b=190 --lr=0.6 --model-ema --aa=rand-m9-mstd0.5-inc1 \ 429 | --num-gpu=8 -j=16 --amp \ 430 | --model=tresnet_m --epochs=300 --mixup=0.2 \ 431 | --sched='cosine' --reprob=0.4 --remode=pixel 432 | ``` 433 | gave accuracy of 80.5%.

434 | 435 | 436 | Also, during the merge request, we had interesting discussions and 437 | insights regarding TResNet design. I am attaching a pdf version the 438 | mentioned discussions. They can shed more light on TResNet design 439 | considerations and directions for the future. 440 | 441 | [TResNet discussion and insights](https://miil-public-eu.oss-eu-central-1.aliyuncs.com/model-zoo/tresnet/TResnet_discussion.pdf) 442 | 443 | (taken with permission from 444 | [here](https://github.com/rwightman/pytorch-image-models/issues/124)) 445 | 446 | 447 | 448 | ## Tips For Working With Inplace-ABN 449 | See 450 | [INPLACE_ABN_TIPS](https://github.com/mrT23/TResNet/blob/master/INPLACE_ABN_TIPS.md). 451 | 452 | 453 | ## Citation 454 | 455 | ``` 456 | @misc{ridnik2020tresnet, 457 | title={TResNet: High Performance GPU-Dedicated Architecture}, 458 | author={Tal Ridnik and Hussam Lawen and Asaf Noy and Itamar Friedman}, 459 | year={2020}, 460 | eprint={2003.13630}, 461 | archivePrefix={arXiv}, 462 | primaryClass={cs.CV} 463 | } 464 | ``` 465 | 466 | ## Contact 467 | Feel free to contact me if there are any questions or issues (Tal 468 | Ridnik, tal.ridnik@alibaba-inc.com). 469 | -------------------------------------------------------------------------------- /figures/herbarium_2020.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Alibaba-MIIL/TResNet/aea5f0f96d0051800379160b994781cae2ad011c/figures/herbarium_2020.png -------------------------------------------------------------------------------- /figures/main_pic.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Alibaba-MIIL/TResNet/aea5f0f96d0051800379160b994781cae2ad011c/figures/main_pic.png -------------------------------------------------------------------------------- /figures/ms_coco_scores.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Alibaba-MIIL/TResNet/aea5f0f96d0051800379160b994781cae2ad011c/figures/ms_coco_scores.png -------------------------------------------------------------------------------- /figures/sotabench.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Alibaba-MIIL/TResNet/aea5f0f96d0051800379160b994781cae2ad011c/figures/sotabench.png -------------------------------------------------------------------------------- /figures/table_1.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Alibaba-MIIL/TResNet/aea5f0f96d0051800379160b994781cae2ad011c/figures/table_1.png -------------------------------------------------------------------------------- /figures/table_3.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Alibaba-MIIL/TResNet/aea5f0f96d0051800379160b994781cae2ad011c/figures/table_3.png -------------------------------------------------------------------------------- /figures/table_4.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Alibaba-MIIL/TResNet/aea5f0f96d0051800379160b994781cae2ad011c/figures/table_4.png -------------------------------------------------------------------------------- /figures/table_5.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Alibaba-MIIL/TResNet/aea5f0f96d0051800379160b994781cae2ad011c/figures/table_5.png -------------------------------------------------------------------------------- /figures/table_6.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Alibaba-MIIL/TResNet/aea5f0f96d0051800379160b994781cae2ad011c/figures/table_6.png -------------------------------------------------------------------------------- /infer.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from src.helper_functions.helper_functions import validate, create_dataloader 3 | from src.models import create_model 4 | import argparse 5 | 6 | torch.backends.cudnn.benchmark = True 7 | 8 | parser = argparse.ArgumentParser(description='PyTorch TResNet ImageNet Inference') 9 | parser.add_argument('--val_dir') 10 | parser.add_argument('--model_path') 11 | parser.add_argument('--model_name', type=str, default='tresnet_m') 12 | parser.add_argument('--num_classes', type=int, default=1000) 13 | parser.add_argument('--input_size', type=int, default=224) 14 | parser.add_argument('--val_zoom_factor', type=int, default=0.875) 15 | parser.add_argument('--batch_size', type=int, default=48) 16 | parser.add_argument('--num_workers', type=int, default=8) 17 | parser.add_argument('--remove_aa_jit', action='store_true', default=False) 18 | 19 | 20 | def main(): 21 | # parsing args 22 | args = parser.parse_args() 23 | 24 | # setup model 25 | print('creating model...') 26 | model = create_model(args).cuda() 27 | from src.models.tresnet_v2.tresnet_v2 import InplacABN_to_ABN 28 | model2 = InplacABN_to_ABN(model) 29 | aaa = torch.jit.script(model2) 30 | state = torch.load(args.model_path, map_location='cpu')['model'] 31 | model.load_state_dict(state, strict=False) 32 | model.eval() 33 | print('done\n') 34 | 35 | # setup data loader 36 | print('creating data loader...') 37 | val_loader = create_dataloader(args) 38 | print('done\n') 39 | 40 | # actual validation process 41 | print('doing validation...') 42 | prec1_f = validate(model, val_loader) 43 | print("final top-1 validation accuracy: {:.2f}".format(prec1_f.avg)) 44 | 45 | 46 | if __name__ == '__main__': 47 | main() 48 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | torch>=1.3 2 | torchvision>=0.4.0 3 | git+https://github.com/mapillary/inplace_abn.git@v1.0.11 -------------------------------------------------------------------------------- /src/helper_functions/helper_functions.py: -------------------------------------------------------------------------------- 1 | import time 2 | import torch 3 | from torchvision.datasets import ImageFolder 4 | from torchvision.transforms import transforms 5 | 6 | 7 | def create_dataloader(args): 8 | val_bs = args.batch_size 9 | if args.input_size == 448: # squish 10 | val_tfms = transforms.Compose( 11 | [transforms.Resize((args.input_size, args.input_size))]) 12 | else: # crop 13 | val_tfms = transforms.Compose( 14 | [transforms.Resize(int(args.input_size / args.val_zoom_factor)), 15 | transforms.CenterCrop(args.input_size)]) 16 | val_tfms.transforms.append(transforms.ToTensor()) 17 | val_dataset = ImageFolder(args.val_dir, val_tfms) 18 | val_loader = torch.utils.data.DataLoader( 19 | val_dataset, batch_size=val_bs, shuffle=False, 20 | num_workers=args.num_workers, pin_memory=True, drop_last=False) 21 | return val_loader 22 | 23 | 24 | def accuracy(output, target, topk=(1,)): 25 | """Computes the precision@k for the specified values of k""" 26 | maxk = max(topk) 27 | batch_size = target.size(0) 28 | 29 | _, pred = output.topk(maxk, 1, True, True) 30 | pred = pred.t() 31 | correct = pred.eq(target.view(1, -1).expand_as(pred)) 32 | 33 | res = [] 34 | for k in topk: 35 | correct_k = correct[:k].view(-1).float().sum(0, keepdim=True) 36 | res.append(correct_k.mul_(100.0 / batch_size)) 37 | return res 38 | 39 | 40 | class AverageMeter(object): 41 | """Computes and stores the average and current value""" 42 | 43 | def __init__(self): self.reset() 44 | 45 | def reset(self): self.val = self.avg = self.sum = self.count = 0 46 | 47 | def update(self, val, n=1): 48 | self.val = val 49 | self.sum += val * n 50 | self.count += n 51 | self.avg = self.sum / self.count 52 | 53 | 54 | def validate(model, val_loader): 55 | prec1_m = AverageMeter() 56 | last_idx = len(val_loader) - 1 57 | 58 | with torch.no_grad(): 59 | for batch_idx, (input, target) in enumerate(val_loader): 60 | last_batch = batch_idx == last_idx 61 | input = input.cuda() 62 | target = target.cuda() 63 | output = model(input) 64 | 65 | prec1 = accuracy(output, target) 66 | prec1_m.update(prec1[0].item(), output.size(0)) 67 | 68 | if (last_batch or batch_idx % 100 == 0): 69 | log_name = 'ImageNet Test' 70 | print( 71 | '{0}: [{1:>4d}/{2}] ' 72 | 'Prec@1: {top1.val:>7.2f} ({top1.avg:>7.2f}) '.format( 73 | log_name, batch_idx, last_idx, 74 | top1=prec1_m)) 75 | return prec1_m 76 | -------------------------------------------------------------------------------- /src/models/__init__.py: -------------------------------------------------------------------------------- 1 | from .utils import create_model 2 | 3 | __all__ = ['create_model'] 4 | -------------------------------------------------------------------------------- /src/models/tresnet/__init__.py: -------------------------------------------------------------------------------- 1 | from .tresnet import TResnetM, TResnetL, TResnetXL 2 | -------------------------------------------------------------------------------- /src/models/tresnet/layers/anti_aliasing.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn.parallel 3 | import numpy as np 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | 7 | 8 | class AntiAliasDownsampleLayer(nn.Module): 9 | def __init__(self, remove_aa_jit: bool = False, filt_size: int = 3, stride: int = 2, 10 | channels: int = 0): 11 | super(AntiAliasDownsampleLayer, self).__init__() 12 | if not remove_aa_jit: 13 | self.op = DownsampleJIT(filt_size, stride, channels) 14 | else: 15 | self.op = Downsample(filt_size, stride, channels) 16 | 17 | def forward(self, x): 18 | return self.op(x) 19 | 20 | 21 | @torch.jit.script 22 | class DownsampleJIT(object): 23 | def __init__(self, filt_size: int = 3, stride: int = 2, channels: int = 0): 24 | self.stride = stride 25 | self.filt_size = filt_size 26 | self.channels = channels 27 | 28 | assert self.filt_size == 3 29 | assert stride == 2 30 | a = torch.tensor([1., 2., 1.]) 31 | 32 | filt = (a[:, None] * a[None, :]).clone().detach() 33 | filt = filt / torch.sum(filt) 34 | self.filt = filt[None, None, :, :].repeat((self.channels, 1, 1, 1)).cuda().half() 35 | 36 | def __call__(self, input: torch.Tensor): 37 | if input.dtype != self.filt.dtype: 38 | self.filt = self.filt.float() 39 | input_pad = F.pad(input, (1, 1, 1, 1), 'reflect') 40 | return F.conv2d(input_pad, self.filt, stride=2, padding=0, groups=input.shape[1]) 41 | 42 | 43 | class Downsample(nn.Module): 44 | def __init__(self, filt_size=3, stride=2, channels=None): 45 | super(Downsample, self).__init__() 46 | self.filt_size = filt_size 47 | self.stride = stride 48 | self.channels = channels 49 | 50 | 51 | assert self.filt_size == 3 52 | a = torch.tensor([1., 2., 1.]) 53 | 54 | filt = (a[:, None] * a[None, :]) 55 | filt = filt / torch.sum(filt) 56 | 57 | # self.filt = filt[None, None, :, :].repeat((self.channels, 1, 1, 1)) 58 | self.register_buffer('filt', filt[None, None, :, :].repeat((self.channels, 1, 1, 1))) 59 | 60 | def forward(self, input): 61 | input_pad = F.pad(input, (1, 1, 1, 1), 'reflect') 62 | return F.conv2d(input_pad, self.filt, stride=self.stride, padding=0, groups=input.shape[1]) 63 | -------------------------------------------------------------------------------- /src/models/tresnet/layers/avg_pool.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | 6 | 7 | class FastGlobalAvgPool2d(nn.Module): 8 | def __init__(self, flatten=False): 9 | super(FastGlobalAvgPool2d, self).__init__() 10 | self.flatten = flatten 11 | 12 | def forward(self, x): 13 | if self.flatten: 14 | in_size = x.size() 15 | return x.view((in_size[0], in_size[1], -1)).mean(dim=2) 16 | else: 17 | return x.view(x.size(0), x.size(1), -1).mean(-1).view(x.size(0), x.size(1), 1, 1) 18 | 19 | 20 | -------------------------------------------------------------------------------- /src/models/tresnet/layers/space_to_depth.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | 5 | class SpaceToDepth(nn.Module): 6 | def __init__(self, block_size=4): 7 | super().__init__() 8 | assert block_size == 4 9 | self.bs = block_size 10 | 11 | def forward(self, x): 12 | N, C, H, W = x.size() 13 | x = x.view(N, C, H // self.bs, self.bs, W // self.bs, self.bs) # (N, C, H//bs, bs, W//bs, bs) 14 | x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # (N, bs, bs, C, H//bs, W//bs) 15 | x = x.view(N, C * (self.bs ** 2), H // self.bs, W // self.bs) # (N, C*bs^2, H//bs, W//bs) 16 | return x 17 | 18 | 19 | @torch.jit.script 20 | class SpaceToDepthJit(object): 21 | def __call__(self, x: torch.Tensor): 22 | # assuming hard-coded that block_size==4 for acceleration 23 | N, C, H, W = x.size() 24 | x = x.view(N, C, H // 4, 4, W // 4, 4) # (N, C, H//bs, bs, W//bs, bs) 25 | x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # (N, bs, bs, C, H//bs, W//bs) 26 | x = x.view(N, C * 16, H // 4, W // 4) # (N, C*bs^2, H//bs, W//bs) 27 | return x 28 | 29 | 30 | class SpaceToDepthModule(nn.Module): 31 | def __init__(self, remove_model_jit=False): 32 | super().__init__() 33 | if not remove_model_jit: 34 | self.op = SpaceToDepthJit() 35 | else: 36 | self.op = SpaceToDepth() 37 | 38 | def forward(self, x): 39 | return self.op(x) 40 | 41 | 42 | class DepthToSpace(nn.Module): 43 | 44 | def __init__(self, block_size): 45 | super().__init__() 46 | self.bs = block_size 47 | 48 | def forward(self, x): 49 | N, C, H, W = x.size() 50 | x = x.view(N, self.bs, self.bs, C // (self.bs ** 2), H, W) # (N, bs, bs, C//bs^2, H, W) 51 | x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # (N, C//bs^2, H, bs, W, bs) 52 | x = x.view(N, C // (self.bs ** 2), H * self.bs, W * self.bs) # (N, C//bs^2, H * bs, W * bs) 53 | return x -------------------------------------------------------------------------------- /src/models/tresnet/layers/squeeze_and_excite.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.nn.functional as F 3 | from src.models.tresnet.layers.avg_pool import FastGlobalAvgPool2d 4 | 5 | 6 | class Flatten(nn.Module): 7 | def forward(self, x): 8 | return x.view(x.size(0), -1) 9 | 10 | 11 | class SEModule(nn.Module): 12 | 13 | def __init__(self, channels, reduction_channels, inplace=True): 14 | super(SEModule, self).__init__() 15 | self.avg_pool = FastGlobalAvgPool2d() 16 | self.fc1 = nn.Conv2d(channels, reduction_channels, kernel_size=1, padding=0, bias=True) 17 | self.relu = nn.ReLU(inplace=inplace) 18 | self.fc2 = nn.Conv2d(reduction_channels, channels, kernel_size=1, padding=0, bias=True) 19 | # self.activation = hard_sigmoid(inplace=inplace) 20 | self.activation = nn.Sigmoid() 21 | 22 | def forward(self, x): 23 | x_se = self.avg_pool(x) 24 | x_se2 = self.fc1(x_se) 25 | x_se2 = self.relu(x_se2) 26 | x_se = self.fc2(x_se2) 27 | x_se = self.activation(x_se) 28 | return x * x_se 29 | 30 | class hard_sigmoid(nn.Module): 31 | def __init__(self, inplace=True): 32 | super(hard_sigmoid, self).__init__() 33 | self.inplace = inplace 34 | 35 | def forward(self, x): 36 | if self.inplace: 37 | return x.add_(3.).clamp_(0., 6.).div_(6.) 38 | else: 39 | return F.relu6(x + 3.) / 6. -------------------------------------------------------------------------------- /src/models/tresnet/tresnet.py: -------------------------------------------------------------------------------- 1 | from functools import partial 2 | 3 | import torch 4 | import torch.nn as nn 5 | from collections import OrderedDict 6 | from src.models.tresnet.layers.anti_aliasing import AntiAliasDownsampleLayer 7 | from .layers.avg_pool import FastGlobalAvgPool2d 8 | from .layers.squeeze_and_excite import SEModule 9 | from src.models.tresnet.layers.space_to_depth import SpaceToDepthModule 10 | from inplace_abn import InPlaceABN 11 | 12 | 13 | def IABN2Float(module: nn.Module) -> nn.Module: 14 | "If `module` is IABN don't use half precision." 15 | if isinstance(module, InPlaceABN): 16 | module.float() 17 | for child in module.children(): IABN2Float(child) 18 | return module 19 | 20 | 21 | def conv2d_ABN(ni, nf, stride, activation="leaky_relu", kernel_size=3, activation_param=1e-2, groups=1): 22 | return nn.Sequential( 23 | nn.Conv2d(ni, nf, kernel_size=kernel_size, stride=stride, padding=kernel_size // 2, groups=groups, 24 | bias=False), 25 | InPlaceABN(num_features=nf, activation=activation, activation_param=activation_param) 26 | ) 27 | 28 | 29 | class BasicBlock(nn.Module): 30 | expansion = 1 31 | 32 | def __init__(self, inplanes, planes, stride=1, downsample=None, use_se=True, anti_alias_layer=None): 33 | super(BasicBlock, self).__init__() 34 | if stride == 1: 35 | self.conv1 = conv2d_ABN(inplanes, planes, stride=1, activation_param=1e-3) 36 | else: 37 | if anti_alias_layer is None: 38 | self.conv1 = conv2d_ABN(inplanes, planes, stride=2, activation_param=1e-3) 39 | else: 40 | self.conv1 = nn.Sequential(conv2d_ABN(inplanes, planes, stride=1, activation_param=1e-3), 41 | anti_alias_layer(channels=planes, filt_size=3, stride=2)) 42 | 43 | self.conv2 = conv2d_ABN(planes, planes, stride=1, activation="identity") 44 | self.relu = nn.ReLU(inplace=True) 45 | self.downsample = downsample 46 | self.stride = stride 47 | reduce_layer_planes = max(planes * self.expansion // 4, 64) 48 | self.se = SEModule(planes * self.expansion, reduce_layer_planes) if use_se else None 49 | 50 | def forward(self, x): 51 | if self.downsample is not None: 52 | residual = self.downsample(x) 53 | else: 54 | residual = x 55 | 56 | out = self.conv1(x) 57 | out = self.conv2(out) 58 | 59 | if self.se is not None: out = self.se(out) 60 | 61 | out += residual 62 | 63 | out = self.relu(out) 64 | 65 | return out 66 | 67 | 68 | class Bottleneck(nn.Module): 69 | expansion = 4 70 | 71 | def __init__(self, inplanes, planes, stride=1, downsample=None, use_se=True, anti_alias_layer=None): 72 | super(Bottleneck, self).__init__() 73 | self.conv1 = conv2d_ABN(inplanes, planes, kernel_size=1, stride=1, activation="leaky_relu", 74 | activation_param=1e-3) 75 | if stride == 1: 76 | self.conv2 = conv2d_ABN(planes, planes, kernel_size=3, stride=1, activation="leaky_relu", 77 | activation_param=1e-3) 78 | else: 79 | if anti_alias_layer is None: 80 | self.conv2 = conv2d_ABN(planes, planes, kernel_size=3, stride=2, activation="leaky_relu", 81 | activation_param=1e-3) 82 | else: 83 | self.conv2 = nn.Sequential(conv2d_ABN(planes, planes, kernel_size=3, stride=1, 84 | activation="leaky_relu", activation_param=1e-3), 85 | anti_alias_layer(channels=planes, filt_size=3, stride=2)) 86 | 87 | self.conv3 = conv2d_ABN(planes, planes * self.expansion, kernel_size=1, stride=1, 88 | activation="identity") 89 | 90 | self.relu = nn.ReLU(inplace=True) 91 | self.downsample = downsample 92 | self.stride = stride 93 | 94 | reduce_layer_planes = max(planes * self.expansion // 8, 64) 95 | self.se = SEModule(planes, reduce_layer_planes) if use_se else None 96 | 97 | def forward(self, x): 98 | if self.downsample is not None: 99 | residual = self.downsample(x) 100 | else: 101 | residual = x 102 | 103 | out = self.conv1(x) 104 | out = self.conv2(out) 105 | if self.se is not None: out = self.se(out) 106 | 107 | out = self.conv3(out) 108 | out = out + residual # no inplace 109 | out = self.relu(out) 110 | 111 | return out 112 | 113 | 114 | class TResNet(nn.Module): 115 | 116 | def __init__(self, layers, in_chans=3, num_classes=1000, width_factor=1.0, remove_aa_jit=False): 117 | super(TResNet, self).__init__() 118 | 119 | # JIT layers 120 | space_to_depth = SpaceToDepthModule() 121 | anti_alias_layer = partial(AntiAliasDownsampleLayer, remove_aa_jit=remove_aa_jit) 122 | global_pool_layer = FastGlobalAvgPool2d(flatten=True) 123 | 124 | # TResnet stages 125 | self.inplanes = int(64 * width_factor) 126 | self.planes = int(64 * width_factor) 127 | conv1 = conv2d_ABN(in_chans * 16, self.planes, stride=1, kernel_size=3) 128 | layer1 = self._make_layer(BasicBlock, self.planes, layers[0], stride=1, use_se=True, 129 | anti_alias_layer=anti_alias_layer) # 56x56 130 | layer2 = self._make_layer(BasicBlock, self.planes * 2, layers[1], stride=2, use_se=True, 131 | anti_alias_layer=anti_alias_layer) # 28x28 132 | layer3 = self._make_layer(Bottleneck, self.planes * 4, layers[2], stride=2, use_se=True, 133 | anti_alias_layer=anti_alias_layer) # 14x14 134 | layer4 = self._make_layer(Bottleneck, self.planes * 8, layers[3], stride=2, use_se=False, 135 | anti_alias_layer=anti_alias_layer) # 7x7 136 | 137 | # body 138 | self.body = nn.Sequential(OrderedDict([ 139 | ('SpaceToDepth', space_to_depth), 140 | ('conv1', conv1), 141 | ('layer1', layer1), 142 | ('layer2', layer2), 143 | ('layer3', layer3), 144 | ('layer4', layer4)])) 145 | 146 | # head 147 | self.embeddings = [] 148 | self.global_pool = nn.Sequential(OrderedDict([('global_pool_layer', global_pool_layer)])) 149 | self.num_features = (self.planes * 8) * Bottleneck.expansion 150 | fc = nn.Linear(self.num_features, num_classes) 151 | self.head = nn.Sequential(OrderedDict([('fc', fc)])) 152 | 153 | # model initilization 154 | for m in self.modules(): 155 | if isinstance(m, nn.Conv2d): 156 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='leaky_relu') 157 | elif isinstance(m, nn.BatchNorm2d) or isinstance(m, InPlaceABN): 158 | nn.init.constant_(m.weight, 1) 159 | nn.init.constant_(m.bias, 0) 160 | 161 | # residual connections special initialization 162 | for m in self.modules(): 163 | if isinstance(m, BasicBlock): 164 | m.conv2[1].weight = nn.Parameter(torch.zeros_like(m.conv2[1].weight)) # BN to zero 165 | if isinstance(m, Bottleneck): 166 | m.conv3[1].weight = nn.Parameter(torch.zeros_like(m.conv3[1].weight)) # BN to zero 167 | if isinstance(m, nn.Linear): m.weight.data.normal_(0, 0.01) 168 | 169 | def _make_layer(self, block, planes, blocks, stride=1, use_se=True, anti_alias_layer=None): 170 | downsample = None 171 | if stride != 1 or self.inplanes != planes * block.expansion: 172 | layers = [] 173 | if stride == 2: 174 | # avg pooling before 1x1 conv 175 | layers.append(nn.AvgPool2d(kernel_size=2, stride=2, ceil_mode=True, count_include_pad=False)) 176 | layers += [conv2d_ABN(self.inplanes, planes * block.expansion, kernel_size=1, stride=1, 177 | activation="identity")] 178 | downsample = nn.Sequential(*layers) 179 | 180 | layers = [] 181 | layers.append(block(self.inplanes, planes, stride, downsample, use_se=use_se, 182 | anti_alias_layer=anti_alias_layer)) 183 | self.inplanes = planes * block.expansion 184 | for i in range(1, blocks): layers.append( 185 | block(self.inplanes, planes, use_se=use_se, anti_alias_layer=anti_alias_layer)) 186 | return nn.Sequential(*layers) 187 | 188 | def forward(self, x): 189 | x = self.body(x) 190 | self.embeddings = self.global_pool(x) 191 | logits = self.head(self.embeddings) 192 | return logits 193 | 194 | 195 | def TResnetM(model_params): 196 | """ Constructs a medium TResnet model. 197 | """ 198 | in_chans = 3 199 | num_classes = model_params['num_classes'] 200 | remove_aa_jit = model_params['remove_aa_jit'] 201 | model = TResNet(layers=[3, 4, 11, 3], num_classes=num_classes, in_chans=in_chans, 202 | remove_aa_jit=remove_aa_jit) 203 | return model 204 | 205 | 206 | def TResnetL(model_params): 207 | """ Constructs a large TResnet model. 208 | """ 209 | in_chans = 3 210 | num_classes = model_params['num_classes'] 211 | remove_aa_jit = model_params['remove_aa_jit'] 212 | model = TResNet(layers=[4, 5, 18, 3], num_classes=num_classes, in_chans=in_chans, width_factor=1.2, 213 | remove_aa_jit=remove_aa_jit) 214 | return model 215 | 216 | 217 | def TResnetXL(model_params): 218 | """ Constructs an extra-large TResnet model. 219 | """ 220 | in_chans = 3 221 | num_classes = model_params['num_classes'] 222 | remove_aa_jit = model_params['remove_aa_jit'] 223 | model = TResNet(layers=[4, 5, 24, 3], num_classes=num_classes, in_chans=in_chans, width_factor=1.3, 224 | remove_aa_jit=remove_aa_jit) 225 | 226 | return model 227 | -------------------------------------------------------------------------------- /src/models/tresnet_v2/__init__.py: -------------------------------------------------------------------------------- 1 | from .tresnet_v2 import TResnetL_V2 2 | -------------------------------------------------------------------------------- /src/models/tresnet_v2/layers/anti_aliasing.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn.parallel 3 | import numpy as np 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | 7 | 8 | class AntiAliasDownsampleLayer(nn.Module): 9 | def __init__(self, remove_aa_jit: bool = False, filt_size: int = 3, stride: int = 2, 10 | channels: int = 0): 11 | super(AntiAliasDownsampleLayer, self).__init__() 12 | if not remove_aa_jit: 13 | self.op = DownsampleJIT(filt_size, stride, channels) 14 | else: 15 | self.op = Downsample(filt_size, stride, channels) 16 | 17 | def forward(self, x): 18 | return self.op(x) 19 | 20 | 21 | @torch.jit.script 22 | class DownsampleJIT(object): 23 | def __init__(self, filt_size: int = 3, stride: int = 2, channels: int = 0): 24 | self.stride = stride 25 | self.filt_size = filt_size 26 | self.channels = channels 27 | 28 | assert self.filt_size == 3 29 | assert stride == 2 30 | a = torch.tensor([1., 2., 1.]) 31 | 32 | filt = (a[:, None] * a[None, :]).clone().detach() 33 | filt = filt / torch.sum(filt) 34 | self.filt = filt[None, None, :, :].repeat((self.channels, 1, 1, 1)).cuda().half() 35 | 36 | def __call__(self, input: torch.Tensor): 37 | if input.dtype != self.filt.dtype: 38 | self.filt = self.filt.float() 39 | input_pad = F.pad(input, (1, 1, 1, 1), 'reflect') 40 | return F.conv2d(input_pad, self.filt, stride=2, padding=0, groups=input.shape[1]) 41 | 42 | 43 | class Downsample(nn.Module): 44 | def __init__(self, filt_size=3, stride=2, channels=None): 45 | super(Downsample, self).__init__() 46 | self.filt_size = filt_size 47 | self.stride = stride 48 | self.channels = channels 49 | 50 | 51 | assert self.filt_size == 3 52 | a = torch.tensor([1., 2., 1.]) 53 | 54 | filt = (a[:, None] * a[None, :]) 55 | filt = filt / torch.sum(filt) 56 | 57 | # self.filt = filt[None, None, :, :].repeat((self.channels, 1, 1, 1)) 58 | self.register_buffer('filt', filt[None, None, :, :].repeat((self.channels, 1, 1, 1))) 59 | 60 | def forward(self, input): 61 | input_pad = F.pad(input, (1, 1, 1, 1), 'reflect') 62 | return F.conv2d(input_pad, self.filt, stride=self.stride, padding=0, groups=input.shape[1]) 63 | -------------------------------------------------------------------------------- /src/models/tresnet_v2/layers/avg_pool.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | 6 | 7 | class FastGlobalAvgPool2d(nn.Module): 8 | def __init__(self, flatten=False): 9 | super(FastGlobalAvgPool2d, self).__init__() 10 | self.flatten = flatten 11 | 12 | def forward(self, x): 13 | if self.flatten: 14 | in_size = x.size() 15 | return x.view((in_size[0], in_size[1], -1)).mean(dim=2) 16 | else: 17 | return x.view(x.size(0), x.size(1), -1).mean(-1).view(x.size(0), x.size(1), 1, 1) 18 | 19 | 20 | -------------------------------------------------------------------------------- /src/models/tresnet_v2/layers/space_to_depth.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | 5 | class SpaceToDepth(nn.Module): 6 | def __init__(self, block_size=4): 7 | super().__init__() 8 | assert block_size == 4 9 | self.bs = block_size 10 | 11 | def forward(self, x): 12 | N, C, H, W = x.size() 13 | x = x.view(N, C, H // self.bs, self.bs, W // self.bs, self.bs) # (N, C, H//bs, bs, W//bs, bs) 14 | x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # (N, bs, bs, C, H//bs, W//bs) 15 | x = x.view(N, C * (self.bs ** 2), H // self.bs, W // self.bs) # (N, C*bs^2, H//bs, W//bs) 16 | return x 17 | 18 | 19 | @torch.jit.script 20 | class SpaceToDepthJit(object): 21 | def __call__(self, x: torch.Tensor): 22 | # assuming hard-coded that block_size==4 for acceleration 23 | N, C, H, W = x.size() 24 | x = x.view(N, C, H // 4, 4, W // 4, 4) # (N, C, H//bs, bs, W//bs, bs) 25 | x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # (N, bs, bs, C, H//bs, W//bs) 26 | x = x.view(N, C * 16, H // 4, W // 4) # (N, C*bs^2, H//bs, W//bs) 27 | return x 28 | 29 | 30 | class SpaceToDepthModule(nn.Module): 31 | def __init__(self, remove_model_jit=False): 32 | super().__init__() 33 | if not remove_model_jit: 34 | self.op = SpaceToDepthJit() 35 | else: 36 | self.op = SpaceToDepth() 37 | 38 | def forward(self, x): 39 | return self.op(x) 40 | 41 | 42 | class DepthToSpace(nn.Module): 43 | 44 | def __init__(self, block_size): 45 | super().__init__() 46 | self.bs = block_size 47 | 48 | def forward(self, x): 49 | N, C, H, W = x.size() 50 | x = x.view(N, self.bs, self.bs, C // (self.bs ** 2), H, W) # (N, bs, bs, C//bs^2, H, W) 51 | x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # (N, C//bs^2, H, bs, W, bs) 52 | x = x.view(N, C // (self.bs ** 2), H * self.bs, W * self.bs) # (N, C//bs^2, H * bs, W * bs) 53 | return x -------------------------------------------------------------------------------- /src/models/tresnet_v2/layers/squeeze_and_excite.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.nn.functional as F 3 | from src.models.tresnet.layers.avg_pool import FastGlobalAvgPool2d 4 | 5 | 6 | class Flatten(nn.Module): 7 | def forward(self, x): 8 | return x.view(x.size(0), -1) 9 | 10 | 11 | class SEModule(nn.Module): 12 | 13 | def __init__(self, channels, reduction_channels, inplace=True): 14 | super(SEModule, self).__init__() 15 | self.avg_pool = FastGlobalAvgPool2d() 16 | self.fc1 = nn.Conv2d(channels, reduction_channels, kernel_size=1, padding=0, bias=True) 17 | self.relu = nn.ReLU(inplace=inplace) 18 | self.fc2 = nn.Conv2d(reduction_channels, channels, kernel_size=1, padding=0, bias=True) 19 | # self.activation = hard_sigmoid(inplace=inplace) 20 | self.activation = nn.Sigmoid() 21 | 22 | def forward(self, x): 23 | x_se = self.avg_pool(x) 24 | x_se2 = self.fc1(x_se) 25 | x_se2 = self.relu(x_se2) 26 | x_se = self.fc2(x_se2) 27 | x_se = self.activation(x_se) 28 | return x * x_se 29 | 30 | class hard_sigmoid(nn.Module): 31 | def __init__(self, inplace=True): 32 | super(hard_sigmoid, self).__init__() 33 | self.inplace = inplace 34 | 35 | def forward(self, x): 36 | if self.inplace: 37 | return x.add_(3.).clamp_(0., 6.).div_(6.) 38 | else: 39 | return F.relu6(x + 3.) / 6. -------------------------------------------------------------------------------- /src/models/tresnet_v2/tresnet_v2.py: -------------------------------------------------------------------------------- 1 | import math 2 | from collections import OrderedDict 3 | from functools import partial 4 | 5 | import torch.nn as nn 6 | import torch 7 | from torch.nn import Module as Module 8 | 9 | from src.models.tresnet.layers.anti_aliasing import AntiAliasDownsampleLayer 10 | from .layers.avg_pool import FastGlobalAvgPool2d 11 | from src.models.tresnet.layers.space_to_depth import SpaceToDepthModule 12 | from .layers.squeeze_and_excite import SEModule 13 | 14 | from inplace_abn import InPlaceABN 15 | from inplace_abn import ABN 16 | 17 | 18 | def InplacABN_to_ABN(module: nn.Module) -> nn.Module: 19 | # convert all InplaceABN layer to bit-accurate ABN layers. 20 | if isinstance(module, InPlaceABN): 21 | module_new = ABN(module.num_features, activation=module.activation, 22 | activation_param=module.activation_param) 23 | for key in module.state_dict(): 24 | module_new.state_dict()[key].copy_(module.state_dict()[key]) 25 | module_new.training = module.training 26 | module_new.weight.data = module_new.weight.abs() + module_new.eps 27 | return module_new 28 | for name, child in reversed(module._modules.items()): 29 | new_child = InplacABN_to_ABN(child) 30 | if new_child != child: 31 | module._modules[name] = new_child 32 | return module 33 | 34 | 35 | def conv3x3(in_planes, out_planes, stride=1): 36 | return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) 37 | 38 | 39 | def conv3x3_depth(planes, stride=1): 40 | return nn.Conv2d(planes, planes, groups=planes, kernel_size=3, stride=stride, padding=1, bias=False) 41 | 42 | 43 | def conv2d(ni, nf, stride): 44 | return nn.Sequential( 45 | nn.Conv2d(ni, nf, kernel_size=3, stride=stride, padding=1, bias=False), 46 | nn.BatchNorm2d(nf), 47 | nn.ReLU(inplace=True) 48 | ) 49 | 50 | 51 | def conv2d_ABN(ni, nf, stride, activation="leaky_relu", kernel_size=3, activation_param=1e-2, groups=1): 52 | activation_param = 1e-6 53 | return nn.Sequential( 54 | nn.Conv2d(ni, nf, kernel_size=kernel_size, stride=stride, padding=kernel_size // 2, groups=groups, 55 | bias=False), 56 | InPlaceABN(num_features=nf, activation=activation, activation_param=activation_param) 57 | ) 58 | 59 | 60 | class BasicBlock(Module): 61 | expansion = 1 62 | 63 | def __init__(self, inplanes, planes, stride=1, downsample=None, use_se=True, anti_alias_layer=None): 64 | super(BasicBlock, self).__init__() 65 | if stride == 1: 66 | self.conv1 = conv2d_ABN(inplanes, planes, stride=1, activation_param=1e-3) 67 | else: 68 | if anti_alias_layer is None: 69 | self.conv1 = conv2d_ABN(inplanes, planes, stride=2, activation_param=1e-3) 70 | else: 71 | self.conv1 = nn.Sequential(conv2d_ABN(inplanes, planes, stride=1, activation_param=1e-3), 72 | anti_alias_layer(channels=planes, filt_size=3, stride=2)) 73 | 74 | self.conv2 = conv2d_ABN(planes, planes, stride=1, activation="identity") 75 | self.relu = nn.ReLU(inplace=True) 76 | self.downsample = downsample 77 | self.stride = stride 78 | reduce_layer_planes = max(planes * self.expansion // 4, 64) 79 | self.se = SEModule(channels=planes * self.expansion, reduction_channels=reduce_layer_planes) if \ 80 | use_se else None 81 | 82 | def forward(self, x): 83 | if self.downsample is not None: 84 | residual = self.downsample(x) 85 | else: 86 | residual = x 87 | 88 | out = self.conv1(x) 89 | out = self.conv2(out) 90 | 91 | if self.se is not None: out = self.se(out) 92 | 93 | out += residual 94 | 95 | out = self.relu(out) 96 | 97 | return out 98 | 99 | 100 | class Bottleneck(Module): 101 | expansion = 4 102 | 103 | def __init__(self, inplanes, planes, stride=1, downsample=None, use_se=True, anti_alias_layer=None): 104 | super(Bottleneck, self).__init__() 105 | self.conv1 = conv2d_ABN(inplanes, planes, kernel_size=1, stride=1, activation="leaky_relu", 106 | activation_param=1e-3) 107 | if stride == 1: 108 | self.conv2 = conv2d_ABN(planes, planes, kernel_size=3, stride=1, activation="leaky_relu", 109 | activation_param=1e-3) 110 | else: 111 | if anti_alias_layer is None: 112 | self.conv2 = conv2d_ABN(planes, planes, kernel_size=3, stride=2, activation="leaky_relu", 113 | activation_param=1e-3) 114 | else: 115 | self.conv2 = nn.Sequential(conv2d_ABN(planes, planes, kernel_size=3, stride=1, 116 | activation="leaky_relu", activation_param=1e-3), 117 | anti_alias_layer(channels=planes, filt_size=3, stride=2)) 118 | 119 | self.conv3 = conv2d_ABN(planes, planes * self.expansion, kernel_size=1, stride=1, 120 | activation="identity") 121 | 122 | self.relu = nn.ReLU(inplace=True) 123 | self.downsample = downsample 124 | self.stride = stride 125 | 126 | reduce_layer_planes = max(planes * self.expansion // 8, 64) 127 | self.se = SEModule(planes, reduce_layer_planes) if use_se else None 128 | 129 | def forward(self, x): 130 | if self.downsample is not None: 131 | residual = self.downsample(x) 132 | else: 133 | residual = x 134 | 135 | out = self.conv1(x) 136 | out = self.conv2(out) 137 | if self.se is not None: out = self.se(out) 138 | 139 | out = self.conv3(out) 140 | out = out + residual # no inplace 141 | out = self.relu(out) 142 | 143 | return out 144 | 145 | 146 | class TResNetV2(Module): 147 | 148 | def __init__(self, layers, in_chans=3, num_classes=1000, width_factor=1.0, remove_model_jit=False): 149 | super(TResNetV2, self).__init__() 150 | ## body 151 | self.inplanes = int(int(64 * width_factor + 4) / 8) * 8 152 | self.planes = int(int(64 * width_factor + 4) / 8) * 8 153 | SpaceToDepth = SpaceToDepthModule(remove_model_jit=remove_model_jit) 154 | conv1 = conv2d_ABN(in_chans * 16, self.planes, stride=1, kernel_size=3) 155 | 156 | anti_alias_layer = partial(AntiAliasDownsampleLayer, remove_aa_jit=remove_model_jit) 157 | global_pool_layer = FastGlobalAvgPool2d(flatten=True) 158 | layer1 = self._make_layer(Bottleneck, self.planes, layers[0], stride=1, use_se=True, 159 | anti_alias_layer=anti_alias_layer) # 56x56 160 | layer2 = self._make_layer(Bottleneck, self.planes * 2, layers[1], stride=2, use_se=True, 161 | anti_alias_layer=anti_alias_layer) # 28x28 162 | layer3 = self._make_layer(Bottleneck, self.planes * 4, layers[2], stride=2, use_se=True, 163 | anti_alias_layer=anti_alias_layer) # 14x14 164 | layer4 = self._make_layer(Bottleneck, self.planes * 8, layers[3], stride=2, use_se=False, 165 | anti_alias_layer=anti_alias_layer) # 7x7 166 | 167 | self.body = nn.Sequential(OrderedDict([ 168 | ('SpaceToDepth', SpaceToDepth), 169 | ('conv1', conv1), 170 | ('layer1', layer1), 171 | ('layer2', layer2), 172 | ('layer3', layer3), 173 | ('layer4', layer4)])) 174 | 175 | # default head 176 | self.num_features = (self.planes * 8) * Bottleneck.expansion 177 | 178 | fc = nn.Linear(self.num_features , num_classes) 179 | 180 | self.global_pool = nn.Sequential(OrderedDict([('global_pool_layer', global_pool_layer)])) 181 | 182 | self.head = nn.Sequential(OrderedDict([('fc', fc)])) 183 | 184 | self.embeddings = [] 185 | 186 | for m in self.modules(): 187 | if isinstance(m, nn.Conv2d): 188 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='leaky_relu') 189 | elif isinstance(m, nn.BatchNorm2d) or isinstance(m, InPlaceABN): 190 | nn.init.constant_(m.weight, 1) 191 | nn.init.constant_(m.bias, 0) 192 | 193 | # initilize resnet in a magic way 194 | for m in self.modules(): 195 | if isinstance(m, BasicBlock): 196 | m.conv2[1].weight = nn.Parameter(torch.zeros_like(m.conv2[1].weight)) # BN to zero 197 | if isinstance(m, Bottleneck): 198 | m.conv3[1].weight = nn.Parameter(torch.zeros_like(m.conv3[1].weight)) # BN to zero 199 | if isinstance(m, nn.Linear): m.weight.data.normal_(0, 0.01) 200 | 201 | def _make_layer(self, block, planes, blocks, stride=1, use_se=True, anti_alias_layer=None): 202 | downsample = None 203 | if stride != 1 or self.inplanes != planes * block.expansion: 204 | layers = [] 205 | if stride == 2: 206 | # avg pooling before 1x1 conv 207 | layers.append( 208 | nn.AvgPool2d(kernel_size=2, stride=2, ceil_mode=True, count_include_pad=False)) 209 | layers += [ 210 | conv2d_ABN(self.inplanes, planes * block.expansion, kernel_size=1, stride=1, 211 | activation="identity")] 212 | downsample = nn.Sequential(*layers) 213 | 214 | layers = [] 215 | layers.append(block(self.inplanes, planes, stride, downsample, use_se=use_se, 216 | anti_alias_layer=anti_alias_layer)) 217 | self.inplanes = planes * block.expansion 218 | for i in range(1, blocks): layers.append( 219 | block(self.inplanes, planes, use_se=use_se, anti_alias_layer=anti_alias_layer)) 220 | return nn.Sequential(*layers) 221 | 222 | def forward(self, x): 223 | x = self.body(x) 224 | # self.embeddings = self.global_pool(x) 225 | logits = self.head(self.global_pool(x)) 226 | return logits 227 | 228 | 229 | def TResnetL_V2(model_params): 230 | """Constructs a large TResnet model. 231 | """ 232 | in_chans = 3 233 | num_classes = model_params['num_classes'] 234 | remove_model_jit = False 235 | layers_list = [3, 4, 23, 3] 236 | width_factor = 1.0 237 | model = TResNetV2(layers=layers_list, num_classes=num_classes, in_chans=in_chans, 238 | width_factor=width_factor, remove_model_jit=remove_model_jit) 239 | return model 240 | -------------------------------------------------------------------------------- /src/models/utils/__init__.py: -------------------------------------------------------------------------------- 1 | from .factory import create_model 2 | __all__ = ['create_model'] -------------------------------------------------------------------------------- /src/models/utils/factory.py: -------------------------------------------------------------------------------- 1 | import logging 2 | 3 | from ..tresnet_v2 import TResnetL_V2 4 | 5 | logger = logging.getLogger(__name__) 6 | 7 | from ..tresnet import TResnetM, TResnetL, TResnetXL 8 | 9 | 10 | def create_model(args): 11 | """Create a model 12 | """ 13 | model_params = {'args': args, 'num_classes': args.num_classes,'remove_aa_jit': args.remove_aa_jit} 14 | args = model_params['args'] 15 | args.model_name = args.model_name.lower() 16 | 17 | if args.model_name=='tresnet_m': 18 | model = TResnetM(model_params) 19 | elif args.model_name=='tresnet_l': 20 | model = TResnetL(model_params) 21 | elif args.model_name=='tresnet_l_v2': 22 | model = TResnetL_V2(model_params) 23 | elif args.model_name=='tresnet_xl': 24 | model = TResnetXL(model_params) 25 | else: 26 | print("model: {} not found !!".format(args.model_name)) 27 | exit(-1) 28 | 29 | return model 30 | -------------------------------------------------------------------------------- /tests/test_TResNetV2: -------------------------------------------------------------------------------- 1 | 2 | # test_TResNetV2.py - Generated by https://www.codium.ai/ 3 | 4 | import unittest 5 | 6 | """ 7 | Code Analysis: 8 | - This class is a subclass of the Module class from the torch.nn library. 9 | - It initializes the TResNetV2 model with the given parameters. 10 | - It creates a convolutional neural network with a body, head, and global pooling layers. 11 | - It uses a SpaceToDepthModule, conv2d_ABN, AntiAliasDownsampleLayer, FastGlobalAvgPool2d, SEModule, InPlaceABN, and ABN. 12 | - It uses a Bottleneck block with a convolutional layer, batch normalization, and ReLU activation. 13 | - It uses a BasicBlock with a convolutional layer, batch normalization, and ReLU activation. 14 | - It initializes the weights and biases of the convolutional layers and linear layers with Kaiming normal and constant values, respectively. 15 | - It sets the conv2 and conv3 weights of the BasicBlock and Bottleneck layers to zero. 16 | - It has a forward method which takes in an input and returns the logits. 17 | """ 18 | 19 | 20 | """ 21 | Test strategies: 22 | - test_init(): tests that the TResNetV2 model is initialized correctly with the given parameters. 23 | - test_conv1(): tests that the conv1 layer is initialized correctly. 24 | - test_anti_alias_layer(): tests that the AntiAliasDownsampleLayer is initialized correctly. 25 | - test_global_pool_layer(): tests that the FastGlobalAvgPool2d is initialized correctly. 26 | - test_bottleneck_block(): tests that the Bottleneck block is initialized correctly. 27 | - test_basic_block(): tests that the BasicBlock is initialized correctly. 28 | - test_weights_and_biases(): tests that the weights and biases of the convolutional layers and linear layers are initialized correctly. 29 | - test_conv2_and_conv3_weights(): tests that the conv2 and conv3 weights of the BasicBlock and Bottleneck layers are set to zero. 30 | - test_forward(): tests that the forward method takes in an input and returns the logits. 31 | """ 32 | 33 | 34 | class TestTResNetV2(unittest.TestCase): 35 | 36 | def setUp(self): 37 | self.layers = [3, 4, 6, 3] 38 | self.in_chans = 3 39 | self.num_classes = 1000 40 | self.width_factor = 1.0 41 | self.remove_model_jit = False 42 | self.model = TResNetV2(self.layers, self.in_chans, self.num_classes, self.width_factor, self.remove_model_jit) 43 | 44 | def test_init(self): 45 | self.assertEqual(self.model.inplanes, 64) 46 | self.assertEqual(self.model.planes, 64) 47 | self.assertEqual(self.model.num_features, 2048) 48 | 49 | def test_conv1(self): 50 | conv1 = self.model.body[1] 51 | self.assertIsInstance(conv1, ABN) 52 | self.assertEqual(conv1.in_channels, 48) 53 | self.assertEqual(conv1.out_channels, 64) 54 | self.assertEqual(conv1.kernel_size, (3, 3)) 55 | self.assertEqual(conv1.stride, (1, 1)) 56 | 57 | def test_anti_alias_layer(self): 58 | layer1 = self.model.body[3] 59 | anti_alias_layer = layer1[0].anti_alias_layer 60 | self.assertIsInstance(anti_alias_layer, partial) 61 | self.assertEqual(anti_alias_layer.func, AntiAliasDownsampleLayer) 62 | self.assertEqual(anti_alias_layer.args, (self.remove_model_jit,)) 63 | 64 | def test_global_pool_layer(self): 65 | global_pool = self.model.global_pool[0] 66 | self.assertIsInstance(global_pool, FastGlobalAvgPool2d) 67 | self.assertTrue(global_pool.flatten) 68 | 69 | def test_bottleneck_block(self): 70 | layer1 = self.model.body[3] 71 | bottleneck = layer1[0] 72 | self.assertIsInstance(bottleneck, Bottleneck) 73 | self.assertEqual(bottleneck.inplanes, 64) 74 | self.assertEqual(bottleneck.planes, 64) 75 | self.assertEqual(bottleneck.stride, 1) 76 | self.assertIsInstance(bottleneck.downsample, nn.Sequential) 77 | self.assertIsInstance(bottleneck.se, SEModule) 78 | 79 | def test_basic_block(self): 80 | layer2 = self.model.body[4] 81 | basicblock = layer2[0] 82 | self.assertIsInstance(basicblock, BasicBlock) 83 | self.assertEqual(basicblock.inplanes, 128) 84 | self.assertEqual(basicblock.planes, 128) 85 | self.assertEqual(basicblock.stride, 2) 86 | self.assertIsInstance(basicblock.downsample, nn.Sequential) 87 | 88 | def test_weights_and_biases(self): 89 | for m in self.model.modules(): 90 | if isinstance(m, nn.Conv2d): 91 | weight = m.weight 92 | fan = math.sqrt(weight[0].numel()) 93 | nn.init.kaiming_normal_(weight, mode='fan_out', nonlinearity='leaky_relu') 94 | for w in weight: 95 | for i in range(w[0].numel()): 96 | self.assertAlmostEqual(w[0][i], 0, delta=0.001 * fan) 97 | 98 | elif isinstance(m, nn.BatchNorm2d) or isinstance(m, InPlaceABN): 99 | weight = m.weight 100 | bias = m.bias 101 | nn.init.constant_(weight, 1) 102 | --------------------------------------------------------------------------------