├── .github └── ISSUE_TEMPLATE │ └── bug_report.md ├── Dockerfile ├── LICENSE ├── README.md ├── cfg ├── yolov3-1cls.cfg ├── yolov3-spp-1cls.cfg ├── yolov3-spp.cfg ├── yolov3-tiny-1cls.cfg ├── yolov3-tiny.cfg └── yolov3.cfg ├── data ├── 5k.shapes ├── 5k.txt ├── coco.data ├── coco.names ├── coco_1000img.data ├── coco_1000img.txt ├── coco_1000val.data ├── coco_1000val.txt ├── coco_16img.data ├── coco_16img.txt ├── coco_1cls.data ├── coco_1cls.txt ├── coco_1img.data ├── coco_1img.txt ├── coco_1k5k.data ├── coco_32img.data ├── coco_32img.txt ├── coco_500img.txt ├── coco_500val.data ├── coco_500val.txt ├── coco_64img.data ├── coco_64img.shapes ├── coco_64img.txt ├── coco_paper.names ├── get_coco_dataset.sh ├── get_coco_dataset_gdrive.sh ├── hand.data ├── samples │ ├── bus.jpg │ └── zidane.jpg ├── trainvalno5k.shapes └── valid_.shapes ├── detect.py ├── models.py ├── prune.py ├── requirements.txt ├── shortcut_prune.py ├── test.py ├── train.py ├── utils ├── adabound.py ├── datasets.py ├── google_utils.py ├── parse_config.py ├── prune_utils.py ├── torch_utils.py └── utils.py └── weights └── download_yolov3_weights.sh /.github/ISSUE_TEMPLATE/bug_report.md: -------------------------------------------------------------------------------- 1 | --- 2 | name: Bug report 3 | about: Create a report to help us improve 4 | title: '' 5 | labels: '' 6 | assignees: '' 7 | 8 | --- 9 | 10 | **Describe the bug** 11 | A clear and concise description of what the bug is. 12 | 13 | **To Reproduce** 14 | Steps to reproduce the behavior: 15 | 1. Go to '...' 16 | 2. Click on '....' 17 | 3. Scroll down to '....' 18 | 4. See error 19 | 20 | **Expected behavior** 21 | A clear and concise description of what you expected to happen. 22 | 23 | **Screenshots** 24 | If applicable, add screenshots to help explain your problem. 25 | 26 | **Desktop (please complete the following information):** 27 | - OS: [e.g. iOS] 28 | - Browser [e.g. chrome, safari] 29 | - Version [e.g. 22] 30 | 31 | **Smartphone (please complete the following information):** 32 | - Device: [e.g. iPhone6] 33 | - OS: [e.g. iOS8.1] 34 | - Browser [e.g. stock browser, safari] 35 | - Version [e.g. 22] 36 | 37 | **Additional context** 38 | Add any other context about the problem here. 39 | -------------------------------------------------------------------------------- /Dockerfile: -------------------------------------------------------------------------------- 1 | # Start from Nvidia PyTorch image https://ngc.nvidia.com/catalog/containers/nvidia:pytorch 2 | FROM nvcr.io/nvidia/pytorch:19.08-py3 3 | 4 | # Install dependencies (pip or conda) 5 | RUN pip install -U gsutil 6 | # RUN pip install -U -r requirements.txt 7 | # RUN conda update -n base -c defaults conda 8 | # RUN conda install -y -c anaconda future numpy opencv matplotlib tqdm pillow 9 | # RUN conda install -y -c conda-forge scikit-image tensorboard pycocotools 10 | 11 | ## Install OpenCV with Gstreamer support 12 | #WORKDIR /usr/src 13 | #RUN pip uninstall -y opencv-python 14 | #RUN apt-get update 15 | #RUN apt-get install -y gstreamer1.0-tools gstreamer1.0-python3-dbg-plugin-loader libgstreamer1.0-dev libgstreamer-plugins-base1.0-dev 16 | #RUN git clone https://github.com/opencv/opencv.git && cd opencv && git checkout 4.1.1 && mkdir build 17 | #RUN git clone https://github.com/opencv/opencv_contrib.git && cd opencv_contrib && git checkout 4.1.1 18 | #RUN cd opencv/build && cmake ../ \ 19 | # -D OPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules \ 20 | # -D BUILD_OPENCV_PYTHON3=ON \ 21 | # -D PYTHON3_EXECUTABLE=/opt/conda/bin/python \ 22 | # -D PYTHON3_INCLUDE_PATH=/opt/conda/include/python3.6m \ 23 | # -D PYTHON3_LIBRARIES=/opt/conda/lib/python3.6/site-packages \ 24 | # -D WITH_GSTREAMER=ON \ 25 | # -D WITH_FFMPEG=OFF \ 26 | # && make && make install && ldconfig 27 | #RUN cd /usr/local/lib/python3.6/site-packages/cv2/python-3.6/ && mv cv2.cpython-36m-x86_64-linux-gnu.so cv2.so 28 | #RUN cd /opt/conda/lib/python3.6/site-packages/ && ln -s /usr/local/lib/python3.6/site-packages/cv2/python-3.6/cv2.so cv2.so 29 | #RUN python3 -c "import cv2; print(cv2.getBuildInformation())" 30 | 31 | # Create working directory 32 | RUN mkdir -p /usr/src/app 33 | WORKDIR /usr/src/app 34 | 35 | # Copy contents 36 | COPY . /usr/src/app 37 | 38 | # Copy weights 39 | #RUN python3 -c "from utils.google_utils import *; \ 40 | # gdrive_download(id='18xqvs_uwAqfTXp-LJCYLYNHBOcrwbrp0', name='weights/darknet53.conv.74'); \ 41 | # gdrive_download(id='1oPCHKsM2JpM-zgyepQciGli9X0MTsJCO', name='weights/yolov3-spp.weights'); \ 42 | # gdrive_download(id='1vFlbJ_dXPvtwaLLOu-twnjK4exdFiQ73', name='weights/yolov3-spp.pt)" 43 | 44 | 45 | # --------------------------------------------------- Extras Below --------------------------------------------------- 46 | 47 | # Build 48 | # rm -rf yolov3 # Warning: remove existing 49 | # git clone https://github.com/ultralytics/yolov3 && cd yolov3 && python3 detect.py 50 | # sudo docker image prune -af && sudo docker build -t ultralytics/yolov3:v0 . 51 | 52 | # Run 53 | # sudo nvidia-docker run --ipc=host ultralytics/yolov3:v0 python3 detect.py 54 | 55 | # Run with local directory access 56 | # sudo nvidia-docker run --ipc=host --mount type=bind,source="$(pwd)"/coco,target=/usr/src/coco ultralytics/yolov3:v0 python3 train.py 57 | 58 | # Build and Push 59 | # export tag=ultralytics/yolov3:v0 && sudo docker build -t $tag . && docker push $tag 60 | 61 | # Kill all 62 | # sudo docker kill $(sudo docker ps -q) 63 | 64 | # Run bash for loop 65 | # sudo nvidia-docker run --ipc=host ultralytics/yolov3:v0 while true; do python3 train.py --evolve; done 66 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # YOLOv3-training-prune 2 | 3 | ## 环境 4 | 5 | Python3.6, 6 | Pytorch 1.1及以上， 7 | numpy>1.16， 8 | tensorboard=1.13以上 9 | 10 | YOLOv3 的训练参考[博客]（https://blog.csdn.net/qq_34795071/article/details/90769094 ） 代码基于的 [ultralytics/yolov3](https://github.com/ultralytics/yolov3) 11 | 12 | 大家也可以看下这个：https://github.com/tanluren/yolov3-channel-and-layer-pruning 13 | 14 | ## 正常训练（Baseline） 15 | 16 | ```bash 17 | python train.py --data data/VHR.data --cfg cfg/yolov3.cfg --weights/yolov3.weights --epochs 100 --batch-size 32 #后面的epochs自行更改 直接加载weights可以更好的收敛 18 | ``` 19 | 20 | ## 剪枝算法介绍 21 | 22 | 本代码基于论文 [Learning Efficient Convolutional Networks Through Network Slimming (ICCV 2017)](http://openaccess.thecvf.com/content_iccv_2017/html/Liu_Learning_Efficient_Convolutional_ICCV_2017_paper.html) 进行改进实现的 channel pruning算法，类似的代码实现还有这个 [yolov3-network-slimming](https://github.com/talebolano/yolov3-network-slimming)。原始论文中的算法是针对分类模型的，基于 BN 层的 gamma 系数进行剪枝的。 23 | 24 | 以下只是算法的大概步骤，具体实现过程中还要做 s 参数的尝试或者需要进行迭代式剪枝等。 25 | 26 | ## 参数设置 27 | 28 | -sr开启稀疏化， 29 | 30 | --s指定稀疏因子大小， 31 | 32 | --prune指定稀疏类型， 33 | 34 | --prune 0为正常剪枝和规整剪枝的稀疏化 35 | 36 | --prune 1为极限剪枝的稀疏化 37 | 38 | ## 进行稀疏化训练 39 | #### baseline后 生成pt 如果你不想加载yolov3.weights进行稀疏化训练 把baseline最后的pt转化为weights 进行稀疏化 40 | 41 | ```bashpython 42 | python train.py --cfg cfg/yolov3.cfg --data data/VHR.data --weights weights/XX.weights --epochs 100 --batch-size 32 -sr --s 0.001 --prune 0 #scale参数默认0.001，在数据分布广类别多的或者稀疏时掉点厉害的适当调小s 43 | ``` 44 | ## 训练过程中模型可视化 45 | ```bash 46 | tensorboard --logdir=runs 47 | ``` 48 | ## 模型剪枝 49 | ```bash 50 | python prune.py --cfg cfg/yolov3.cfg --data data/VHR.data --weights weights/last.pt --percent 0.5 51 | ``` 52 | ## shortcut_prune剪枝 53 | ```bash 54 | python shortcut_prune.py --cfg/yolov3.cfg --data data/VHR.data --weights weights/last.pt --percent 0.5 55 | ``` 56 | ## 模型进行微调 57 | ```bash 58 | python train.py --cfg cfg/prune_0.5_yolov3_cfg.cfg --data data/VHR.data --weights weights/prune_0.5_last.weights --epochs 100 --batch-size 32 59 | ``` 60 | ## convert cfg/pytorch model to darknet weights 61 | ```bash 62 | python -c "from models import *; convert('cfg/yolov3.cfg', 'weights/yolov3.pt')" 63 | Success: converted 'weights/yolov3.pt' to 'converted.weights' 64 | ``` 65 | ## 参考 66 | https://github.com/Lam1360/YOLOv3-model-pruning 67 | 68 | https://github.com/ultralytics/yolov3 69 | -------------------------------------------------------------------------------- /cfg/yolov3-1cls.cfg: -------------------------------------------------------------------------------- 1 | [net] 2 | # Testing 3 | #batch=1 4 | #subdivisions=1 5 | # Training 6 | batch=16 7 | subdivisions=1 8 | width=416 9 | height=416 10 | channels=3 11 | momentum=0.9 12 | decay=0.0005 13 | angle=0 14 | saturation = 1.5 15 | exposure = 1.5 16 | hue=.1 17 | 18 | learning_rate=0.001 19 | burn_in=1000 20 | max_batches = 500200 21 | policy=steps 22 | steps=400000,450000 23 | scales=.1,.1 24 | 25 | [convolutional] 26 | batch_normalize=1 27 | filters=32 28 | size=3 29 | stride=1 30 | pad=1 31 | activation=leaky 32 | 33 | # Downsample 34 | 35 | [convolutional] 36 | batch_normalize=1 37 | filters=64 38 | size=3 39 | stride=2 40 | pad=1 41 | activation=leaky 42 | 43 | [convolutional] 44 | batch_normalize=1 45 | filters=32 46 | size=1 47 | stride=1 48 | pad=1 49 | activation=leaky 50 | 51 | [convolutional] 52 | batch_normalize=1 53 | filters=64 54 | size=3 55 | stride=1 56 | pad=1 57 | activation=leaky 58 | 59 | [shortcut] 60 | from=-3 61 | activation=linear 62 | 63 | # Downsample 64 | 65 | [convolutional] 66 | batch_normalize=1 67 | filters=128 68 | size=3 69 | stride=2 70 | pad=1 71 | activation=leaky 72 | 73 | [convolutional] 74 | batch_normalize=1 75 | filters=64 76 | size=1 77 | stride=1 78 | pad=1 79 | activation=leaky 80 | 81 | [convolutional] 82 | batch_normalize=1 83 | filters=128 84 | size=3 85 | stride=1 86 | pad=1 87 | activation=leaky 88 | 89 | [shortcut] 90 | from=-3 91 | activation=linear 92 | 93 | [convolutional] 94 | batch_normalize=1 95 | filters=64 96 | size=1 97 | stride=1 98 | pad=1 99 | activation=leaky 100 | 101 | [convolutional] 102 | batch_normalize=1 103 | filters=128 104 | size=3 105 | stride=1 106 | pad=1 107 | activation=leaky 108 | 109 | [shortcut] 110 | from=-3 111 | activation=linear 112 | 113 | # Downsample 114 | 115 | [convolutional] 116 | batch_normalize=1 117 | filters=256 118 | size=3 119 | stride=2 120 | pad=1 121 | activation=leaky 122 | 123 | [convolutional] 124 | batch_normalize=1 125 | filters=128 126 | size=1 127 | stride=1 128 | pad=1 129 | activation=leaky 130 | 131 | [convolutional] 132 | batch_normalize=1 133 | filters=256 134 | size=3 135 | stride=1 136 | pad=1 137 | activation=leaky 138 | 139 | [shortcut] 140 | from=-3 141 | activation=linear 142 | 143 | [convolutional] 144 | batch_normalize=1 145 | filters=128 146 | size=1 147 | stride=1 148 | pad=1 149 | activation=leaky 150 | 151 | [convolutional] 152 | batch_normalize=1 153 | filters=256 154 | size=3 155 | stride=1 156 | pad=1 157 | activation=leaky 158 | 159 | [shortcut] 160 | from=-3 161 | activation=linear 162 | 163 | [convolutional] 164 | batch_normalize=1 165 | filters=128 166 | size=1 167 | stride=1 168 | pad=1 169 | activation=leaky 170 | 171 | [convolutional] 172 | batch_normalize=1 173 | filters=256 174 | size=3 175 | stride=1 176 | pad=1 177 | activation=leaky 178 | 179 | [shortcut] 180 | from=-3 181 | activation=linear 182 | 183 | [convolutional] 184 | batch_normalize=1 185 | filters=128 186 | size=1 187 | stride=1 188 | pad=1 189 | activation=leaky 190 | 191 | [convolutional] 192 | batch_normalize=1 193 | filters=256 194 | size=3 195 | stride=1 196 | pad=1 197 | activation=leaky 198 | 199 | [shortcut] 200 | from=-3 201 | activation=linear 202 | 203 | 204 | [convolutional] 205 | batch_normalize=1 206 | filters=128 207 | size=1 208 | stride=1 209 | pad=1 210 | activation=leaky 211 | 212 | [convolutional] 213 | batch_normalize=1 214 | filters=256 215 | size=3 216 | stride=1 217 | pad=1 218 | activation=leaky 219 | 220 | [shortcut] 221 | from=-3 222 | activation=linear 223 | 224 | [convolutional] 225 | batch_normalize=1 226 | filters=128 227 | size=1 228 | stride=1 229 | pad=1 230 | activation=leaky 231 | 232 | [convolutional] 233 | batch_normalize=1 234 | filters=256 235 | size=3 236 | stride=1 237 | pad=1 238 | activation=leaky 239 | 240 | [shortcut] 241 | from=-3 242 | activation=linear 243 | 244 | [convolutional] 245 | batch_normalize=1 246 | filters=128 247 | size=1 248 | stride=1 249 | pad=1 250 | activation=leaky 251 | 252 | [convolutional] 253 | batch_normalize=1 254 | filters=256 255 | size=3 256 | stride=1 257 | pad=1 258 | activation=leaky 259 | 260 | [shortcut] 261 | from=-3 262 | activation=linear 263 | 264 | [convolutional] 265 | batch_normalize=1 266 | filters=128 267 | size=1 268 | stride=1 269 | pad=1 270 | activation=leaky 271 | 272 | [convolutional] 273 | batch_normalize=1 274 | filters=256 275 | size=3 276 | stride=1 277 | pad=1 278 | activation=leaky 279 | 280 | [shortcut] 281 | from=-3 282 | activation=linear 283 | 284 | # Downsample 285 | 286 | [convolutional] 287 | batch_normalize=1 288 | filters=512 289 | size=3 290 | stride=2 291 | pad=1 292 | activation=leaky 293 | 294 | [convolutional] 295 | batch_normalize=1 296 | filters=256 297 | size=1 298 | stride=1 299 | pad=1 300 | activation=leaky 301 | 302 | [convolutional] 303 | batch_normalize=1 304 | filters=512 305 | size=3 306 | stride=1 307 | pad=1 308 | activation=leaky 309 | 310 | [shortcut] 311 | from=-3 312 | activation=linear 313 | 314 | 315 | [convolutional] 316 | batch_normalize=1 317 | filters=256 318 | size=1 319 | stride=1 320 | pad=1 321 | activation=leaky 322 | 323 | [convolutional] 324 | batch_normalize=1 325 | filters=512 326 | size=3 327 | stride=1 328 | pad=1 329 | activation=leaky 330 | 331 | [shortcut] 332 | from=-3 333 | activation=linear 334 | 335 | 336 | [convolutional] 337 | batch_normalize=1 338 | filters=256 339 | size=1 340 | stride=1 341 | pad=1 342 | activation=leaky 343 | 344 | [convolutional] 345 | batch_normalize=1 346 | filters=512 347 | size=3 348 | stride=1 349 | pad=1 350 | activation=leaky 351 | 352 | [shortcut] 353 | from=-3 354 | activation=linear 355 | 356 | 357 | [convolutional] 358 | batch_normalize=1 359 | filters=256 360 | size=1 361 | stride=1 362 | pad=1 363 | activation=leaky 364 | 365 | [convolutional] 366 | batch_normalize=1 367 | filters=512 368 | size=3 369 | stride=1 370 | pad=1 371 | activation=leaky 372 | 373 | [shortcut] 374 | from=-3 375 | activation=linear 376 | 377 | [convolutional] 378 | batch_normalize=1 379 | filters=256 380 | size=1 381 | stride=1 382 | pad=1 383 | activation=leaky 384 | 385 | [convolutional] 386 | batch_normalize=1 387 | filters=512 388 | size=3 389 | stride=1 390 | pad=1 391 | activation=leaky 392 | 393 | [shortcut] 394 | from=-3 395 | activation=linear 396 | 397 | 398 | [convolutional] 399 | batch_normalize=1 400 | filters=256 401 | size=1 402 | stride=1 403 | pad=1 404 | activation=leaky 405 | 406 | [convolutional] 407 | batch_normalize=1 408 | filters=512 409 | size=3 410 | stride=1 411 | pad=1 412 | activation=leaky 413 | 414 | [shortcut] 415 | from=-3 416 | activation=linear 417 | 418 | 419 | [convolutional] 420 | batch_normalize=1 421 | filters=256 422 | size=1 423 | stride=1 424 | pad=1 425 | activation=leaky 426 | 427 | [convolutional] 428 | batch_normalize=1 429 | filters=512 430 | size=3 431 | stride=1 432 | pad=1 433 | activation=leaky 434 | 435 | [shortcut] 436 | from=-3 437 | activation=linear 438 | 439 | [convolutional] 440 | batch_normalize=1 441 | filters=256 442 | size=1 443 | stride=1 444 | pad=1 445 | activation=leaky 446 | 447 | [convolutional] 448 | batch_normalize=1 449 | filters=512 450 | size=3 451 | stride=1 452 | pad=1 453 | activation=leaky 454 | 455 | [shortcut] 456 | from=-3 457 | activation=linear 458 | 459 | # Downsample 460 | 461 | [convolutional] 462 | batch_normalize=1 463 | filters=1024 464 | size=3 465 | stride=2 466 | pad=1 467 | activation=leaky 468 | 469 | [convolutional] 470 | batch_normalize=1 471 | filters=512 472 | size=1 473 | stride=1 474 | pad=1 475 | activation=leaky 476 | 477 | [convolutional] 478 | batch_normalize=1 479 | filters=1024 480 | size=3 481 | stride=1 482 | pad=1 483 | activation=leaky 484 | 485 | [shortcut] 486 | from=-3 487 | activation=linear 488 | 489 | [convolutional] 490 | batch_normalize=1 491 | filters=512 492 | size=1 493 | stride=1 494 | pad=1 495 | activation=leaky 496 | 497 | [convolutional] 498 | batch_normalize=1 499 | filters=1024 500 | size=3 501 | stride=1 502 | pad=1 503 | activation=leaky 504 | 505 | [shortcut] 506 | from=-3 507 | activation=linear 508 | 509 | [convolutional] 510 | batch_normalize=1 511 | filters=512 512 | size=1 513 | stride=1 514 | pad=1 515 | activation=leaky 516 | 517 | [convolutional] 518 | batch_normalize=1 519 | filters=1024 520 | size=3 521 | stride=1 522 | pad=1 523 | activation=leaky 524 | 525 | [shortcut] 526 | from=-3 527 | activation=linear 528 | 529 | [convolutional] 530 | batch_normalize=1 531 | filters=512 532 | size=1 533 | stride=1 534 | pad=1 535 | activation=leaky 536 | 537 | [convolutional] 538 | batch_normalize=1 539 | filters=1024 540 | size=3 541 | stride=1 542 | pad=1 543 | activation=leaky 544 | 545 | [shortcut] 546 | from=-3 547 | activation=linear 548 | 549 | ###################### 550 | 551 | [convolutional] 552 | batch_normalize=1 553 | filters=512 554 | size=1 555 | stride=1 556 | pad=1 557 | activation=leaky 558 | 559 | [convolutional] 560 | batch_normalize=1 561 | size=3 562 | stride=1 563 | pad=1 564 | filters=1024 565 | activation=leaky 566 | 567 | [convolutional] 568 | batch_normalize=1 569 | filters=512 570 | size=1 571 | stride=1 572 | pad=1 573 | activation=leaky 574 | 575 | [convolutional] 576 | batch_normalize=1 577 | size=3 578 | stride=1 579 | pad=1 580 | filters=1024 581 | activation=leaky 582 | 583 | [convolutional] 584 | batch_normalize=1 585 | filters=512 586 | size=1 587 | stride=1 588 | pad=1 589 | activation=leaky 590 | 591 | [convolutional] 592 | batch_normalize=1 593 | size=3 594 | stride=1 595 | pad=1 596 | filters=1024 597 | activation=leaky 598 | 599 | [convolutional] 600 | size=1 601 | stride=1 602 | pad=1 603 | filters=18 604 | activation=linear 605 | 606 | 607 | [yolo] 608 | mask = 6,7,8 609 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 610 | classes=1 611 | num=9 612 | jitter=.3 613 | ignore_thresh = .7 614 | truth_thresh = 1 615 | random=1 616 | 617 | 618 | [route] 619 | layers = -4 620 | 621 | [convolutional] 622 | batch_normalize=1 623 | filters=256 624 | size=1 625 | stride=1 626 | pad=1 627 | activation=leaky 628 | 629 | [upsample] 630 | stride=2 631 | 632 | [route] 633 | layers = -1, 61 634 | 635 | 636 | 637 | [convolutional] 638 | batch_normalize=1 639 | filters=256 640 | size=1 641 | stride=1 642 | pad=1 643 | activation=leaky 644 | 645 | [convolutional] 646 | batch_normalize=1 647 | size=3 648 | stride=1 649 | pad=1 650 | filters=512 651 | activation=leaky 652 | 653 | [convolutional] 654 | batch_normalize=1 655 | filters=256 656 | size=1 657 | stride=1 658 | pad=1 659 | activation=leaky 660 | 661 | [convolutional] 662 | batch_normalize=1 663 | size=3 664 | stride=1 665 | pad=1 666 | filters=512 667 | activation=leaky 668 | 669 | [convolutional] 670 | batch_normalize=1 671 | filters=256 672 | size=1 673 | stride=1 674 | pad=1 675 | activation=leaky 676 | 677 | [convolutional] 678 | batch_normalize=1 679 | size=3 680 | stride=1 681 | pad=1 682 | filters=512 683 | activation=leaky 684 | 685 | [convolutional] 686 | size=1 687 | stride=1 688 | pad=1 689 | filters=18 690 | activation=linear 691 | 692 | 693 | [yolo] 694 | mask = 3,4,5 695 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 696 | classes=1 697 | num=9 698 | jitter=.3 699 | ignore_thresh = .7 700 | truth_thresh = 1 701 | random=1 702 | 703 | 704 | 705 | [route] 706 | layers = -4 707 | 708 | [convolutional] 709 | batch_normalize=1 710 | filters=128 711 | size=1 712 | stride=1 713 | pad=1 714 | activation=leaky 715 | 716 | [upsample] 717 | stride=2 718 | 719 | [route] 720 | layers = -1, 36 721 | 722 | 723 | 724 | [convolutional] 725 | batch_normalize=1 726 | filters=128 727 | size=1 728 | stride=1 729 | pad=1 730 | activation=leaky 731 | 732 | [convolutional] 733 | batch_normalize=1 734 | size=3 735 | stride=1 736 | pad=1 737 | filters=256 738 | activation=leaky 739 | 740 | [convolutional] 741 | batch_normalize=1 742 | filters=128 743 | size=1 744 | stride=1 745 | pad=1 746 | activation=leaky 747 | 748 | [convolutional] 749 | batch_normalize=1 750 | size=3 751 | stride=1 752 | pad=1 753 | filters=256 754 | activation=leaky 755 | 756 | [convolutional] 757 | batch_normalize=1 758 | filters=128 759 | size=1 760 | stride=1 761 | pad=1 762 | activation=leaky 763 | 764 | [convolutional] 765 | batch_normalize=1 766 | size=3 767 | stride=1 768 | pad=1 769 | filters=256 770 | activation=leaky 771 | 772 | [convolutional] 773 | size=1 774 | stride=1 775 | pad=1 776 | filters=18 777 | activation=linear 778 | 779 | 780 | [yolo] 781 | mask = 0,1,2 782 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 783 | classes=1 784 | num=9 785 | jitter=.3 786 | ignore_thresh = .7 787 | truth_thresh = 1 788 | random=1 789 | -------------------------------------------------------------------------------- /cfg/yolov3-spp-1cls.cfg: -------------------------------------------------------------------------------- 1 | [net] 2 | # Testing 3 | # batch=1 4 | # subdivisions=1 5 | # Training 6 | batch=64 7 | subdivisions=16 8 | width=608 9 | height=608 10 | channels=3 11 | momentum=0.9 12 | decay=0.0005 13 | angle=0 14 | saturation = 1.5 15 | exposure = 1.5 16 | hue=.1 17 | 18 | learning_rate=0.001 19 | burn_in=100 20 | max_batches = 5000 21 | policy=steps 22 | steps=4000,4500 23 | scales=.1,.1 24 | 25 | [convolutional] 26 | batch_normalize=1 27 | filters=32 28 | size=3 29 | stride=1 30 | pad=1 31 | activation=leaky 32 | 33 | # Downsample 34 | 35 | [convolutional] 36 | batch_normalize=1 37 | filters=64 38 | size=3 39 | stride=2 40 | pad=1 41 | activation=leaky 42 | 43 | [convolutional] 44 | batch_normalize=1 45 | filters=32 46 | size=1 47 | stride=1 48 | pad=1 49 | activation=leaky 50 | 51 | [convolutional] 52 | batch_normalize=1 53 | filters=64 54 | size=3 55 | stride=1 56 | pad=1 57 | activation=leaky 58 | 59 | [shortcut] 60 | from=-3 61 | activation=linear 62 | 63 | # Downsample 64 | 65 | [convolutional] 66 | batch_normalize=1 67 | filters=128 68 | size=3 69 | stride=2 70 | pad=1 71 | activation=leaky 72 | 73 | [convolutional] 74 | batch_normalize=1 75 | filters=64 76 | size=1 77 | stride=1 78 | pad=1 79 | activation=leaky 80 | 81 | [convolutional] 82 | batch_normalize=1 83 | filters=128 84 | size=3 85 | stride=1 86 | pad=1 87 | activation=leaky 88 | 89 | [shortcut] 90 | from=-3 91 | activation=linear 92 | 93 | [convolutional] 94 | batch_normalize=1 95 | filters=64 96 | size=1 97 | stride=1 98 | pad=1 99 | activation=leaky 100 | 101 | [convolutional] 102 | batch_normalize=1 103 | filters=128 104 | size=3 105 | stride=1 106 | pad=1 107 | activation=leaky 108 | 109 | [shortcut] 110 | from=-3 111 | activation=linear 112 | 113 | # Downsample 114 | 115 | [convolutional] 116 | batch_normalize=1 117 | filters=256 118 | size=3 119 | stride=2 120 | pad=1 121 | activation=leaky 122 | 123 | [convolutional] 124 | batch_normalize=1 125 | filters=128 126 | size=1 127 | stride=1 128 | pad=1 129 | activation=leaky 130 | 131 | [convolutional] 132 | batch_normalize=1 133 | filters=256 134 | size=3 135 | stride=1 136 | pad=1 137 | activation=leaky 138 | 139 | [shortcut] 140 | from=-3 141 | activation=linear 142 | 143 | [convolutional] 144 | batch_normalize=1 145 | filters=128 146 | size=1 147 | stride=1 148 | pad=1 149 | activation=leaky 150 | 151 | [convolutional] 152 | batch_normalize=1 153 | filters=256 154 | size=3 155 | stride=1 156 | pad=1 157 | activation=leaky 158 | 159 | [shortcut] 160 | from=-3 161 | activation=linear 162 | 163 | [convolutional] 164 | batch_normalize=1 165 | filters=128 166 | size=1 167 | stride=1 168 | pad=1 169 | activation=leaky 170 | 171 | [convolutional] 172 | batch_normalize=1 173 | filters=256 174 | size=3 175 | stride=1 176 | pad=1 177 | activation=leaky 178 | 179 | [shortcut] 180 | from=-3 181 | activation=linear 182 | 183 | [convolutional] 184 | batch_normalize=1 185 | filters=128 186 | size=1 187 | stride=1 188 | pad=1 189 | activation=leaky 190 | 191 | [convolutional] 192 | batch_normalize=1 193 | filters=256 194 | size=3 195 | stride=1 196 | pad=1 197 | activation=leaky 198 | 199 | [shortcut] 200 | from=-3 201 | activation=linear 202 | 203 | 204 | [convolutional] 205 | batch_normalize=1 206 | filters=128 207 | size=1 208 | stride=1 209 | pad=1 210 | activation=leaky 211 | 212 | [convolutional] 213 | batch_normalize=1 214 | filters=256 215 | size=3 216 | stride=1 217 | pad=1 218 | activation=leaky 219 | 220 | [shortcut] 221 | from=-3 222 | activation=linear 223 | 224 | [convolutional] 225 | batch_normalize=1 226 | filters=128 227 | size=1 228 | stride=1 229 | pad=1 230 | activation=leaky 231 | 232 | [convolutional] 233 | batch_normalize=1 234 | filters=256 235 | size=3 236 | stride=1 237 | pad=1 238 | activation=leaky 239 | 240 | [shortcut] 241 | from=-3 242 | activation=linear 243 | 244 | [convolutional] 245 | batch_normalize=1 246 | filters=128 247 | size=1 248 | stride=1 249 | pad=1 250 | activation=leaky 251 | 252 | [convolutional] 253 | batch_normalize=1 254 | filters=256 255 | size=3 256 | stride=1 257 | pad=1 258 | activation=leaky 259 | 260 | [shortcut] 261 | from=-3 262 | activation=linear 263 | 264 | [convolutional] 265 | batch_normalize=1 266 | filters=128 267 | size=1 268 | stride=1 269 | pad=1 270 | activation=leaky 271 | 272 | [convolutional] 273 | batch_normalize=1 274 | filters=256 275 | size=3 276 | stride=1 277 | pad=1 278 | activation=leaky 279 | 280 | [shortcut] 281 | from=-3 282 | activation=linear 283 | 284 | # Downsample 285 | 286 | [convolutional] 287 | batch_normalize=1 288 | filters=512 289 | size=3 290 | stride=2 291 | pad=1 292 | activation=leaky 293 | 294 | [convolutional] 295 | batch_normalize=1 296 | filters=256 297 | size=1 298 | stride=1 299 | pad=1 300 | activation=leaky 301 | 302 | [convolutional] 303 | batch_normalize=1 304 | filters=512 305 | size=3 306 | stride=1 307 | pad=1 308 | activation=leaky 309 | 310 | [shortcut] 311 | from=-3 312 | activation=linear 313 | 314 | 315 | [convolutional] 316 | batch_normalize=1 317 | filters=256 318 | size=1 319 | stride=1 320 | pad=1 321 | activation=leaky 322 | 323 | [convolutional] 324 | batch_normalize=1 325 | filters=512 326 | size=3 327 | stride=1 328 | pad=1 329 | activation=leaky 330 | 331 | [shortcut] 332 | from=-3 333 | activation=linear 334 | 335 | 336 | [convolutional] 337 | batch_normalize=1 338 | filters=256 339 | size=1 340 | stride=1 341 | pad=1 342 | activation=leaky 343 | 344 | [convolutional] 345 | batch_normalize=1 346 | filters=512 347 | size=3 348 | stride=1 349 | pad=1 350 | activation=leaky 351 | 352 | [shortcut] 353 | from=-3 354 | activation=linear 355 | 356 | 357 | [convolutional] 358 | batch_normalize=1 359 | filters=256 360 | size=1 361 | stride=1 362 | pad=1 363 | activation=leaky 364 | 365 | [convolutional] 366 | batch_normalize=1 367 | filters=512 368 | size=3 369 | stride=1 370 | pad=1 371 | activation=leaky 372 | 373 | [shortcut] 374 | from=-3 375 | activation=linear 376 | 377 | [convolutional] 378 | batch_normalize=1 379 | filters=256 380 | size=1 381 | stride=1 382 | pad=1 383 | activation=leaky 384 | 385 | [convolutional] 386 | batch_normalize=1 387 | filters=512 388 | size=3 389 | stride=1 390 | pad=1 391 | activation=leaky 392 | 393 | [shortcut] 394 | from=-3 395 | activation=linear 396 | 397 | 398 | [convolutional] 399 | batch_normalize=1 400 | filters=256 401 | size=1 402 | stride=1 403 | pad=1 404 | activation=leaky 405 | 406 | [convolutional] 407 | batch_normalize=1 408 | filters=512 409 | size=3 410 | stride=1 411 | pad=1 412 | activation=leaky 413 | 414 | [shortcut] 415 | from=-3 416 | activation=linear 417 | 418 | 419 | [convolutional] 420 | batch_normalize=1 421 | filters=256 422 | size=1 423 | stride=1 424 | pad=1 425 | activation=leaky 426 | 427 | [convolutional] 428 | batch_normalize=1 429 | filters=512 430 | size=3 431 | stride=1 432 | pad=1 433 | activation=leaky 434 | 435 | [shortcut] 436 | from=-3 437 | activation=linear 438 | 439 | [convolutional] 440 | batch_normalize=1 441 | filters=256 442 | size=1 443 | stride=1 444 | pad=1 445 | activation=leaky 446 | 447 | [convolutional] 448 | batch_normalize=1 449 | filters=512 450 | size=3 451 | stride=1 452 | pad=1 453 | activation=leaky 454 | 455 | [shortcut] 456 | from=-3 457 | activation=linear 458 | 459 | # Downsample 460 | 461 | [convolutional] 462 | batch_normalize=1 463 | filters=1024 464 | size=3 465 | stride=2 466 | pad=1 467 | activation=leaky 468 | 469 | [convolutional] 470 | batch_normalize=1 471 | filters=512 472 | size=1 473 | stride=1 474 | pad=1 475 | activation=leaky 476 | 477 | [convolutional] 478 | batch_normalize=1 479 | filters=1024 480 | size=3 481 | stride=1 482 | pad=1 483 | activation=leaky 484 | 485 | [shortcut] 486 | from=-3 487 | activation=linear 488 | 489 | [convolutional] 490 | batch_normalize=1 491 | filters=512 492 | size=1 493 | stride=1 494 | pad=1 495 | activation=leaky 496 | 497 | [convolutional] 498 | batch_normalize=1 499 | filters=1024 500 | size=3 501 | stride=1 502 | pad=1 503 | activation=leaky 504 | 505 | [shortcut] 506 | from=-3 507 | activation=linear 508 | 509 | [convolutional] 510 | batch_normalize=1 511 | filters=512 512 | size=1 513 | stride=1 514 | pad=1 515 | activation=leaky 516 | 517 | [convolutional] 518 | batch_normalize=1 519 | filters=1024 520 | size=3 521 | stride=1 522 | pad=1 523 | activation=leaky 524 | 525 | [shortcut] 526 | from=-3 527 | activation=linear 528 | 529 | [convolutional] 530 | batch_normalize=1 531 | filters=512 532 | size=1 533 | stride=1 534 | pad=1 535 | activation=leaky 536 | 537 | [convolutional] 538 | batch_normalize=1 539 | filters=1024 540 | size=3 541 | stride=1 542 | pad=1 543 | activation=leaky 544 | 545 | [shortcut] 546 | from=-3 547 | activation=linear 548 | 549 | ###################### 550 | 551 | [convolutional] 552 | batch_normalize=1 553 | filters=512 554 | size=1 555 | stride=1 556 | pad=1 557 | activation=leaky 558 | 559 | [convolutional] 560 | batch_normalize=1 561 | size=3 562 | stride=1 563 | pad=1 564 | filters=1024 565 | activation=leaky 566 | 567 | [convolutional] 568 | batch_normalize=1 569 | filters=512 570 | size=1 571 | stride=1 572 | pad=1 573 | activation=leaky 574 | 575 | ### SPP ### 576 | [maxpool] 577 | stride=1 578 | size=5 579 | 580 | [route] 581 | layers=-2 582 | 583 | [maxpool] 584 | stride=1 585 | size=9 586 | 587 | [route] 588 | layers=-4 589 | 590 | [maxpool] 591 | stride=1 592 | size=13 593 | 594 | [route] 595 | layers=-1,-3,-5,-6 596 | 597 | ### End SPP ### 598 | 599 | [convolutional] 600 | batch_normalize=1 601 | filters=512 602 | size=1 603 | stride=1 604 | pad=1 605 | activation=leaky 606 | 607 | 608 | [convolutional] 609 | batch_normalize=1 610 | size=3 611 | stride=1 612 | pad=1 613 | filters=1024 614 | activation=leaky 615 | 616 | [convolutional] 617 | batch_normalize=1 618 | filters=512 619 | size=1 620 | stride=1 621 | pad=1 622 | activation=leaky 623 | 624 | [convolutional] 625 | batch_normalize=1 626 | size=3 627 | stride=1 628 | pad=1 629 | filters=1024 630 | activation=leaky 631 | 632 | [convolutional] 633 | size=1 634 | stride=1 635 | pad=1 636 | filters=18 637 | activation=linear 638 | 639 | 640 | [yolo] 641 | mask = 6,7,8 642 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 643 | classes=1 644 | num=9 645 | jitter=.3 646 | ignore_thresh = .7 647 | truth_thresh = 1 648 | random=1 649 | 650 | 651 | [route] 652 | layers = -4 653 | 654 | [convolutional] 655 | batch_normalize=1 656 | filters=256 657 | size=1 658 | stride=1 659 | pad=1 660 | activation=leaky 661 | 662 | [upsample] 663 | stride=2 664 | 665 | [route] 666 | layers = -1, 61 667 | 668 | 669 | 670 | [convolutional] 671 | batch_normalize=1 672 | filters=256 673 | size=1 674 | stride=1 675 | pad=1 676 | activation=leaky 677 | 678 | [convolutional] 679 | batch_normalize=1 680 | size=3 681 | stride=1 682 | pad=1 683 | filters=512 684 | activation=leaky 685 | 686 | [convolutional] 687 | batch_normalize=1 688 | filters=256 689 | size=1 690 | stride=1 691 | pad=1 692 | activation=leaky 693 | 694 | [convolutional] 695 | batch_normalize=1 696 | size=3 697 | stride=1 698 | pad=1 699 | filters=512 700 | activation=leaky 701 | 702 | [convolutional] 703 | batch_normalize=1 704 | filters=256 705 | size=1 706 | stride=1 707 | pad=1 708 | activation=leaky 709 | 710 | [convolutional] 711 | batch_normalize=1 712 | size=3 713 | stride=1 714 | pad=1 715 | filters=512 716 | activation=leaky 717 | 718 | [convolutional] 719 | size=1 720 | stride=1 721 | pad=1 722 | filters=18 723 | activation=linear 724 | 725 | 726 | [yolo] 727 | mask = 3,4,5 728 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 729 | classes=1 730 | num=9 731 | jitter=.3 732 | ignore_thresh = .7 733 | truth_thresh = 1 734 | random=1 735 | 736 | 737 | 738 | [route] 739 | layers = -4 740 | 741 | [convolutional] 742 | batch_normalize=1 743 | filters=128 744 | size=1 745 | stride=1 746 | pad=1 747 | activation=leaky 748 | 749 | [upsample] 750 | stride=2 751 | 752 | [route] 753 | layers = -1, 36 754 | 755 | 756 | 757 | [convolutional] 758 | batch_normalize=1 759 | filters=128 760 | size=1 761 | stride=1 762 | pad=1 763 | activation=leaky 764 | 765 | [convolutional] 766 | batch_normalize=1 767 | size=3 768 | stride=1 769 | pad=1 770 | filters=256 771 | activation=leaky 772 | 773 | [convolutional] 774 | batch_normalize=1 775 | filters=128 776 | size=1 777 | stride=1 778 | pad=1 779 | activation=leaky 780 | 781 | [convolutional] 782 | batch_normalize=1 783 | size=3 784 | stride=1 785 | pad=1 786 | filters=256 787 | activation=leaky 788 | 789 | [convolutional] 790 | batch_normalize=1 791 | filters=128 792 | size=1 793 | stride=1 794 | pad=1 795 | activation=leaky 796 | 797 | [convolutional] 798 | batch_normalize=1 799 | size=3 800 | stride=1 801 | pad=1 802 | filters=256 803 | activation=leaky 804 | 805 | [convolutional] 806 | size=1 807 | stride=1 808 | pad=1 809 | filters=18 810 | activation=linear 811 | 812 | 813 | [yolo] 814 | mask = 0,1,2 815 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 816 | classes=1 817 | num=9 818 | jitter=.3 819 | ignore_thresh = .7 820 | truth_thresh = 1 821 | random=1 822 | -------------------------------------------------------------------------------- /cfg/yolov3-spp.cfg: -------------------------------------------------------------------------------- 1 | [net] 2 | # Testing 3 | # batch=1 4 | # subdivisions=1 5 | # Training 6 | batch=64 7 | subdivisions=16 8 | width=608 9 | height=608 10 | channels=3 11 | momentum=0.9 12 | decay=0.0005 13 | angle=0 14 | saturation = 1.5 15 | exposure = 1.5 16 | hue=.1 17 | 18 | learning_rate=0.001 19 | burn_in=1000 20 | max_batches = 500200 21 | policy=steps 22 | steps=400000,450000 23 | scales=.1,.1 24 | 25 | [convolutional] 26 | batch_normalize=1 27 | filters=32 28 | size=3 29 | stride=1 30 | pad=1 31 | activation=leaky 32 | 33 | # Downsample 34 | 35 | [convolutional] 36 | batch_normalize=1 37 | filters=64 38 | size=3 39 | stride=2 40 | pad=1 41 | activation=leaky 42 | 43 | [convolutional] 44 | batch_normalize=1 45 | filters=32 46 | size=1 47 | stride=1 48 | pad=1 49 | activation=leaky 50 | 51 | [convolutional] 52 | batch_normalize=1 53 | filters=64 54 | size=3 55 | stride=1 56 | pad=1 57 | activation=leaky 58 | 59 | [shortcut] 60 | from=-3 61 | activation=linear 62 | 63 | # Downsample 64 | 65 | [convolutional] 66 | batch_normalize=1 67 | filters=128 68 | size=3 69 | stride=2 70 | pad=1 71 | activation=leaky 72 | 73 | [convolutional] 74 | batch_normalize=1 75 | filters=64 76 | size=1 77 | stride=1 78 | pad=1 79 | activation=leaky 80 | 81 | [convolutional] 82 | batch_normalize=1 83 | filters=128 84 | size=3 85 | stride=1 86 | pad=1 87 | activation=leaky 88 | 89 | [shortcut] 90 | from=-3 91 | activation=linear 92 | 93 | [convolutional] 94 | batch_normalize=1 95 | filters=64 96 | size=1 97 | stride=1 98 | pad=1 99 | activation=leaky 100 | 101 | [convolutional] 102 | batch_normalize=1 103 | filters=128 104 | size=3 105 | stride=1 106 | pad=1 107 | activation=leaky 108 | 109 | [shortcut] 110 | from=-3 111 | activation=linear 112 | 113 | # Downsample 114 | 115 | [convolutional] 116 | batch_normalize=1 117 | filters=256 118 | size=3 119 | stride=2 120 | pad=1 121 | activation=leaky 122 | 123 | [convolutional] 124 | batch_normalize=1 125 | filters=128 126 | size=1 127 | stride=1 128 | pad=1 129 | activation=leaky 130 | 131 | [convolutional] 132 | batch_normalize=1 133 | filters=256 134 | size=3 135 | stride=1 136 | pad=1 137 | activation=leaky 138 | 139 | [shortcut] 140 | from=-3 141 | activation=linear 142 | 143 | [convolutional] 144 | batch_normalize=1 145 | filters=128 146 | size=1 147 | stride=1 148 | pad=1 149 | activation=leaky 150 | 151 | [convolutional] 152 | batch_normalize=1 153 | filters=256 154 | size=3 155 | stride=1 156 | pad=1 157 | activation=leaky 158 | 159 | [shortcut] 160 | from=-3 161 | activation=linear 162 | 163 | [convolutional] 164 | batch_normalize=1 165 | filters=128 166 | size=1 167 | stride=1 168 | pad=1 169 | activation=leaky 170 | 171 | [convolutional] 172 | batch_normalize=1 173 | filters=256 174 | size=3 175 | stride=1 176 | pad=1 177 | activation=leaky 178 | 179 | [shortcut] 180 | from=-3 181 | activation=linear 182 | 183 | [convolutional] 184 | batch_normalize=1 185 | filters=128 186 | size=1 187 | stride=1 188 | pad=1 189 | activation=leaky 190 | 191 | [convolutional] 192 | batch_normalize=1 193 | filters=256 194 | size=3 195 | stride=1 196 | pad=1 197 | activation=leaky 198 | 199 | [shortcut] 200 | from=-3 201 | activation=linear 202 | 203 | 204 | [convolutional] 205 | batch_normalize=1 206 | filters=128 207 | size=1 208 | stride=1 209 | pad=1 210 | activation=leaky 211 | 212 | [convolutional] 213 | batch_normalize=1 214 | filters=256 215 | size=3 216 | stride=1 217 | pad=1 218 | activation=leaky 219 | 220 | [shortcut] 221 | from=-3 222 | activation=linear 223 | 224 | [convolutional] 225 | batch_normalize=1 226 | filters=128 227 | size=1 228 | stride=1 229 | pad=1 230 | activation=leaky 231 | 232 | [convolutional] 233 | batch_normalize=1 234 | filters=256 235 | size=3 236 | stride=1 237 | pad=1 238 | activation=leaky 239 | 240 | [shortcut] 241 | from=-3 242 | activation=linear 243 | 244 | [convolutional] 245 | batch_normalize=1 246 | filters=128 247 | size=1 248 | stride=1 249 | pad=1 250 | activation=leaky 251 | 252 | [convolutional] 253 | batch_normalize=1 254 | filters=256 255 | size=3 256 | stride=1 257 | pad=1 258 | activation=leaky 259 | 260 | [shortcut] 261 | from=-3 262 | activation=linear 263 | 264 | [convolutional] 265 | batch_normalize=1 266 | filters=128 267 | size=1 268 | stride=1 269 | pad=1 270 | activation=leaky 271 | 272 | [convolutional] 273 | batch_normalize=1 274 | filters=256 275 | size=3 276 | stride=1 277 | pad=1 278 | activation=leaky 279 | 280 | [shortcut] 281 | from=-3 282 | activation=linear 283 | 284 | # Downsample 285 | 286 | [convolutional] 287 | batch_normalize=1 288 | filters=512 289 | size=3 290 | stride=2 291 | pad=1 292 | activation=leaky 293 | 294 | [convolutional] 295 | batch_normalize=1 296 | filters=256 297 | size=1 298 | stride=1 299 | pad=1 300 | activation=leaky 301 | 302 | [convolutional] 303 | batch_normalize=1 304 | filters=512 305 | size=3 306 | stride=1 307 | pad=1 308 | activation=leaky 309 | 310 | [shortcut] 311 | from=-3 312 | activation=linear 313 | 314 | 315 | [convolutional] 316 | batch_normalize=1 317 | filters=256 318 | size=1 319 | stride=1 320 | pad=1 321 | activation=leaky 322 | 323 | [convolutional] 324 | batch_normalize=1 325 | filters=512 326 | size=3 327 | stride=1 328 | pad=1 329 | activation=leaky 330 | 331 | [shortcut] 332 | from=-3 333 | activation=linear 334 | 335 | 336 | [convolutional] 337 | batch_normalize=1 338 | filters=256 339 | size=1 340 | stride=1 341 | pad=1 342 | activation=leaky 343 | 344 | [convolutional] 345 | batch_normalize=1 346 | filters=512 347 | size=3 348 | stride=1 349 | pad=1 350 | activation=leaky 351 | 352 | [shortcut] 353 | from=-3 354 | activation=linear 355 | 356 | 357 | [convolutional] 358 | batch_normalize=1 359 | filters=256 360 | size=1 361 | stride=1 362 | pad=1 363 | activation=leaky 364 | 365 | [convolutional] 366 | batch_normalize=1 367 | filters=512 368 | size=3 369 | stride=1 370 | pad=1 371 | activation=leaky 372 | 373 | [shortcut] 374 | from=-3 375 | activation=linear 376 | 377 | [convolutional] 378 | batch_normalize=1 379 | filters=256 380 | size=1 381 | stride=1 382 | pad=1 383 | activation=leaky 384 | 385 | [convolutional] 386 | batch_normalize=1 387 | filters=512 388 | size=3 389 | stride=1 390 | pad=1 391 | activation=leaky 392 | 393 | [shortcut] 394 | from=-3 395 | activation=linear 396 | 397 | 398 | [convolutional] 399 | batch_normalize=1 400 | filters=256 401 | size=1 402 | stride=1 403 | pad=1 404 | activation=leaky 405 | 406 | [convolutional] 407 | batch_normalize=1 408 | filters=512 409 | size=3 410 | stride=1 411 | pad=1 412 | activation=leaky 413 | 414 | [shortcut] 415 | from=-3 416 | activation=linear 417 | 418 | 419 | [convolutional] 420 | batch_normalize=1 421 | filters=256 422 | size=1 423 | stride=1 424 | pad=1 425 | activation=leaky 426 | 427 | [convolutional] 428 | batch_normalize=1 429 | filters=512 430 | size=3 431 | stride=1 432 | pad=1 433 | activation=leaky 434 | 435 | [shortcut] 436 | from=-3 437 | activation=linear 438 | 439 | [convolutional] 440 | batch_normalize=1 441 | filters=256 442 | size=1 443 | stride=1 444 | pad=1 445 | activation=leaky 446 | 447 | [convolutional] 448 | batch_normalize=1 449 | filters=512 450 | size=3 451 | stride=1 452 | pad=1 453 | activation=leaky 454 | 455 | [shortcut] 456 | from=-3 457 | activation=linear 458 | 459 | # Downsample 460 | 461 | [convolutional] 462 | batch_normalize=1 463 | filters=1024 464 | size=3 465 | stride=2 466 | pad=1 467 | activation=leaky 468 | 469 | [convolutional] 470 | batch_normalize=1 471 | filters=512 472 | size=1 473 | stride=1 474 | pad=1 475 | activation=leaky 476 | 477 | [convolutional] 478 | batch_normalize=1 479 | filters=1024 480 | size=3 481 | stride=1 482 | pad=1 483 | activation=leaky 484 | 485 | [shortcut] 486 | from=-3 487 | activation=linear 488 | 489 | [convolutional] 490 | batch_normalize=1 491 | filters=512 492 | size=1 493 | stride=1 494 | pad=1 495 | activation=leaky 496 | 497 | [convolutional] 498 | batch_normalize=1 499 | filters=1024 500 | size=3 501 | stride=1 502 | pad=1 503 | activation=leaky 504 | 505 | [shortcut] 506 | from=-3 507 | activation=linear 508 | 509 | [convolutional] 510 | batch_normalize=1 511 | filters=512 512 | size=1 513 | stride=1 514 | pad=1 515 | activation=leaky 516 | 517 | [convolutional] 518 | batch_normalize=1 519 | filters=1024 520 | size=3 521 | stride=1 522 | pad=1 523 | activation=leaky 524 | 525 | [shortcut] 526 | from=-3 527 | activation=linear 528 | 529 | [convolutional] 530 | batch_normalize=1 531 | filters=512 532 | size=1 533 | stride=1 534 | pad=1 535 | activation=leaky 536 | 537 | [convolutional] 538 | batch_normalize=1 539 | filters=1024 540 | size=3 541 | stride=1 542 | pad=1 543 | activation=leaky 544 | 545 | [shortcut] 546 | from=-3 547 | activation=linear 548 | 549 | ###################### 550 | 551 | [convolutional] 552 | batch_normalize=1 553 | filters=512 554 | size=1 555 | stride=1 556 | pad=1 557 | activation=leaky 558 | 559 | [convolutional] 560 | batch_normalize=1 561 | size=3 562 | stride=1 563 | pad=1 564 | filters=1024 565 | activation=leaky 566 | 567 | [convolutional] 568 | batch_normalize=1 569 | filters=512 570 | size=1 571 | stride=1 572 | pad=1 573 | activation=leaky 574 | 575 | ### SPP ### 576 | [maxpool] 577 | stride=1 578 | size=5 579 | 580 | [route] 581 | layers=-2 582 | 583 | [maxpool] 584 | stride=1 585 | size=9 586 | 587 | [route] 588 | layers=-4 589 | 590 | [maxpool] 591 | stride=1 592 | size=13 593 | 594 | [route] 595 | layers=-1,-3,-5,-6 596 | 597 | ### End SPP ### 598 | 599 | [convolutional] 600 | batch_normalize=1 601 | filters=512 602 | size=1 603 | stride=1 604 | pad=1 605 | activation=leaky 606 | 607 | 608 | [convolutional] 609 | batch_normalize=1 610 | size=3 611 | stride=1 612 | pad=1 613 | filters=1024 614 | activation=leaky 615 | 616 | [convolutional] 617 | batch_normalize=1 618 | filters=512 619 | size=1 620 | stride=1 621 | pad=1 622 | activation=leaky 623 | 624 | [convolutional] 625 | batch_normalize=1 626 | size=3 627 | stride=1 628 | pad=1 629 | filters=1024 630 | activation=leaky 631 | 632 | [convolutional] 633 | size=1 634 | stride=1 635 | pad=1 636 | filters=255 637 | activation=linear 638 | 639 | 640 | [yolo] 641 | mask = 6,7,8 642 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 643 | classes=80 644 | num=9 645 | jitter=.3 646 | ignore_thresh = .7 647 | truth_thresh = 1 648 | random=1 649 | 650 | 651 | [route] 652 | layers = -4 653 | 654 | [convolutional] 655 | batch_normalize=1 656 | filters=256 657 | size=1 658 | stride=1 659 | pad=1 660 | activation=leaky 661 | 662 | [upsample] 663 | stride=2 664 | 665 | [route] 666 | layers = -1, 61 667 | 668 | 669 | 670 | [convolutional] 671 | batch_normalize=1 672 | filters=256 673 | size=1 674 | stride=1 675 | pad=1 676 | activation=leaky 677 | 678 | [convolutional] 679 | batch_normalize=1 680 | size=3 681 | stride=1 682 | pad=1 683 | filters=512 684 | activation=leaky 685 | 686 | [convolutional] 687 | batch_normalize=1 688 | filters=256 689 | size=1 690 | stride=1 691 | pad=1 692 | activation=leaky 693 | 694 | [convolutional] 695 | batch_normalize=1 696 | size=3 697 | stride=1 698 | pad=1 699 | filters=512 700 | activation=leaky 701 | 702 | [convolutional] 703 | batch_normalize=1 704 | filters=256 705 | size=1 706 | stride=1 707 | pad=1 708 | activation=leaky 709 | 710 | [convolutional] 711 | batch_normalize=1 712 | size=3 713 | stride=1 714 | pad=1 715 | filters=512 716 | activation=leaky 717 | 718 | [convolutional] 719 | size=1 720 | stride=1 721 | pad=1 722 | filters=255 723 | activation=linear 724 | 725 | 726 | [yolo] 727 | mask = 3,4,5 728 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 729 | classes=80 730 | num=9 731 | jitter=.3 732 | ignore_thresh = .7 733 | truth_thresh = 1 734 | random=1 735 | 736 | 737 | 738 | [route] 739 | layers = -4 740 | 741 | [convolutional] 742 | batch_normalize=1 743 | filters=128 744 | size=1 745 | stride=1 746 | pad=1 747 | activation=leaky 748 | 749 | [upsample] 750 | stride=2 751 | 752 | [route] 753 | layers = -1, 36 754 | 755 | 756 | 757 | [convolutional] 758 | batch_normalize=1 759 | filters=128 760 | size=1 761 | stride=1 762 | pad=1 763 | activation=leaky 764 | 765 | [convolutional] 766 | batch_normalize=1 767 | size=3 768 | stride=1 769 | pad=1 770 | filters=256 771 | activation=leaky 772 | 773 | [convolutional] 774 | batch_normalize=1 775 | filters=128 776 | size=1 777 | stride=1 778 | pad=1 779 | activation=leaky 780 | 781 | [convolutional] 782 | batch_normalize=1 783 | size=3 784 | stride=1 785 | pad=1 786 | filters=256 787 | activation=leaky 788 | 789 | [convolutional] 790 | batch_normalize=1 791 | filters=128 792 | size=1 793 | stride=1 794 | pad=1 795 | activation=leaky 796 | 797 | [convolutional] 798 | batch_normalize=1 799 | size=3 800 | stride=1 801 | pad=1 802 | filters=256 803 | activation=leaky 804 | 805 | [convolutional] 806 | size=1 807 | stride=1 808 | pad=1 809 | filters=255 810 | activation=linear 811 | 812 | 813 | [yolo] 814 | mask = 0,1,2 815 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 816 | classes=80 817 | num=9 818 | jitter=.3 819 | ignore_thresh = .7 820 | truth_thresh = 1 821 | random=1 822 | -------------------------------------------------------------------------------- /cfg/yolov3-tiny-1cls.cfg: -------------------------------------------------------------------------------- 1 | [net] 2 | # Testing 3 | batch=1 4 | subdivisions=1 5 | # Training 6 | # batch=64 7 | # subdivisions=2 8 | width=416 9 | height=416 10 | channels=3 11 | momentum=0.9 12 | decay=0.0005 13 | angle=0 14 | saturation = 1.5 15 | exposure = 1.5 16 | hue=.1 17 | 18 | learning_rate=0.001 19 | burn_in=1000 20 | max_batches = 500200 21 | policy=steps 22 | steps=400000,450000 23 | scales=.1,.1 24 | 25 | [convolutional] 26 | batch_normalize=1 27 | filters=16 28 | size=3 29 | stride=1 30 | pad=1 31 | activation=leaky 32 | 33 | [maxpool] 34 | size=2 35 | stride=2 36 | 37 | [convolutional] 38 | batch_normalize=1 39 | filters=32 40 | size=3 41 | stride=1 42 | pad=1 43 | activation=leaky 44 | 45 | [maxpool] 46 | size=2 47 | stride=2 48 | 49 | [convolutional] 50 | batch_normalize=1 51 | filters=64 52 | size=3 53 | stride=1 54 | pad=1 55 | activation=leaky 56 | 57 | [maxpool] 58 | size=2 59 | stride=2 60 | 61 | [convolutional] 62 | batch_normalize=1 63 | filters=128 64 | size=3 65 | stride=1 66 | pad=1 67 | activation=leaky 68 | 69 | [maxpool] 70 | size=2 71 | stride=2 72 | 73 | [convolutional] 74 | batch_normalize=1 75 | filters=256 76 | size=3 77 | stride=1 78 | pad=1 79 | activation=leaky 80 | 81 | [maxpool] 82 | size=2 83 | stride=2 84 | 85 | [convolutional] 86 | batch_normalize=1 87 | filters=512 88 | size=3 89 | stride=1 90 | pad=1 91 | activation=leaky 92 | 93 | [maxpool] 94 | size=2 95 | stride=1 96 | 97 | [convolutional] 98 | batch_normalize=1 99 | filters=1024 100 | size=3 101 | stride=1 102 | pad=1 103 | activation=leaky 104 | 105 | ########### 106 | 107 | [convolutional] 108 | batch_normalize=1 109 | filters=256 110 | size=1 111 | stride=1 112 | pad=1 113 | activation=leaky 114 | 115 | [convolutional] 116 | batch_normalize=1 117 | filters=512 118 | size=3 119 | stride=1 120 | pad=1 121 | activation=leaky 122 | 123 | [convolutional] 124 | size=1 125 | stride=1 126 | pad=1 127 | filters=18 128 | activation=linear 129 | 130 | 131 | 132 | [yolo] 133 | mask = 3,4,5 134 | anchors = 10,14, 23,27, 37,58, 81,82, 135,169, 344,319 135 | classes=1 136 | num=6 137 | jitter=.3 138 | ignore_thresh = .7 139 | truth_thresh = 1 140 | random=1 141 | 142 | [route] 143 | layers = -4 144 | 145 | [convolutional] 146 | batch_normalize=1 147 | filters=128 148 | size=1 149 | stride=1 150 | pad=1 151 | activation=leaky 152 | 153 | [upsample] 154 | stride=2 155 | 156 | [route] 157 | layers = -1, 8 158 | 159 | [convolutional] 160 | batch_normalize=1 161 | filters=256 162 | size=3 163 | stride=1 164 | pad=1 165 | activation=leaky 166 | 167 | [convolutional] 168 | size=1 169 | stride=1 170 | pad=1 171 | filters=18 172 | activation=linear 173 | 174 | [yolo] 175 | mask = 0,1,2 176 | anchors = 10,14, 23,27, 37,58, 81,82, 135,169, 344,319 177 | classes=1 178 | num=6 179 | jitter=.3 180 | ignore_thresh = .7 181 | truth_thresh = 1 182 | random=1 183 | -------------------------------------------------------------------------------- /cfg/yolov3-tiny.cfg: -------------------------------------------------------------------------------- 1 | [net] 2 | # Testing 3 | batch=1 4 | subdivisions=1 5 | # Training 6 | # batch=64 7 | # subdivisions=2 8 | width=416 9 | height=416 10 | channels=3 11 | momentum=0.9 12 | decay=0.0005 13 | angle=0 14 | saturation = 1.5 15 | exposure = 1.5 16 | hue=.1 17 | 18 | learning_rate=0.001 19 | burn_in=1000 20 | max_batches = 500200 21 | policy=steps 22 | steps=400000,450000 23 | scales=.1,.1 24 | 25 | [convolutional] 26 | batch_normalize=1 27 | filters=16 28 | size=3 29 | stride=1 30 | pad=1 31 | activation=leaky 32 | 33 | [maxpool] 34 | size=2 35 | stride=2 36 | 37 | [convolutional] 38 | batch_normalize=1 39 | filters=32 40 | size=3 41 | stride=1 42 | pad=1 43 | activation=leaky 44 | 45 | [maxpool] 46 | size=2 47 | stride=2 48 | 49 | [convolutional] 50 | batch_normalize=1 51 | filters=64 52 | size=3 53 | stride=1 54 | pad=1 55 | activation=leaky 56 | 57 | [maxpool] 58 | size=2 59 | stride=2 60 | 61 | [convolutional] 62 | batch_normalize=1 63 | filters=128 64 | size=3 65 | stride=1 66 | pad=1 67 | activation=leaky 68 | 69 | [maxpool] 70 | size=2 71 | stride=2 72 | 73 | [convolutional] 74 | batch_normalize=1 75 | filters=256 76 | size=3 77 | stride=1 78 | pad=1 79 | activation=leaky 80 | 81 | [maxpool] 82 | size=2 83 | stride=2 84 | 85 | [convolutional] 86 | batch_normalize=1 87 | filters=512 88 | size=3 89 | stride=1 90 | pad=1 91 | activation=leaky 92 | 93 | [maxpool] 94 | size=2 95 | stride=1 96 | 97 | [convolutional] 98 | batch_normalize=1 99 | filters=1024 100 | size=3 101 | stride=1 102 | pad=1 103 | activation=leaky 104 | 105 | ########### 106 | 107 | [convolutional] 108 | batch_normalize=1 109 | filters=256 110 | size=1 111 | stride=1 112 | pad=1 113 | activation=leaky 114 | 115 | [convolutional] 116 | batch_normalize=1 117 | filters=512 118 | size=3 119 | stride=1 120 | pad=1 121 | activation=leaky 122 | 123 | [convolutional] 124 | size=1 125 | stride=1 126 | pad=1 127 | filters=255 128 | activation=linear 129 | 130 | 131 | 132 | [yolo] 133 | mask = 3,4,5 134 | anchors = 10,14, 23,27, 37,58, 81,82, 135,169, 344,319 135 | classes=80 136 | num=6 137 | jitter=.3 138 | ignore_thresh = .7 139 | truth_thresh = 1 140 | random=1 141 | 142 | [route] 143 | layers = -4 144 | 145 | [convolutional] 146 | batch_normalize=1 147 | filters=128 148 | size=1 149 | stride=1 150 | pad=1 151 | activation=leaky 152 | 153 | [upsample] 154 | stride=2 155 | 156 | [route] 157 | layers = -1, 8 158 | 159 | [convolutional] 160 | batch_normalize=1 161 | filters=256 162 | size=3 163 | stride=1 164 | pad=1 165 | activation=leaky 166 | 167 | [convolutional] 168 | size=1 169 | stride=1 170 | pad=1 171 | filters=255 172 | activation=linear 173 | 174 | [yolo] 175 | mask = 1,2,3 176 | anchors = 10,14, 23,27, 37,58, 81,82, 135,169, 344,319 177 | classes=80 178 | num=6 179 | jitter=.3 180 | ignore_thresh = .7 181 | truth_thresh = 1 182 | random=1 183 | -------------------------------------------------------------------------------- /cfg/yolov3.cfg: -------------------------------------------------------------------------------- 1 | [net] 2 | # Testing 3 | #batch=1 4 | #subdivisions=1 5 | # Training 6 | batch=16 7 | subdivisions=1 8 | width=416 9 | height=416 10 | channels=3 11 | momentum=0.9 12 | decay=0.0005 13 | angle=0 14 | saturation = 1.5 15 | exposure = 1.5 16 | hue=.1 17 | 18 | learning_rate=0.001 19 | burn_in=1000 20 | max_batches = 500200 21 | policy=steps 22 | steps=400000,450000 23 | scales=.1,.1 24 | 25 | [convolutional] 26 | batch_normalize=1 27 | filters=32 28 | size=3 29 | stride=1 30 | pad=1 31 | activation=leaky 32 | 33 | # Downsample 34 | 35 | [convolutional] 36 | batch_normalize=1 37 | filters=64 38 | size=3 39 | stride=2 40 | pad=1 41 | activation=leaky 42 | 43 | [convolutional] 44 | batch_normalize=1 45 | filters=32 46 | size=1 47 | stride=1 48 | pad=1 49 | activation=leaky 50 | 51 | [convolutional] 52 | batch_normalize=1 53 | filters=64 54 | size=3 55 | stride=1 56 | pad=1 57 | activation=leaky 58 | 59 | [shortcut] 60 | from=-3 61 | activation=linear 62 | 63 | # Downsample 64 | 65 | [convolutional] 66 | batch_normalize=1 67 | filters=128 68 | size=3 69 | stride=2 70 | pad=1 71 | activation=leaky 72 | 73 | [convolutional] 74 | batch_normalize=1 75 | filters=64 76 | size=1 77 | stride=1 78 | pad=1 79 | activation=leaky 80 | 81 | [convolutional] 82 | batch_normalize=1 83 | filters=128 84 | size=3 85 | stride=1 86 | pad=1 87 | activation=leaky 88 | 89 | [shortcut] 90 | from=-3 91 | activation=linear 92 | 93 | [convolutional] 94 | batch_normalize=1 95 | filters=64 96 | size=1 97 | stride=1 98 | pad=1 99 | activation=leaky 100 | 101 | [convolutional] 102 | batch_normalize=1 103 | filters=128 104 | size=3 105 | stride=1 106 | pad=1 107 | activation=leaky 108 | 109 | [shortcut] 110 | from=-3 111 | activation=linear 112 | 113 | # Downsample 114 | 115 | [convolutional] 116 | batch_normalize=1 117 | filters=256 118 | size=3 119 | stride=2 120 | pad=1 121 | activation=leaky 122 | 123 | [convolutional] 124 | batch_normalize=1 125 | filters=128 126 | size=1 127 | stride=1 128 | pad=1 129 | activation=leaky 130 | 131 | [convolutional] 132 | batch_normalize=1 133 | filters=256 134 | size=3 135 | stride=1 136 | pad=1 137 | activation=leaky 138 | 139 | [shortcut] 140 | from=-3 141 | activation=linear 142 | 143 | [convolutional] 144 | batch_normalize=1 145 | filters=128 146 | size=1 147 | stride=1 148 | pad=1 149 | activation=leaky 150 | 151 | [convolutional] 152 | batch_normalize=1 153 | filters=256 154 | size=3 155 | stride=1 156 | pad=1 157 | activation=leaky 158 | 159 | [shortcut] 160 | from=-3 161 | activation=linear 162 | 163 | [convolutional] 164 | batch_normalize=1 165 | filters=128 166 | size=1 167 | stride=1 168 | pad=1 169 | activation=leaky 170 | 171 | [convolutional] 172 | batch_normalize=1 173 | filters=256 174 | size=3 175 | stride=1 176 | pad=1 177 | activation=leaky 178 | 179 | [shortcut] 180 | from=-3 181 | activation=linear 182 | 183 | [convolutional] 184 | batch_normalize=1 185 | filters=128 186 | size=1 187 | stride=1 188 | pad=1 189 | activation=leaky 190 | 191 | [convolutional] 192 | batch_normalize=1 193 | filters=256 194 | size=3 195 | stride=1 196 | pad=1 197 | activation=leaky 198 | 199 | [shortcut] 200 | from=-3 201 | activation=linear 202 | 203 | 204 | [convolutional] 205 | batch_normalize=1 206 | filters=128 207 | size=1 208 | stride=1 209 | pad=1 210 | activation=leaky 211 | 212 | [convolutional] 213 | batch_normalize=1 214 | filters=256 215 | size=3 216 | stride=1 217 | pad=1 218 | activation=leaky 219 | 220 | [shortcut] 221 | from=-3 222 | activation=linear 223 | 224 | [convolutional] 225 | batch_normalize=1 226 | filters=128 227 | size=1 228 | stride=1 229 | pad=1 230 | activation=leaky 231 | 232 | [convolutional] 233 | batch_normalize=1 234 | filters=256 235 | size=3 236 | stride=1 237 | pad=1 238 | activation=leaky 239 | 240 | [shortcut] 241 | from=-3 242 | activation=linear 243 | 244 | [convolutional] 245 | batch_normalize=1 246 | filters=128 247 | size=1 248 | stride=1 249 | pad=1 250 | activation=leaky 251 | 252 | [convolutional] 253 | batch_normalize=1 254 | filters=256 255 | size=3 256 | stride=1 257 | pad=1 258 | activation=leaky 259 | 260 | [shortcut] 261 | from=-3 262 | activation=linear 263 | 264 | [convolutional] 265 | batch_normalize=1 266 | filters=128 267 | size=1 268 | stride=1 269 | pad=1 270 | activation=leaky 271 | 272 | [convolutional] 273 | batch_normalize=1 274 | filters=256 275 | size=3 276 | stride=1 277 | pad=1 278 | activation=leaky 279 | 280 | [shortcut] 281 | from=-3 282 | activation=linear 283 | 284 | # Downsample 285 | 286 | [convolutional] 287 | batch_normalize=1 288 | filters=512 289 | size=3 290 | stride=2 291 | pad=1 292 | activation=leaky 293 | 294 | [convolutional] 295 | batch_normalize=1 296 | filters=256 297 | size=1 298 | stride=1 299 | pad=1 300 | activation=leaky 301 | 302 | [convolutional] 303 | batch_normalize=1 304 | filters=512 305 | size=3 306 | stride=1 307 | pad=1 308 | activation=leaky 309 | 310 | [shortcut] 311 | from=-3 312 | activation=linear 313 | 314 | 315 | [convolutional] 316 | batch_normalize=1 317 | filters=256 318 | size=1 319 | stride=1 320 | pad=1 321 | activation=leaky 322 | 323 | [convolutional] 324 | batch_normalize=1 325 | filters=512 326 | size=3 327 | stride=1 328 | pad=1 329 | activation=leaky 330 | 331 | [shortcut] 332 | from=-3 333 | activation=linear 334 | 335 | 336 | [convolutional] 337 | batch_normalize=1 338 | filters=256 339 | size=1 340 | stride=1 341 | pad=1 342 | activation=leaky 343 | 344 | [convolutional] 345 | batch_normalize=1 346 | filters=512 347 | size=3 348 | stride=1 349 | pad=1 350 | activation=leaky 351 | 352 | [shortcut] 353 | from=-3 354 | activation=linear 355 | 356 | 357 | [convolutional] 358 | batch_normalize=1 359 | filters=256 360 | size=1 361 | stride=1 362 | pad=1 363 | activation=leaky 364 | 365 | [convolutional] 366 | batch_normalize=1 367 | filters=512 368 | size=3 369 | stride=1 370 | pad=1 371 | activation=leaky 372 | 373 | [shortcut] 374 | from=-3 375 | activation=linear 376 | 377 | [convolutional] 378 | batch_normalize=1 379 | filters=256 380 | size=1 381 | stride=1 382 | pad=1 383 | activation=leaky 384 | 385 | [convolutional] 386 | batch_normalize=1 387 | filters=512 388 | size=3 389 | stride=1 390 | pad=1 391 | activation=leaky 392 | 393 | [shortcut] 394 | from=-3 395 | activation=linear 396 | 397 | 398 | [convolutional] 399 | batch_normalize=1 400 | filters=256 401 | size=1 402 | stride=1 403 | pad=1 404 | activation=leaky 405 | 406 | [convolutional] 407 | batch_normalize=1 408 | filters=512 409 | size=3 410 | stride=1 411 | pad=1 412 | activation=leaky 413 | 414 | [shortcut] 415 | from=-3 416 | activation=linear 417 | 418 | 419 | [convolutional] 420 | batch_normalize=1 421 | filters=256 422 | size=1 423 | stride=1 424 | pad=1 425 | activation=leaky 426 | 427 | [convolutional] 428 | batch_normalize=1 429 | filters=512 430 | size=3 431 | stride=1 432 | pad=1 433 | activation=leaky 434 | 435 | [shortcut] 436 | from=-3 437 | activation=linear 438 | 439 | [convolutional] 440 | batch_normalize=1 441 | filters=256 442 | size=1 443 | stride=1 444 | pad=1 445 | activation=leaky 446 | 447 | [convolutional] 448 | batch_normalize=1 449 | filters=512 450 | size=3 451 | stride=1 452 | pad=1 453 | activation=leaky 454 | 455 | [shortcut] 456 | from=-3 457 | activation=linear 458 | 459 | # Downsample 460 | 461 | [convolutional] 462 | batch_normalize=1 463 | filters=1024 464 | size=3 465 | stride=2 466 | pad=1 467 | activation=leaky 468 | 469 | [convolutional] 470 | batch_normalize=1 471 | filters=512 472 | size=1 473 | stride=1 474 | pad=1 475 | activation=leaky 476 | 477 | [convolutional] 478 | batch_normalize=1 479 | filters=1024 480 | size=3 481 | stride=1 482 | pad=1 483 | activation=leaky 484 | 485 | [shortcut] 486 | from=-3 487 | activation=linear 488 | 489 | [convolutional] 490 | batch_normalize=1 491 | filters=512 492 | size=1 493 | stride=1 494 | pad=1 495 | activation=leaky 496 | 497 | [convolutional] 498 | batch_normalize=1 499 | filters=1024 500 | size=3 501 | stride=1 502 | pad=1 503 | activation=leaky 504 | 505 | [shortcut] 506 | from=-3 507 | activation=linear 508 | 509 | [convolutional] 510 | batch_normalize=1 511 | filters=512 512 | size=1 513 | stride=1 514 | pad=1 515 | activation=leaky 516 | 517 | [convolutional] 518 | batch_normalize=1 519 | filters=1024 520 | size=3 521 | stride=1 522 | pad=1 523 | activation=leaky 524 | 525 | [shortcut] 526 | from=-3 527 | activation=linear 528 | 529 | [convolutional] 530 | batch_normalize=1 531 | filters=512 532 | size=1 533 | stride=1 534 | pad=1 535 | activation=leaky 536 | 537 | [convolutional] 538 | batch_normalize=1 539 | filters=1024 540 | size=3 541 | stride=1 542 | pad=1 543 | activation=leaky 544 | 545 | [shortcut] 546 | from=-3 547 | activation=linear 548 | 549 | ###################### 550 | 551 | [convolutional] 552 | batch_normalize=1 553 | filters=512 554 | size=1 555 | stride=1 556 | pad=1 557 | activation=leaky 558 | 559 | [convolutional] 560 | batch_normalize=1 561 | size=3 562 | stride=1 563 | pad=1 564 | filters=1024 565 | activation=leaky 566 | 567 | [convolutional] 568 | batch_normalize=1 569 | filters=512 570 | size=1 571 | stride=1 572 | pad=1 573 | activation=leaky 574 | 575 | [convolutional] 576 | batch_normalize=1 577 | size=3 578 | stride=1 579 | pad=1 580 | filters=1024 581 | activation=leaky 582 | 583 | [convolutional] 584 | batch_normalize=1 585 | filters=512 586 | size=1 587 | stride=1 588 | pad=1 589 | activation=leaky 590 | 591 | [convolutional] 592 | batch_normalize=1 593 | size=3 594 | stride=1 595 | pad=1 596 | filters=1024 597 | activation=leaky 598 | 599 | [convolutional] 600 | size=1 601 | stride=1 602 | pad=1 603 | filters=255 604 | activation=linear 605 | 606 | 607 | [yolo] 608 | mask = 6,7,8 609 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 610 | classes=80 611 | num=9 612 | jitter=.3 613 | ignore_thresh = .7 614 | truth_thresh = 1 615 | random=1 616 | 617 | 618 | [route] 619 | layers = -4 620 | 621 | [convolutional] 622 | batch_normalize=1 623 | filters=256 624 | size=1 625 | stride=1 626 | pad=1 627 | activation=leaky 628 | 629 | [upsample] 630 | stride=2 631 | 632 | [route] 633 | layers = -1, 61 634 | 635 | 636 | 637 | [convolutional] 638 | batch_normalize=1 639 | filters=256 640 | size=1 641 | stride=1 642 | pad=1 643 | activation=leaky 644 | 645 | [convolutional] 646 | batch_normalize=1 647 | size=3 648 | stride=1 649 | pad=1 650 | filters=512 651 | activation=leaky 652 | 653 | [convolutional] 654 | batch_normalize=1 655 | filters=256 656 | size=1 657 | stride=1 658 | pad=1 659 | activation=leaky 660 | 661 | [convolutional] 662 | batch_normalize=1 663 | size=3 664 | stride=1 665 | pad=1 666 | filters=512 667 | activation=leaky 668 | 669 | [convolutional] 670 | batch_normalize=1 671 | filters=256 672 | size=1 673 | stride=1 674 | pad=1 675 | activation=leaky 676 | 677 | [convolutional] 678 | batch_normalize=1 679 | size=3 680 | stride=1 681 | pad=1 682 | filters=512 683 | activation=leaky 684 | 685 | [convolutional] 686 | size=1 687 | stride=1 688 | pad=1 689 | filters=255 690 | activation=linear 691 | 692 | 693 | [yolo] 694 | mask = 3,4,5 695 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 696 | classes=80 697 | num=9 698 | jitter=.3 699 | ignore_thresh = .7 700 | truth_thresh = 1 701 | random=1 702 | 703 | 704 | 705 | [route] 706 | layers = -4 707 | 708 | [convolutional] 709 | batch_normalize=1 710 | filters=128 711 | size=1 712 | stride=1 713 | pad=1 714 | activation=leaky 715 | 716 | [upsample] 717 | stride=2 718 | 719 | [route] 720 | layers = -1, 36 721 | 722 | 723 | 724 | [convolutional] 725 | batch_normalize=1 726 | filters=128 727 | size=1 728 | stride=1 729 | pad=1 730 | activation=leaky 731 | 732 | [convolutional] 733 | batch_normalize=1 734 | size=3 735 | stride=1 736 | pad=1 737 | filters=256 738 | activation=leaky 739 | 740 | [convolutional] 741 | batch_normalize=1 742 | filters=128 743 | size=1 744 | stride=1 745 | pad=1 746 | activation=leaky 747 | 748 | [convolutional] 749 | batch_normalize=1 750 | size=3 751 | stride=1 752 | pad=1 753 | filters=256 754 | activation=leaky 755 | 756 | [convolutional] 757 | batch_normalize=1 758 | filters=128 759 | size=1 760 | stride=1 761 | pad=1 762 | activation=leaky 763 | 764 | [convolutional] 765 | batch_normalize=1 766 | size=3 767 | stride=1 768 | pad=1 769 | filters=256 770 | activation=leaky 771 | 772 | [convolutional] 773 | size=1 774 | stride=1 775 | pad=1 776 | filters=255 777 | activation=linear 778 | 779 | 780 | [yolo] 781 | mask = 0,1,2 782 | anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 783 | classes=80 784 | num=9 785 | jitter=.3 786 | ignore_thresh = .7 787 | truth_thresh = 1 788 | random=1 789 | -------------------------------------------------------------------------------- /data/coco.data: -------------------------------------------------------------------------------- 1 | classes=80 2 | train=../coco/trainvalno5k.txt 3 | valid=../coco/5k.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco.names: -------------------------------------------------------------------------------- 1 | person 2 | bicycle 3 | car 4 | motorcycle 5 | airplane 6 | bus 7 | train 8 | truck 9 | boat 10 | traffic light 11 | fire hydrant 12 | stop sign 13 | parking meter 14 | bench 15 | bird 16 | cat 17 | dog 18 | horse 19 | sheep 20 | cow 21 | elephant 22 | bear 23 | zebra 24 | giraffe 25 | backpack 26 | umbrella 27 | handbag 28 | tie 29 | suitcase 30 | frisbee 31 | skis 32 | snowboard 33 | sports ball 34 | kite 35 | baseball bat 36 | baseball glove 37 | skateboard 38 | surfboard 39 | tennis racket 40 | bottle 41 | wine glass 42 | cup 43 | fork 44 | knife 45 | spoon 46 | bowl 47 | banana 48 | apple 49 | sandwich 50 | orange 51 | broccoli 52 | carrot 53 | hot dog 54 | pizza 55 | donut 56 | cake 57 | chair 58 | couch 59 | potted plant 60 | bed 61 | dining table 62 | toilet 63 | tv 64 | laptop 65 | mouse 66 | remote 67 | keyboard 68 | cell phone 69 | microwave 70 | oven 71 | toaster 72 | sink 73 | refrigerator 74 | book 75 | clock 76 | vase 77 | scissors 78 | teddy bear 79 | hair drier 80 | toothbrush 81 | -------------------------------------------------------------------------------- /data/coco_1000img.data: -------------------------------------------------------------------------------- 1 | classes=80 2 | train=./data/coco_1000img.txt 3 | valid=./data/coco_1000img.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco_1000val.data: -------------------------------------------------------------------------------- 1 | classes=80 2 | train=./data/coco_1000img.txt 3 | valid=./data/coco_1000val.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco_16img.data: -------------------------------------------------------------------------------- 1 | classes=80 2 | train=./data/coco_16img.txt 3 | valid=./data/coco_16img.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco_16img.txt: -------------------------------------------------------------------------------- 1 | ../coco/images/train2014/COCO_train2014_000000000009.jpg 2 | ../coco/images/train2014/COCO_train2014_000000000025.jpg 3 | ../coco/images/train2014/COCO_train2014_000000000030.jpg 4 | ../coco/images/train2014/COCO_train2014_000000000034.jpg 5 | ../coco/images/train2014/COCO_train2014_000000000036.jpg 6 | ../coco/images/train2014/COCO_train2014_000000000049.jpg 7 | ../coco/images/train2014/COCO_train2014_000000000061.jpg 8 | ../coco/images/train2014/COCO_train2014_000000000064.jpg 9 | ../coco/images/train2014/COCO_train2014_000000000071.jpg 10 | ../coco/images/train2014/COCO_train2014_000000000072.jpg 11 | ../coco/images/train2014/COCO_train2014_000000000077.jpg 12 | ../coco/images/train2014/COCO_train2014_000000000078.jpg 13 | ../coco/images/train2014/COCO_train2014_000000000081.jpg 14 | ../coco/images/train2014/COCO_train2014_000000000086.jpg 15 | ../coco/images/train2014/COCO_train2014_000000000089.jpg 16 | ../coco/images/train2014/COCO_train2014_000000000092.jpg 17 | -------------------------------------------------------------------------------- /data/coco_1cls.data: -------------------------------------------------------------------------------- 1 | classes=1 2 | train=./data/coco_1cls.txt 3 | valid=./data/coco_1cls.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco_1cls.txt: -------------------------------------------------------------------------------- 1 | ../coco/images/val2014/COCO_val2014_000000013992.jpg 2 | ../coco/images/val2014/COCO_val2014_000000047226.jpg 3 | ../coco/images/val2014/COCO_val2014_000000050324.jpg 4 | ../coco/images/val2014/COCO_val2014_000000121497.jpg 5 | ../coco/images/val2014/COCO_val2014_000000001464.jpg 6 | -------------------------------------------------------------------------------- /data/coco_1img.data: -------------------------------------------------------------------------------- 1 | classes=80 2 | train=./data/coco_1img.txt 3 | valid=./data/coco_1img.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco_1img.txt: -------------------------------------------------------------------------------- 1 | ../coco/images/val2014/COCO_val2014_000000581886.jpg 2 | -------------------------------------------------------------------------------- /data/coco_1k5k.data: -------------------------------------------------------------------------------- 1 | classes=80 2 | train=./data/coco_1000img.txt 3 | valid=./data/5k.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco_32img.data: -------------------------------------------------------------------------------- 1 | classes=80 2 | train=./data/coco_32img.txt 3 | valid=./data/coco_32img.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco_32img.txt: -------------------------------------------------------------------------------- 1 | ../coco/images/train2014/COCO_train2014_000000000009.jpg 2 | ../coco/images/train2014/COCO_train2014_000000000025.jpg 3 | ../coco/images/train2014/COCO_train2014_000000000030.jpg 4 | ../coco/images/train2014/COCO_train2014_000000000034.jpg 5 | ../coco/images/train2014/COCO_train2014_000000000036.jpg 6 | ../coco/images/train2014/COCO_train2014_000000000049.jpg 7 | ../coco/images/train2014/COCO_train2014_000000000061.jpg 8 | ../coco/images/train2014/COCO_train2014_000000000064.jpg 9 | ../coco/images/train2014/COCO_train2014_000000000071.jpg 10 | ../coco/images/train2014/COCO_train2014_000000000072.jpg 11 | ../coco/images/train2014/COCO_train2014_000000000077.jpg 12 | ../coco/images/train2014/COCO_train2014_000000000078.jpg 13 | ../coco/images/train2014/COCO_train2014_000000000081.jpg 14 | ../coco/images/train2014/COCO_train2014_000000000086.jpg 15 | ../coco/images/train2014/COCO_train2014_000000000089.jpg 16 | ../coco/images/train2014/COCO_train2014_000000000092.jpg 17 | ../coco/images/train2014/COCO_train2014_000000000094.jpg 18 | ../coco/images/train2014/COCO_train2014_000000000109.jpg 19 | ../coco/images/train2014/COCO_train2014_000000000110.jpg 20 | ../coco/images/train2014/COCO_train2014_000000000113.jpg 21 | ../coco/images/train2014/COCO_train2014_000000000127.jpg 22 | ../coco/images/train2014/COCO_train2014_000000000138.jpg 23 | ../coco/images/train2014/COCO_train2014_000000000142.jpg 24 | ../coco/images/train2014/COCO_train2014_000000000144.jpg 25 | ../coco/images/train2014/COCO_train2014_000000000149.jpg 26 | ../coco/images/train2014/COCO_train2014_000000000151.jpg 27 | ../coco/images/train2014/COCO_train2014_000000000154.jpg 28 | ../coco/images/train2014/COCO_train2014_000000000165.jpg 29 | ../coco/images/train2014/COCO_train2014_000000000194.jpg 30 | ../coco/images/train2014/COCO_train2014_000000000201.jpg 31 | ../coco/images/train2014/COCO_train2014_000000000247.jpg 32 | ../coco/images/train2014/COCO_train2014_000000000260.jpg 33 | -------------------------------------------------------------------------------- /data/coco_500val.data: -------------------------------------------------------------------------------- 1 | classes=80 2 | train=./data/coco_500img.txt 3 | valid=./data/coco_500val.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco_64img.data: -------------------------------------------------------------------------------- 1 | classes=80 2 | train=./data/coco_64img.txt 3 | valid=./data/coco_64img.txt 4 | names=data/coco.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/coco_64img.shapes: -------------------------------------------------------------------------------- 1 | 640 480 2 | 640 426 3 | 640 428 4 | 640 425 5 | 481 640 6 | 381 500 7 | 640 488 8 | 480 640 9 | 640 426 10 | 427 640 11 | 500 375 12 | 612 612 13 | 640 425 14 | 512 640 15 | 640 480 16 | 640 427 17 | 640 427 18 | 640 416 19 | 640 480 20 | 416 640 21 | 640 481 22 | 640 573 23 | 480 640 24 | 640 480 25 | 640 428 26 | 480 640 27 | 427 640 28 | 640 536 29 | 640 480 30 | 640 428 31 | 640 424 32 | 500 333 33 | 591 640 34 | 640 480 35 | 640 426 36 | 600 600 37 | 640 427 38 | 640 427 39 | 640 480 40 | 640 481 41 | 640 427 42 | 640 480 43 | 640 480 44 | 480 640 45 | 480 640 46 | 640 480 47 | 446 640 48 | 640 480 49 | 640 611 50 | 426 640 51 | 640 480 52 | 640 389 53 | 427 640 54 | 640 480 55 | 640 480 56 | 480 640 57 | 640 480 58 | 640 427 59 | 500 495 60 | 500 313 61 | 640 480 62 | 360 640 63 | 427 640 64 | 640 480 65 | -------------------------------------------------------------------------------- /data/coco_64img.txt: -------------------------------------------------------------------------------- 1 | ../coco/images/train2014/COCO_train2014_000000000009.jpg 2 | ../coco/images/train2014/COCO_train2014_000000000025.jpg 3 | ../coco/images/train2014/COCO_train2014_000000000030.jpg 4 | ../coco/images/train2014/COCO_train2014_000000000034.jpg 5 | ../coco/images/train2014/COCO_train2014_000000000036.jpg 6 | ../coco/images/train2014/COCO_train2014_000000000049.jpg 7 | ../coco/images/train2014/COCO_train2014_000000000061.jpg 8 | ../coco/images/train2014/COCO_train2014_000000000064.jpg 9 | ../coco/images/train2014/COCO_train2014_000000000071.jpg 10 | ../coco/images/train2014/COCO_train2014_000000000072.jpg 11 | ../coco/images/train2014/COCO_train2014_000000000077.jpg 12 | ../coco/images/train2014/COCO_train2014_000000000078.jpg 13 | ../coco/images/train2014/COCO_train2014_000000000081.jpg 14 | ../coco/images/train2014/COCO_train2014_000000000086.jpg 15 | ../coco/images/train2014/COCO_train2014_000000000089.jpg 16 | ../coco/images/train2014/COCO_train2014_000000000092.jpg 17 | ../coco/images/train2014/COCO_train2014_000000000094.jpg 18 | ../coco/images/train2014/COCO_train2014_000000000109.jpg 19 | ../coco/images/train2014/COCO_train2014_000000000110.jpg 20 | ../coco/images/train2014/COCO_train2014_000000000113.jpg 21 | ../coco/images/train2014/COCO_train2014_000000000127.jpg 22 | ../coco/images/train2014/COCO_train2014_000000000138.jpg 23 | ../coco/images/train2014/COCO_train2014_000000000142.jpg 24 | ../coco/images/train2014/COCO_train2014_000000000144.jpg 25 | ../coco/images/train2014/COCO_train2014_000000000149.jpg 26 | ../coco/images/train2014/COCO_train2014_000000000151.jpg 27 | ../coco/images/train2014/COCO_train2014_000000000154.jpg 28 | ../coco/images/train2014/COCO_train2014_000000000165.jpg 29 | ../coco/images/train2014/COCO_train2014_000000000194.jpg 30 | ../coco/images/train2014/COCO_train2014_000000000201.jpg 31 | ../coco/images/train2014/COCO_train2014_000000000247.jpg 32 | ../coco/images/train2014/COCO_train2014_000000000260.jpg 33 | ../coco/images/train2014/COCO_train2014_000000000263.jpg 34 | ../coco/images/train2014/COCO_train2014_000000000307.jpg 35 | ../coco/images/train2014/COCO_train2014_000000000308.jpg 36 | ../coco/images/train2014/COCO_train2014_000000000309.jpg 37 | ../coco/images/train2014/COCO_train2014_000000000312.jpg 38 | ../coco/images/train2014/COCO_train2014_000000000315.jpg 39 | ../coco/images/train2014/COCO_train2014_000000000321.jpg 40 | ../coco/images/train2014/COCO_train2014_000000000322.jpg 41 | ../coco/images/train2014/COCO_train2014_000000000326.jpg 42 | ../coco/images/train2014/COCO_train2014_000000000332.jpg 43 | ../coco/images/train2014/COCO_train2014_000000000349.jpg 44 | ../coco/images/train2014/COCO_train2014_000000000368.jpg 45 | ../coco/images/train2014/COCO_train2014_000000000370.jpg 46 | ../coco/images/train2014/COCO_train2014_000000000382.jpg 47 | ../coco/images/train2014/COCO_train2014_000000000384.jpg 48 | ../coco/images/train2014/COCO_train2014_000000000389.jpg 49 | ../coco/images/train2014/COCO_train2014_000000000394.jpg 50 | ../coco/images/train2014/COCO_train2014_000000000404.jpg 51 | ../coco/images/train2014/COCO_train2014_000000000419.jpg 52 | ../coco/images/train2014/COCO_train2014_000000000431.jpg 53 | ../coco/images/train2014/COCO_train2014_000000000436.jpg 54 | ../coco/images/train2014/COCO_train2014_000000000438.jpg 55 | ../coco/images/train2014/COCO_train2014_000000000443.jpg 56 | ../coco/images/train2014/COCO_train2014_000000000446.jpg 57 | ../coco/images/train2014/COCO_train2014_000000000450.jpg 58 | ../coco/images/train2014/COCO_train2014_000000000471.jpg 59 | ../coco/images/train2014/COCO_train2014_000000000490.jpg 60 | ../coco/images/train2014/COCO_train2014_000000000491.jpg 61 | ../coco/images/train2014/COCO_train2014_000000000510.jpg 62 | ../coco/images/train2014/COCO_train2014_000000000514.jpg 63 | ../coco/images/train2014/COCO_train2014_000000000529.jpg 64 | ../coco/images/train2014/COCO_train2014_000000000531.jpg 65 | -------------------------------------------------------------------------------- /data/coco_paper.names: -------------------------------------------------------------------------------- 1 | person 2 | bicycle 3 | car 4 | motorcycle 5 | airplane 6 | bus 7 | train 8 | truck 9 | boat 10 | traffic light 11 | fire hydrant 12 | street sign 13 | stop sign 14 | parking meter 15 | bench 16 | bird 17 | cat 18 | dog 19 | horse 20 | sheep 21 | cow 22 | elephant 23 | bear 24 | zebra 25 | giraffe 26 | hat 27 | backpack 28 | umbrella 29 | shoe 30 | eye glasses 31 | handbag 32 | tie 33 | suitcase 34 | frisbee 35 | skis 36 | snowboard 37 | sports ball 38 | kite 39 | baseball bat 40 | baseball glove 41 | skateboard 42 | surfboard 43 | tennis racket 44 | bottle 45 | plate 46 | wine glass 47 | cup 48 | fork 49 | knife 50 | spoon 51 | bowl 52 | banana 53 | apple 54 | sandwich 55 | orange 56 | broccoli 57 | carrot 58 | hot dog 59 | pizza 60 | donut 61 | cake 62 | chair 63 | couch 64 | potted plant 65 | bed 66 | mirror 67 | dining table 68 | window 69 | desk 70 | toilet 71 | door 72 | tv 73 | laptop 74 | mouse 75 | remote 76 | keyboard 77 | cell phone 78 | microwave 79 | oven 80 | toaster 81 | sink 82 | refrigerator 83 | blender 84 | book 85 | clock 86 | vase 87 | scissors 88 | teddy bear 89 | hair drier 90 | toothbrush 91 | hair brush -------------------------------------------------------------------------------- /data/get_coco_dataset.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # CREDIT: https://github.com/pjreddie/darknet/tree/master/scripts/get_coco_dataset.sh 3 | 4 | # Clone COCO API 5 | git clone https://github.com/pdollar/coco && cd coco 6 | 7 | # Download Images 8 | mkdir images && cd images 9 | wget -c https://pjreddie.com/media/files/train2014.zip 10 | wget -c https://pjreddie.com/media/files/val2014.zip 11 | 12 | # Unzip 13 | unzip -q train2014.zip 14 | unzip -q val2014.zip 15 | 16 | # (optional) Delete zip files 17 | rm -rf *.zip 18 | 19 | cd .. 20 | 21 | # Download COCO Metadata 22 | wget -c https://pjreddie.com/media/files/instances_train-val2014.zip 23 | wget -c https://pjreddie.com/media/files/coco/5k.part 24 | wget -c https://pjreddie.com/media/files/coco/trainvalno5k.part 25 | wget -c https://pjreddie.com/media/files/coco/labels.tgz 26 | tar xzf labels.tgz 27 | unzip -q instances_train-val2014.zip 28 | 29 | # Set Up Image Lists 30 | paste <(awk "{print \"$PWD\"}" <5k.part) 5k.part | tr -d '\t' > 5k.txt 31 | paste <(awk "{print \"$PWD\"}" trainvalno5k.txt 32 | 33 | # get xview training data 34 | # wget -O train_images.tgz 'https://d307kc0mrhucc3.cloudfront.net/train_images.tgz?Expires=1530124049&Signature=JrQoxipmsETvb7eQHCfDFUO-QEHJGAayUv0i-ParmS-1hn7hl9D~bzGuHWG82imEbZSLUARTtm0wOJ7EmYMGmG5PtLKz9H5qi6DjoSUuFc13NQ-~6yUhE~NfPaTnehUdUMCa3On2wl1h1ZtRG~0Jq1P-AJbpe~oQxbyBrs1KccaMa7FK4F4oMM6sMnNgoXx8-3O77kYw~uOpTMFmTaQdHln6EztW0Lx17i57kK3ogbSUpXgaUTqjHCRA1dWIl7PY1ngQnLslkLhZqmKcaL-BvWf0ZGjHxCDQBpnUjIlvMu5NasegkwD9Jjc0ClgTxsttSkmbapVqaVC8peR0pO619Q__&Key-Pair-Id=APKAIKGDJB5C3XUL2DXQ' 35 | # tar -xvzf train_images.tgz 36 | # sudo rm -rf train_images/._* 37 | # lastly convert each .tif to a .bmp for faster loading in cv2 38 | 39 | # ./coco/images/train2014/COCO_train2014_000000167126.jpg # corrupted image 40 | -------------------------------------------------------------------------------- /data/get_coco_dataset_gdrive.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # https://stackoverflow.com/questions/48133080/how-to-download-a-google-drive-url-via-curl-or-wget/48133859 3 | 4 | # Zip coco folder 5 | # zip -r coco.zip coco 6 | # tar -czvf coco.tar.gz coco 7 | 8 | # Set fileid and filename 9 | filename="coco.zip" 10 | fileid="1HaXkef9z6y5l4vUnCYgdmEAj61c6bfWO" # coco.zip 11 | 12 | # Download from Google Drive, accepting presented query 13 | curl -c ./cookie -s -L "https://drive.google.com/uc?export=download&id=${fileid}" > /dev/null 14 | curl -Lb ./cookie "https://drive.google.com/uc?export=download&confirm=`awk '/download/ {print $NF}' ./cookie`&id=${fileid}" -o ${filename} 15 | rm ./cookie 16 | 17 | # Unzip 18 | unzip -q ${filename} # for coco.zip 19 | # tar -xzf ${filename} # for coco.tar.gz 20 | -------------------------------------------------------------------------------- /data/hand.data: -------------------------------------------------------------------------------- 1 | classes=1 2 | train=D:/dl/YOLOv3-model-pruning/data/train_.txt 3 | valid=D:/dl/YOLOv3-model-pruning/data/valid_.txt 4 | names=D:/dl/YOLOv3-model-pruning/data/oxfordhand.names 5 | backup=backup/ 6 | eval=coco 7 | -------------------------------------------------------------------------------- /data/samples/bus.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zbyuan/pruning_yolov3/25e13596a218ec55375f4781aa3dae621b5961cb/data/samples/bus.jpg -------------------------------------------------------------------------------- /data/samples/zidane.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zbyuan/pruning_yolov3/25e13596a218ec55375f4781aa3dae621b5961cb/data/samples/zidane.jpg -------------------------------------------------------------------------------- /data/valid_.shapes: -------------------------------------------------------------------------------- 1 | 500 375 2 | 500 375 3 | 500 375 4 | 500 375 5 | 500 375 6 | 500 375 7 | 375 500 8 | 500 333 9 | 333 500 10 | 500 375 11 | 500 434 12 | 500 375 13 | 333 500 14 | 500 375 15 | 500 331 16 | 500 375 17 | 500 375 18 | 500 374 19 | 500 375 20 | 500 375 21 | 375 500 22 | 500 333 23 | 500 333 24 | 500 375 25 | 500 332 26 | 500 486 27 | 500 375 28 | 500 375 29 | 375 500 30 | 375 500 31 | 500 375 32 | 500 375 33 | 500 375 34 | 355 500 35 | 375 500 36 | 500 333 37 | 500 375 38 | 500 377 39 | 375 500 40 | 500 375 41 | 500 375 42 | 500 375 43 | 500 375 44 | 333 500 45 | 500 375 46 | 500 333 47 | 500 346 48 | 500 375 49 | 476 500 50 | 500 333 51 | 500 420 52 | 500 333 53 | 500 333 54 | 500 333 55 | 333 500 56 | 333 500 57 | 375 500 58 | 500 379 59 | 500 375 60 | 500 375 61 | 500 357 62 | 375 500 63 | 500 393 64 | 333 500 65 | 500 375 66 | 500 375 67 | 500 333 68 | 333 500 69 | 327 500 70 | 500 375 71 | 500 375 72 | 500 345 73 | 333 500 74 | 375 500 75 | 500 380 76 | 500 375 77 | 487 377 78 | 500 375 79 | 500 333 80 | 500 333 81 | 333 500 82 | 500 375 83 | 375 500 84 | 500 375 85 | 500 375 86 | 500 375 87 | 375 500 88 | 500 375 89 | 500 332 90 | 333 500 91 | 480 360 92 | 500 334 93 | 500 375 94 | 500 375 95 | 333 500 96 | 500 333 97 | 375 500 98 | 500 375 99 | 500 375 100 | 500 375 101 | 500 375 102 | 500 375 103 | 332 500 104 | 500 375 105 | 500 375 106 | 375 500 107 | 500 333 108 | 500 331 109 | 500 375 110 | 333 500 111 | 333 500 112 | 486 500 113 | 500 375 114 | 375 500 115 | 356 500 116 | 500 375 117 | 375 500 118 | 500 375 119 | 500 375 120 | 268 400 121 | 389 500 122 | 333 500 123 | 500 375 124 | 500 375 125 | 500 367 126 | 500 375 127 | 500 375 128 | 500 334 129 | 495 500 130 | 319 480 131 | 500 375 132 | 500 375 133 | 333 500 134 | 500 375 135 | 500 375 136 | 500 375 137 | 326 500 138 | 500 375 139 | 500 375 140 | 500 375 141 | 500 400 142 | 332 500 143 | 500 375 144 | 500 375 145 | 360 331 146 | 333 500 147 | 500 332 148 | 500 374 149 | 500 375 150 | 375 500 151 | 500 375 152 | 500 375 153 | 500 375 154 | 500 367 155 | 500 375 156 | 500 375 157 | 500 375 158 | 375 500 159 | 500 375 160 | 500 375 161 | 500 390 162 | 500 358 163 | 500 397 164 | 500 341 165 | 375 500 166 | 500 333 167 | 500 375 168 | 332 500 169 | 500 375 170 | 500 375 171 | 500 375 172 | 375 500 173 | 240 320 174 | 450 480 175 | 417 500 176 | 500 400 177 | 500 375 178 | 500 411 179 | 338 500 180 | 500 375 181 | 500 375 182 | 379 500 183 | 500 375 184 | 333 500 185 | 500 332 186 | 500 375 187 | 500 375 188 | 500 375 189 | 500 375 190 | 332 500 191 | 469 500 192 | 500 375 193 | 333 500 194 | 500 375 195 | 500 375 196 | 500 376 197 | 500 375 198 | 500 334 199 | 500 375 200 | 500 375 201 | 500 341 202 | 500 333 203 | 500 375 204 | 500 375 205 | 500 334 206 | 500 375 207 | 500 375 208 | 500 357 209 | 375 500 210 | 500 375 211 | 500 375 212 | 375 500 213 | 500 375 214 | 500 497 215 | 375 500 216 | 375 500 217 | 500 334 218 | 500 375 219 | 500 375 220 | 500 375 221 | 500 375 222 | 500 375 223 | 333 500 224 | 500 375 225 | 375 500 226 | 500 375 227 | 500 375 228 | 375 500 229 | 500 375 230 | 334 500 231 | 500 375 232 | 364 500 233 | 375 500 234 | 494 500 235 | 484 500 236 | 500 333 237 | 500 375 238 | 500 443 239 | 375 500 240 | 500 375 241 | 500 334 242 | 500 375 243 | 375 500 244 | 500 375 245 | 500 333 246 | 500 375 247 | 313 500 248 | 500 375 249 | 400 300 250 | 375 500 251 | 375 500 252 | 500 375 253 | 333 500 254 | 500 337 255 | 375 500 256 | 500 290 257 | 500 375 258 | 500 312 259 | 500 333 260 | 500 375 261 | 375 500 262 | 500 333 263 | 500 333 264 | 500 333 265 | 500 375 266 | 500 375 267 | 500 375 268 | 500 333 269 | 500 375 270 | 500 375 271 | 500 375 272 | 500 375 273 | 500 375 274 | 375 500 275 | 375 500 276 | 500 375 277 | 500 374 278 | 333 500 279 | 375 500 280 | 500 375 281 | 500 375 282 | 500 375 283 | 500 375 284 | 333 500 285 | 500 375 286 | 500 375 287 | 500 375 288 | 500 333 289 | 294 500 290 | 500 375 291 | 500 375 292 | 500 375 293 | 500 334 294 | 375 500 295 | 333 500 296 | 500 375 297 | 333 500 298 | 500 375 299 | 500 221 300 | 500 374 301 | 500 375 302 | 333 500 303 | 500 333 304 | 500 375 305 | 270 360 306 | 500 371 307 | 500 333 308 | 500 335 309 | 358 500 310 | 220 500 311 | 500 375 312 | 500 375 313 | 375 500 314 | 500 375 315 | 500 375 316 | 375 500 317 | 366 500 318 | 500 375 319 | 500 379 320 | 500 375 321 | 500 489 322 | 500 333 323 | 500 375 324 | 500 375 325 | 500 333 326 | 500 375 327 | 500 375 328 | 334 500 329 | 500 395 330 | 333 500 331 | 500 369 332 | 500 375 333 | 375 500 334 | 500 375 335 | 500 375 336 | 375 500 337 | 500 333 338 | 500 332 339 | 500 375 340 | 500 375 341 | 375 500 342 | 375 500 343 | 500 379 344 | 500 395 345 | 500 333 346 | 500 375 347 | 500 375 348 | 378 500 349 | 500 333 350 | 500 335 351 | 500 333 352 | 375 500 353 | 375 500 354 | 281 500 355 | 500 336 356 | 500 333 357 | 500 375 358 | 500 245 359 | 500 375 360 | 500 375 361 | 500 333 362 | 500 375 363 | 500 334 364 | 500 375 365 | 500 419 366 | 500 375 367 | 500 333 368 | 500 375 369 | 500 375 370 | 375 500 371 | 500 375 372 | 500 375 373 | 500 375 374 | 375 500 375 | 500 332 376 | 500 333 377 | 500 277 378 | 500 333 379 | 500 333 380 | 375 500 381 | 500 334 382 | 500 375 383 | 500 375 384 | 500 333 385 | 335 500 386 | 500 375 387 | 500 375 388 | 332 500 389 | 500 375 390 | 500 375 391 | 500 375 392 | 500 375 393 | 500 375 394 | 500 313 395 | 500 375 396 | 500 375 397 | 333 500 398 | 500 375 399 | 500 375 400 | 335 500 401 | 500 375 402 | 500 375 403 | 500 375 404 | 375 500 405 | 500 335 406 | 375 500 407 | 500 375 408 | 375 500 409 | 500 500 410 | 500 375 411 | 500 375 412 | 500 333 413 | 500 375 414 | 500 375 415 | 500 375 416 | 500 375 417 | 333 500 418 | 500 375 419 | 500 375 420 | 500 375 421 | 500 332 422 | 500 375 423 | 334 500 424 | 332 500 425 | 375 500 426 | 500 333 427 | 500 405 428 | 333 500 429 | 500 334 430 | 500 333 431 | 500 375 432 | 500 375 433 | 500 332 434 | 333 500 435 | 368 500 436 | 375 500 437 | 500 375 438 | 500 375 439 | 500 375 440 | 500 375 441 | 500 375 442 | 288 432 443 | 375 500 444 | 500 375 445 | 500 333 446 | 500 375 447 | 500 333 448 | 375 500 449 | 500 375 450 | 500 281 451 | 333 500 452 | 500 333 453 | 500 375 454 | 500 333 455 | 500 375 456 | 500 334 457 | 500 375 458 | 375 500 459 | 375 500 460 | 375 500 461 | 500 333 462 | 500 333 463 | 500 375 464 | 500 375 465 | 500 375 466 | 500 375 467 | 500 375 468 | 500 356 469 | 474 500 470 | 500 375 471 | 500 375 472 | 500 326 473 | 360 480 474 | 500 375 475 | 500 375 476 | 500 488 477 | 500 375 478 | 442 500 479 | 500 333 480 | 450 349 481 | 375 500 482 | 500 375 483 | 375 500 484 | 306 500 485 | 500 338 486 | 500 333 487 | 500 375 488 | 375 500 489 | 500 375 490 | 500 333 491 | 375 500 492 | 500 375 493 | 375 500 494 | 378 500 495 | 500 375 496 | 500 375 497 | 500 375 498 | 500 405 499 | 500 333 500 | 500 375 501 | 500 500 502 | 500 375 503 | 450 300 504 | 500 375 505 | 500 375 506 | 500 375 507 | 333 500 508 | 500 375 509 | 500 375 510 | 500 375 511 | 500 375 512 | 500 377 513 | 500 375 514 | 500 333 515 | 500 375 516 | 375 500 517 | 375 500 518 | 500 375 519 | 500 375 520 | 500 375 521 | 375 500 522 | 500 375 523 | 500 332 524 | 500 375 525 | 500 375 526 | 500 375 527 | 334 500 528 | 500 375 529 | 500 375 530 | 332 500 531 | 500 333 532 | 500 375 533 | 375 500 534 | 375 500 535 | 333 500 536 | 500 332 537 | 500 375 538 | 500 375 539 | 500 375 540 | 500 364 541 | 333 500 542 | 500 375 543 | 500 333 544 | 500 375 545 | 500 375 546 | 500 333 547 | 375 500 548 | 500 375 549 | 500 333 550 | 500 336 551 | 500 375 552 | 500 359 553 | 500 333 554 | 500 375 555 | 500 375 556 | 500 375 557 | 500 332 558 | 500 375 559 | 500 333 560 | 500 375 561 | 282 500 562 | 500 375 563 | 500 375 564 | 500 375 565 | 500 375 566 | 500 375 567 | 375 500 568 | 500 375 569 | 375 500 570 | 500 375 571 | 500 375 572 | 500 375 573 | 500 390 574 | 334 500 575 | 332 500 576 | 500 375 577 | 500 375 578 | 425 319 579 | 500 333 580 | 500 334 581 | 500 375 582 | 375 500 583 | 500 333 584 | 500 375 585 | 375 500 586 | 500 332 587 | 500 375 588 | 448 336 589 | 500 375 590 | 500 375 591 | 500 375 592 | 500 375 593 | 500 375 594 | 500 375 595 | 500 335 596 | 500 333 597 | 375 500 598 | 375 500 599 | 500 333 600 | 500 375 601 | 333 500 602 | 500 375 603 | 500 375 604 | 500 375 605 | 375 500 606 | 500 375 607 | 500 375 608 | 500 375 609 | 333 500 610 | 375 500 611 | 500 439 612 | 375 500 613 | 500 375 614 | 500 375 615 | 334 500 616 | 374 500 617 | 500 375 618 | 321 500 619 | 500 400 620 | 500 375 621 | 500 375 622 | 500 375 623 | 500 375 624 | 500 410 625 | 500 333 626 | 500 375 627 | 500 375 628 | 500 334 629 | 500 375 630 | 500 375 631 | 500 326 632 | 500 375 633 | 500 357 634 | 500 374 635 | 500 375 636 | 500 374 637 | 500 333 638 | 500 375 639 | 500 375 640 | 500 334 641 | 375 500 642 | 500 375 643 | 500 334 644 | 500 375 645 | 500 333 646 | 500 375 647 | 500 375 648 | 500 375 649 | 480 361 650 | 375 500 651 | 333 500 652 | 500 333 653 | 333 500 654 | 500 333 655 | 500 372 656 | 375 500 657 | 500 375 658 | 375 500 659 | 500 319 660 | 500 333 661 | 500 375 662 | 375 500 663 | 500 377 664 | 400 498 665 | 500 393 666 | 500 334 667 | 333 500 668 | 500 333 669 | 500 375 670 | 500 375 671 | 500 375 672 | 500 375 673 | 375 500 674 | 375 500 675 | 333 500 676 | 500 333 677 | 500 333 678 | 333 500 679 | 306 500 680 | 500 375 681 | 500 334 682 | 320 448 683 | 333 500 684 | 375 500 685 | 500 334 686 | 270 360 687 | 361 500 688 | 500 393 689 | 500 375 690 | 396 500 691 | 500 333 692 | 500 333 693 | 320 480 694 | 500 375 695 | 500 375 696 | 500 375 697 | 500 375 698 | 399 500 699 | 500 375 700 | 500 375 701 | 500 333 702 | 500 332 703 | 500 375 704 | 320 480 705 | 500 375 706 | 375 500 707 | 500 334 708 | 451 500 709 | 500 375 710 | 500 375 711 | 500 368 712 | 500 375 713 | 375 500 714 | 500 375 715 | 500 375 716 | 319 500 717 | 500 375 718 | 450 500 719 | 375 500 720 | 500 375 721 | 375 500 722 | 500 374 723 | 500 375 724 | 500 375 725 | 500 375 726 | 345 500 727 | 500 375 728 | 500 375 729 | 500 325 730 | 500 375 731 | 500 379 732 | 500 333 733 | 500 375 734 | 375 500 735 | 500 375 736 | 500 375 737 | 354 500 738 | 500 375 739 | 375 500 740 | 500 375 741 | 333 500 742 | 375 500 743 | 500 333 744 | 225 417 745 | 333 500 746 | 500 333 747 | 375 500 748 | 500 332 749 | 500 334 750 | 400 500 751 | 500 333 752 | 333 500 753 | 500 375 754 | 375 500 755 | 333 500 756 | 500 333 757 | 340 500 758 | 500 375 759 | 375 500 760 | 333 500 761 | 500 375 762 | 500 375 763 | 500 375 764 | 500 375 765 | 500 402 766 | 375 500 767 | 500 333 768 | 500 333 769 | 374 500 770 | 500 333 771 | 375 500 772 | 500 333 773 | 500 375 774 | 500 375 775 | 500 294 776 | 500 375 777 | 375 500 778 | 500 375 779 | 500 332 780 | 332 500 781 | 358 500 782 | 500 333 783 | 380 472 784 | 500 375 785 | 375 500 786 | 500 375 787 | 500 375 788 | 500 375 789 | 453 500 790 | 375 500 791 | 500 333 792 | 500 500 793 | 375 500 794 | 500 375 795 | 375 500 796 | 500 375 797 | 375 500 798 | 402 500 799 | 459 288 800 | 273 500 801 | 415 500 802 | 346 336 803 | 320 500 804 | 500 375 805 | 500 333 806 | 500 333 807 | 500 410 808 | 500 375 809 | 500 458 810 | 500 333 811 | 500 375 812 | 443 437 813 | 341 251 814 | 375 500 815 | 500 375 816 | 252 400 817 | 288 432 818 | 500 334 819 | 375 500 820 | 375 500 821 | 377 500 822 | -------------------------------------------------------------------------------- /detect.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from sys import platform 3 | 4 | from models import * # set ONNX_EXPORT in models.py 5 | from utils.datasets import * 6 | from utils.utils import * 7 | 8 | 9 | def detect(save_txt=False, save_img=False): 10 | img_size = (320, 192) if ONNX_EXPORT else opt.img_size # (320, 192) or (416, 256) or (608, 352) for (height, width) 11 | out, source, weights, half, view_img = opt.output, opt.source, opt.weights, opt.half, opt.view_img 12 | webcam = source == '0' or source.startswith('rtsp') or source.startswith('http') or source.endswith('.txt') 13 | 14 | # Initialize 15 | device = torch_utils.select_device(device='cpu' if ONNX_EXPORT else opt.device) 16 | if os.path.exists(out): 17 | shutil.rmtree(out) # delete output folder 18 | os.makedirs(out) # make new output folder 19 | 20 | # Initialize model 21 | model = Darknet(opt.cfg, img_size) 22 | 23 | # Load weights 24 | attempt_download(weights) 25 | if weights.endswith('.pt'): # pytorch format 26 | model.load_state_dict(torch.load(weights, map_location=device)['model']) 27 | else: # darknet format 28 | _ = load_darknet_weights(model, weights) 29 | 30 | # Fuse Conv2d + BatchNorm2d layers 31 | # model.fuse() 32 | 33 | # Eval mode 34 | model.to(device).eval() 35 | 36 | # Export mode 37 | if ONNX_EXPORT: 38 | img = torch.zeros((1, 3) + img_size) # (1, 3, 320, 192) 39 | torch.onnx.export(model, img, 'weights/export.onnx', verbose=True) 40 | return 41 | 42 | # Half precision 43 | half = half and device.type != 'cpu' # half precision only supported on CUDA 44 | if half: 45 | model.half() 46 | 47 | # Set Dataloader 48 | vid_path, vid_writer = None, None 49 | if webcam: 50 | view_img = True 51 | torch.backends.cudnn.benchmark = True # set True to speed up constant image size inference 52 | dataset = LoadStreams(source, img_size=img_size, half=half) 53 | else: 54 | save_img = True 55 | dataset = LoadImages(source, img_size=img_size, half=half) 56 | 57 | # Get classes and colors 58 | classes = load_classes(parse_data_cfg(opt.data)['names']) 59 | colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(len(classes))] 60 | 61 | # Run inference 62 | t0 = time.time() 63 | for path, img, im0s, vid_cap in dataset: 64 | t = time.time() 65 | 66 | # Get detections 67 | img = torch.from_numpy(img).to(device) 68 | if img.ndimension() == 3: 69 | img = img.unsqueeze(0) 70 | pred, _ = model(img) 71 | 72 | if opt.half: 73 | pred = pred.float() 74 | 75 | for i, det in enumerate(non_max_suppression(pred, opt.conf_thres, opt.nms_thres)): # detections per image 76 | if webcam: # batch_size >= 1 77 | p, s, im0 = path[i], '%g: ' % i, im0s[i] 78 | else: 79 | p, s, im0 = path, '', im0s 80 | 81 | save_path = str(Path(out) / Path(p).name) 82 | s += '%gx%g ' % img.shape[2:] # print string 83 | if det is not None and len(det): 84 | # Rescale boxes from img_size to im0 size 85 | det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round() 86 | 87 | # Print results 88 | for c in det[:, -1].unique(): 89 | n = (det[:, -1] == c).sum() # detections per class 90 | s += '%g %ss, ' % (n, classes[int(c)]) # add to string 91 | 92 | # Write results 93 | for *xyxy, conf, _, cls in det: 94 | if save_txt: # Write to file 95 | with open(save_path + '.txt', 'a') as file: 96 | file.write(('%g ' * 6 + '\n') % (*xyxy, cls, conf)) 97 | 98 | if save_img or view_img: # Add bbox to image 99 | label = '%s %.2f' % (classes[int(cls)], conf) 100 | plot_one_box(xyxy, im0, label=label, color=colors[int(cls)]) 101 | 102 | print('%sDone. (%.3fs)' % (s, time.time() - t)) 103 | 104 | # Stream results 105 | if view_img: 106 | cv2.imshow(p, im0) 107 | 108 | # Save results (image with detections) 109 | if save_img: 110 | if dataset.mode == 'images': 111 | cv2.imwrite(save_path, im0) 112 | else: 113 | if vid_path != save_path: # new video 114 | vid_path = save_path 115 | if isinstance(vid_writer, cv2.VideoWriter): 116 | vid_writer.release() # release previous video writer 117 | 118 | fps = vid_cap.get(cv2.CAP_PROP_FPS) 119 | w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) 120 | h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) 121 | vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*opt.fourcc), fps, (w, h)) 122 | vid_writer.write(im0) 123 | 124 | if save_txt or save_img: 125 | print('Results saved to %s' % os.getcwd() + os.sep + out) 126 | if platform == 'darwin': # MacOS 127 | os.system('open ' + out + ' ' + save_path) 128 | 129 | print('Done. (%.3fs)' % (time.time() - t0)) 130 | 131 | 132 | if __name__ == '__main__': 133 | parser = argparse.ArgumentParser() 134 | parser.add_argument('--cfg', type=str, default='cfg/yolov3-spp.cfg', help='cfg file path') 135 | parser.add_argument('--data', type=str, default='data/coco.data', help='coco.data file path') 136 | parser.add_argument('--weights', type=str, default='weights/yolov3-spp.weights', help='path to weights file') 137 | parser.add_argument('--source', type=str, default='data/samples', help='source') # input file/folder, 0 for webcam 138 | parser.add_argument('--output', type=str, default='output', help='output folder') # output folder 139 | parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)') 140 | parser.add_argument('--conf-thres', type=float, default=0.3, help='object confidence threshold') 141 | parser.add_argument('--nms-thres', type=float, default=0.5, help='iou threshold for non-maximum suppression') 142 | parser.add_argument('--fourcc', type=str, default='mp4v', help='output video codec (verify ffmpeg support)') 143 | parser.add_argument('--half', action='store_true', help='half precision FP16 inference') 144 | parser.add_argument('--device', default='', help='device id (i.e. 0 or 0,1) or cpu') 145 | parser.add_argument('--view-img', action='store_true', help='display results') 146 | opt = parser.parse_args() 147 | print(opt) 148 | 149 | with torch.no_grad(): 150 | detect() 151 | -------------------------------------------------------------------------------- /models.py: -------------------------------------------------------------------------------- 1 | import torch.nn.functional as F 2 | 3 | from utils.google_utils import * 4 | from utils.parse_config import * 5 | from utils.utils import * 6 | 7 | ONNX_EXPORT = False 8 | 9 | 10 | def create_modules(module_defs, img_size, arc): 11 | # Constructs module list of layer blocks from module configuration in module_defs 12 | 13 | hyperparams = module_defs.pop(0) 14 | output_filters = [int(hyperparams['channels'])] 15 | module_list = nn.ModuleList() 16 | routs = [] # list of layers which rout to deeper layes 17 | yolo_index = -1 18 | 19 | for i, mdef in enumerate(module_defs): 20 | modules = nn.Sequential() 21 | 22 | if mdef['type'] == 'convolutional': 23 | bn = int(mdef['batch_normalize']) 24 | filters = int(mdef['filters']) 25 | kernel_size = int(mdef['size']) 26 | pad = (kernel_size - 1) // 2 if int(mdef['pad']) else 0 27 | modules.add_module('Conv2d', nn.Conv2d(in_channels=output_filters[-1], 28 | out_channels=filters, 29 | kernel_size=kernel_size, 30 | stride=int(mdef['stride']), 31 | padding=pad, 32 | bias=not bn)) 33 | if bn: 34 | modules.add_module('BatchNorm2d', nn.BatchNorm2d(filters, momentum=0.1)) 35 | if mdef['activation'] == 'leaky': # TODO: activation study https://github.com/ultralytics/yolov3/issues/441 36 | modules.add_module('activation', nn.LeakyReLU(0.1, inplace=True)) 37 | # modules.add_module('activation', nn.PReLU(num_parameters=1, init=0.10)) 38 | # modules.add_module('activation', Swish()) 39 | 40 | elif mdef['type'] == 'maxpool': 41 | kernel_size = int(mdef['size']) 42 | stride = int(mdef['stride']) 43 | maxpool = nn.MaxPool2d(kernel_size=kernel_size, stride=stride, padding=int((kernel_size - 1) // 2)) 44 | if kernel_size == 2 and stride == 1: # yolov3-tiny 45 | modules.add_module('ZeroPad2d', nn.ZeroPad2d((0, 1, 0, 1))) 46 | modules.add_module('MaxPool2d', maxpool) 47 | else: 48 | modules = maxpool 49 | 50 | elif mdef['type'] == 'upsample': 51 | modules = nn.Upsample(scale_factor=int(mdef['stride']), mode='nearest') 52 | 53 | elif mdef['type'] == 'route': # nn.Sequential() placeholder for 'route' layer 54 | layers = [int(x) for x in mdef['layers'].split(',')] 55 | filters = sum([output_filters[i + 1 if i > 0 else i] for i in layers]) 56 | routs.extend([l if l > 0 else l + i for l in layers]) 57 | # if mdef[i+1]['type'] == 'reorg3d': 58 | # modules = nn.Upsample(scale_factor=1/float(mdef[i+1]['stride']), mode='nearest') # reorg3d 59 | 60 | elif mdef['type'] == 'shortcut': # nn.Sequential() placeholder for 'shortcut' layer 61 | filters = output_filters[int(mdef['from'])] 62 | layer = int(mdef['from']) 63 | routs.extend([i + layer if layer < 0 else layer]) 64 | 65 | elif mdef['type'] == 'reorg3d': # yolov3-spp-pan-scale 66 | # torch.Size([16, 128, 104, 104]) 67 | # torch.Size([16, 64, 208, 208]) <-- # stride 2 interpolate dimensions 2 and 3 to cat with prior layer 68 | pass 69 | 70 | elif mdef['type'] == 'yolo': 71 | yolo_index += 1 72 | mask = [int(x) for x in mdef['mask'].split(',')] # anchor mask 73 | modules = YOLOLayer(anchors=mdef['anchors'][mask], # anchor list 74 | nc=int(mdef['classes']), # number of classes 75 | img_size=img_size, # (416, 416) 76 | yolo_index=yolo_index, # 0, 1 or 2 77 | arc=arc) # yolo architecture 78 | 79 | # Initialize preceding Conv2d() bias (https://arxiv.org/pdf/1708.02002.pdf section 3.3) 80 | try: 81 | if arc == 'defaultpw' or arc == 'Fdefaultpw': # default with positive weights 82 | b = [-4, -3.6] # obj, cls 83 | elif arc == 'default': # default no pw (40 cls, 80 obj) 84 | b = [-5.5, -4.0] 85 | elif arc == 'uBCE': # unified BCE (80 classes) 86 | b = [0, -8.5] 87 | elif arc == 'uCE': # unified CE (1 background + 80 classes) 88 | b = [10, -0.1] 89 | elif arc == 'Fdefault': # Focal default no pw (28 cls, 21 obj, no pw) 90 | b = [-2.1, -1.8] 91 | elif arc == 'uFBCE' or arc == 'uFBCEpw': # unified FocalBCE (5120 obj, 80 classes) 92 | b = [0, -6.5] 93 | elif arc == 'uFCE': # unified FocalCE (64 cls, 1 background + 80 classes) 94 | b = [7.7, -1.1] 95 | 96 | bias = module_list[-1][0].bias.view(len(mask), -1) # 255 to 3x85 97 | bias[:, 4] += b[0] - bias[:, 4].mean() # obj 98 | bias[:, 5:] += b[1] - bias[:, 5:].mean() # cls 99 | # bias = torch.load('weights/yolov3-spp.bias.pt')[yolo_index] # list of tensors [3x85, 3x85, 3x85] 100 | module_list[-1][0].bias = torch.nn.Parameter(bias.view(-1)) 101 | # utils.print_model_biases(model) 102 | except: 103 | print('WARNING: smart bias initialization failure.') 104 | 105 | else: 106 | print('Warning: Unrecognized Layer Type: ' + mdef['type']) 107 | 108 | # Register module list and number of output filters 109 | module_list.append(modules) 110 | output_filters.append(filters) 111 | 112 | return module_list, routs, hyperparams 113 | 114 | 115 | class Swish(nn.Module): 116 | def __init__(self): 117 | super(Swish, self).__init__() 118 | 119 | def forward(self, x): 120 | return x * torch.sigmoid(x) 121 | 122 | 123 | class YOLOLayer(nn.Module): 124 | def __init__(self, anchors, nc, img_size, yolo_index, arc): 125 | super(YOLOLayer, self).__init__() 126 | 127 | self.anchors = torch.Tensor(anchors) 128 | self.na = len(anchors) # number of anchors (3) 129 | self.nc = nc # number of classes (80) 130 | self.nx = 0 # initialize number of x gridpoints 131 | self.ny = 0 # initialize number of y gridpoints 132 | self.arc = arc 133 | 134 | if ONNX_EXPORT: # grids must be computed in __init__ 135 | stride = [32, 16, 8][yolo_index] # stride of this layer 136 | nx = int(img_size[1] / stride) # number x grid points 137 | ny = int(img_size[0] / stride) # number y grid points 138 | create_grids(self, img_size, (nx, ny)) 139 | 140 | def forward(self, p, img_size, var=None): 141 | if ONNX_EXPORT: 142 | bs = 1 # batch size 143 | else: 144 | bs, ny, nx = p.shape[0], p.shape[-2], p.shape[-1] 145 | if (self.nx, self.ny) != (nx, ny): 146 | create_grids(self, img_size, (nx, ny), p.device, p.dtype) 147 | 148 | # p.view(bs, 255, 13, 13) -- > (bs, 3, 13, 13, 85) # (bs, anchors, grid, grid, classes + xywh) 149 | p = p.view(bs, self.na, self.nc + 5, self.ny, self.nx).permute(0, 1, 3, 4, 2).contiguous() # prediction 150 | 151 | if self.training: 152 | return p 153 | 154 | elif ONNX_EXPORT: 155 | # Constants CAN NOT BE BROADCAST, ensure correct shape! 156 | ngu = self.ng.repeat((1, self.na * self.nx * self.ny, 1)) 157 | grid_xy = self.grid_xy.repeat((1, self.na, 1, 1, 1)).view((1, -1, 2)) 158 | anchor_wh = self.anchor_wh.repeat((1, 1, self.nx, self.ny, 1)).view((1, -1, 2)) / ngu 159 | 160 | p = p.view(-1, 5 + self.nc) 161 | xy = torch.sigmoid(p[..., 0:2]) + grid_xy[0] # x, y 162 | wh = torch.exp(p[..., 2:4]) * anchor_wh[0] # width, height 163 | p_conf = torch.sigmoid(p[:, 4:5]) # Conf 164 | p_cls = F.softmax(p[:, 5:85], 1) * p_conf # SSD-like conf 165 | return torch.cat((xy / ngu[0], wh, p_conf, p_cls), 1).t() 166 | 167 | # p = p.view(1, -1, 5 + self.nc) 168 | # xy = torch.sigmoid(p[..., 0:2]) + grid_xy # x, y 169 | # wh = torch.exp(p[..., 2:4]) * anchor_wh # width, height 170 | # p_conf = torch.sigmoid(p[..., 4:5]) # Conf 171 | # p_cls = p[..., 5:5 + self.nc] 172 | # # Broadcasting only supported on first dimension in CoreML. See onnx-coreml/_operators.py 173 | # # p_cls = F.softmax(p_cls, 2) * p_conf # SSD-like conf 174 | # p_cls = torch.exp(p_cls).permute((2, 1, 0)) 175 | # p_cls = p_cls / p_cls.sum(0).unsqueeze(0) * p_conf.permute((2, 1, 0)) # F.softmax() equivalent 176 | # p_cls = p_cls.permute(2, 1, 0) 177 | # return torch.cat((xy / ngu, wh, p_conf, p_cls), 2).squeeze().t() 178 | 179 | else: # inference 180 | # s = 1.5 # scale_xy (pxy = pxy * s - (s - 1) / 2) 181 | io = p.clone() # inference output 182 | io[..., 0:2] = torch.sigmoid(io[..., 0:2]) + self.grid_xy # xy 183 | io[..., 2:4] = torch.exp(io[..., 2:4]) * self.anchor_wh # wh yolo method 184 | # io[..., 2:4] = ((torch.sigmoid(io[..., 2:4]) * 2) ** 3) * self.anchor_wh # wh power method 185 | io[..., :4] *= self.stride 186 | 187 | if 'default' in self.arc: # seperate obj and cls 188 | torch.sigmoid_(io[..., 4:]) 189 | elif 'BCE' in self.arc: # unified BCE (80 classes) 190 | torch.sigmoid_(io[..., 5:]) 191 | io[..., 4] = 1 192 | elif 'CE' in self.arc: # unified CE (1 background + 80 classes) 193 | io[..., 4:] = F.softmax(io[..., 4:], dim=4) 194 | io[..., 4] = 1 195 | 196 | if self.nc == 1: 197 | io[..., 5] = 1 # single-class model https://github.com/ultralytics/yolov3/issues/235 198 | 199 | # reshape from [1, 3, 13, 13, 85] to [1, 507, 85] 200 | return io.view(bs, -1, 5 + self.nc), p 201 | 202 | 203 | class Darknet(nn.Module): 204 | # YOLOv3 object detection model 205 | 206 | def __init__(self, cfg, img_size=(416, 416), arc='default'): 207 | super(Darknet, self).__init__() 208 | 209 | if isinstance(cfg, str): 210 | self.module_defs = parse_model_cfg(cfg) 211 | elif isinstance(cfg, list): 212 | self.module_defs = cfg 213 | 214 | self.module_list, self.routs, self.hyperparams = create_modules(self.module_defs, img_size, arc) 215 | self.yolo_layers = get_yolo_layers(self) 216 | 217 | # Darknet Header https://github.com/AlexeyAB/darknet/issues/2914#issuecomment-496675346 218 | self.version = np.array([0, 2, 5], dtype=np.int32) # (int32) version info: major, minor, revision 219 | self.seen = np.array([0], dtype=np.int64) # (int64) number of images seen during training 220 | 221 | def forward(self, x, var=None): 222 | img_size = x.shape[-2:] 223 | layer_outputs = [] 224 | output = [] 225 | 226 | for i, (mdef, module) in enumerate(zip(self.module_defs, self.module_list)): 227 | mtype = mdef['type'] 228 | if mtype in ['convolutional', 'upsample', 'maxpool']: 229 | x = module(x) 230 | elif mtype == 'route': 231 | layers = [int(x) for x in mdef['layers'].split(',')] 232 | if len(layers) == 1: 233 | x = layer_outputs[layers[0]] 234 | else: 235 | try: 236 | x = torch.cat([layer_outputs[i] for i in layers], 1) 237 | except: # apply stride 2 for darknet reorg layer 238 | layer_outputs[layers[1]] = F.interpolate(layer_outputs[layers[1]], scale_factor=[0.5, 0.5]) 239 | x = torch.cat([layer_outputs[i] for i in layers], 1) 240 | # print(''), [print(layer_outputs[i].shape) for i in layers], print(x.shape) 241 | elif mtype == 'shortcut': 242 | x = x + layer_outputs[int(mdef['from'])] 243 | elif mtype == 'yolo': 244 | x = module(x, img_size) 245 | output.append(x) 246 | layer_outputs.append(x if i in self.routs else []) 247 | 248 | if self.training: 249 | return output 250 | elif ONNX_EXPORT: 251 | output = torch.cat(output, 1) # cat 3 layers 85 x (507, 2028, 8112) to 85 x 10647 252 | nc = self.module_list[self.yolo_layers[0]].nc # number of classes 253 | return output[5:5 + nc].t(), output[:4].t() # ONNX scores, boxes 254 | else: 255 | io, p = list(zip(*output)) # inference output, training output 256 | return torch.cat(io, 1), p 257 | 258 | def fuse(self): 259 | # Fuse Conv2d + BatchNorm2d layers throughout model 260 | fused_list = nn.ModuleList() 261 | for a in list(self.children())[0]: 262 | if isinstance(a, nn.Sequential): 263 | for i, b in enumerate(a): 264 | if isinstance(b, nn.modules.batchnorm.BatchNorm2d): 265 | # fuse this bn layer with the previous conv2d layer 266 | conv = a[i - 1] 267 | fused = torch_utils.fuse_conv_and_bn(conv, b) 268 | a = nn.Sequential(fused, *list(a.children())[i + 1:]) 269 | break 270 | fused_list.append(a) 271 | self.module_list = fused_list 272 | # model_info(self) # yolov3-spp reduced from 225 to 152 layers 273 | 274 | 275 | def get_yolo_layers(model): 276 | return [i for i, x in enumerate(model.module_defs) if x['type'] == 'yolo'] # [82, 94, 106] for yolov3 277 | 278 | 279 | def create_grids(self, img_size=416, ng=(13, 13), device='cpu', type=torch.float32): 280 | nx, ny = ng # x and y grid size 281 | self.img_size = max(img_size) 282 | self.stride = self.img_size / max(ng) 283 | 284 | # build xy offsets 285 | yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) 286 | self.grid_xy = torch.stack((xv, yv), 2).to(device).type(type).view((1, 1, ny, nx, 2)) 287 | 288 | # build wh gains 289 | self.anchor_vec = self.anchors.to(device) / self.stride 290 | self.anchor_wh = self.anchor_vec.view(1, self.na, 1, 1, 2).to(device).type(type) 291 | self.ng = torch.Tensor(ng).to(device) 292 | self.nx = nx 293 | self.ny = ny 294 | 295 | 296 | def load_darknet_weights(self, weights, cutoff=-1): 297 | # Parses and loads the weights stored in 'weights' 298 | 299 | # Establish cutoffs (load layers between 0 and cutoff. if cutoff = -1 all are loaded) 300 | file = Path(weights).name 301 | if file == 'darknet53.conv.74': 302 | cutoff = 75 303 | elif file == 'yolov3-tiny.conv.15': 304 | cutoff = 15 305 | 306 | # Read weights file 307 | with open(weights, 'rb') as f: 308 | # Read Header https://github.com/AlexeyAB/darknet/issues/2914#issuecomment-496675346 309 | self.version = np.fromfile(f, dtype=np.int32, count=3) # (int32) version info: major, minor, revision 310 | self.seen = np.fromfile(f, dtype=np.int64, count=1) # (int64) number of images seen during training 311 | 312 | weights = np.fromfile(f, dtype=np.float32) # The rest are weights 313 | 314 | ptr = 0 315 | for i, (mdef, module) in enumerate(zip(self.module_defs[:cutoff], self.module_list[:cutoff])): 316 | if mdef['type'] == 'convolutional': 317 | conv_layer = module[0] 318 | if mdef['batch_normalize']: 319 | # Load BN bias, weights, running mean and running variance 320 | bn_layer = module[1] 321 | num_b = bn_layer.bias.numel() # Number of biases 322 | # Bias 323 | bn_b = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.bias) 324 | bn_layer.bias.data.copy_(bn_b) 325 | ptr += num_b 326 | # Weight 327 | bn_w = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.weight) 328 | bn_layer.weight.data.copy_(bn_w) 329 | ptr += num_b 330 | # Running Mean 331 | bn_rm = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.running_mean) 332 | bn_layer.running_mean.data.copy_(bn_rm) 333 | ptr += num_b 334 | # Running Var 335 | bn_rv = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.running_var) 336 | bn_layer.running_var.data.copy_(bn_rv) 337 | ptr += num_b 338 | # Load conv. weights 339 | num_w = conv_layer.weight.numel() 340 | conv_w = torch.from_numpy(weights[ptr:ptr + num_w]).view_as(conv_layer.weight) 341 | conv_layer.weight.data.copy_(conv_w) 342 | ptr += num_w 343 | else: 344 | if "yolov3.weights" in file: 345 | num_b = 255 346 | ptr += num_b 347 | num_w = int(self.module_defs[i-1]["filters"]) * 255 348 | ptr += num_w 349 | else: 350 | # Load conv. bias 351 | num_b = conv_layer.bias.numel() 352 | conv_b = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(conv_layer.bias) 353 | conv_layer.bias.data.copy_(conv_b) 354 | ptr += num_b 355 | # Load conv. weights 356 | num_w = conv_layer.weight.numel() 357 | conv_w = torch.from_numpy(weights[ptr:ptr + num_w]).view_as(conv_layer.weight) 358 | conv_layer.weight.data.copy_(conv_w) 359 | ptr += num_w 360 | assert ptr == len(weights) 361 | return cutoff 362 | 363 | 364 | def save_weights(self, path='model.weights', cutoff=-1): 365 | # Converts a PyTorch model to Darket format (*.pt to *.weights) 366 | # Note: Does not work if model.fuse() is applied 367 | with open(path, 'wb') as f: 368 | # Write Header https://github.com/AlexeyAB/darknet/issues/2914#issuecomment-496675346 369 | self.version.tofile(f) # (int32) version info: major, minor, revision 370 | self.seen.tofile(f) # (int64) number of images seen during training 371 | 372 | # Iterate through layers 373 | for i, (mdef, module) in enumerate(zip(self.module_defs[:cutoff], self.module_list[:cutoff])): 374 | if mdef['type'] == 'convolutional': 375 | conv_layer = module[0] 376 | # If batch norm, load bn first 377 | if mdef['batch_normalize']: 378 | bn_layer = module[1] 379 | bn_layer.bias.data.cpu().numpy().tofile(f) 380 | bn_layer.weight.data.cpu().numpy().tofile(f) 381 | bn_layer.running_mean.data.cpu().numpy().tofile(f) 382 | bn_layer.running_var.data.cpu().numpy().tofile(f) 383 | # Load conv bias 384 | else: 385 | conv_layer.bias.data.cpu().numpy().tofile(f) 386 | # Load conv weights 387 | conv_layer.weight.data.cpu().numpy().tofile(f) 388 | 389 | 390 | def convert(cfg='cfg/yolov3-spp.cfg', weights='weights/yolov3-spp.weights'): 391 | # Converts between PyTorch and Darknet format per extension (i.e. *.weights convert to *.pt and vice versa) 392 | # from models import *; convert('cfg/yolov3-spp.cfg', 'weights/yolov3-spp.weights') 393 | 394 | # Initialize model 395 | model = Darknet(cfg) 396 | 397 | # Load weights and save 398 | if weights.endswith('.pt'): # if PyTorch format 399 | model.load_state_dict(torch.load(weights, map_location='cpu')['model']) 400 | save_weights(model, path='converted.weights', cutoff=-1) 401 | print("Success: converted '%s' to 'converted.weights'" % weights) 402 | 403 | elif weights.endswith('.weights'): # darknet format 404 | _ = load_darknet_weights(model, weights) 405 | 406 | chkpt = {'epoch': -1, 407 | 'best_fitness': None, 408 | 'training_results': None, 409 | 'model': model.state_dict(), 410 | 'optimizer': None} 411 | 412 | torch.save(chkpt, 'converted.pt') 413 | print("Success: converted '%s' to 'converted.pt'" % weights) 414 | 415 | else: 416 | print('Error: extension not supported.') 417 | 418 | 419 | def attempt_download(weights): 420 | # Attempt to download pretrained weights if not found locally 421 | 422 | msg = weights + ' missing, download from https://drive.google.com/drive/folders/1uxgUBemJVw9wZsdpboYbzUN4bcRhsuAI' 423 | if weights and not os.path.isfile(weights): 424 | file = Path(weights).name 425 | 426 | if file == 'yolov3-spp.weights': 427 | gdrive_download(id='1oPCHKsM2JpM-zgyepQciGli9X0MTsJCO', name=weights) 428 | elif file == 'yolov3-spp.pt': 429 | gdrive_download(id='1vFlbJ_dXPvtwaLLOu-twnjK4exdFiQ73', name=weights) 430 | elif file == 'yolov3.pt': 431 | gdrive_download(id='11uy0ybbOXA2hc-NJkJbbbkDwNX1QZDlz', name=weights) 432 | elif file == 'yolov3-tiny.pt': 433 | gdrive_download(id='1qKSgejNeNczgNNiCn9ZF_o55GFk1DjY_', name=weights) 434 | elif file == 'darknet53.conv.74': 435 | gdrive_download(id='18xqvs_uwAqfTXp-LJCYLYNHBOcrwbrp0', name=weights) 436 | elif file == 'yolov3-tiny.conv.15': 437 | gdrive_download(id='140PnSedCsGGgu3rOD6Ez4oI6cdDzerLC', name=weights) 438 | 439 | else: 440 | try: # download from pjreddie.com 441 | url = 'https://pjreddie.com/media/files/' + file 442 | print('Downloading ' + url) 443 | os.system('curl -f ' + url + ' -o ' + weights) 444 | except IOError: 445 | print(msg) 446 | os.system('rm ' + weights) # remove partial downloads 447 | 448 | assert os.path.exists(weights), msg # download missing weights from Google Drive 449 | -------------------------------------------------------------------------------- /prune.py: -------------------------------------------------------------------------------- 1 | from models import * 2 | from utils.utils import * 3 | import numpy as np 4 | from copy import deepcopy 5 | from test import test 6 | from terminaltables import AsciiTable 7 | import time 8 | from utils.prune_utils import * 9 | import argparse 10 | 11 | 12 | 13 | if __name__ == '__main__': 14 | parser = argparse.ArgumentParser() 15 | parser.add_argument('--cfg', type=str, default='cfg/yolov3.cfg', help='cfg file path') 16 | parser.add_argument('--data', type=str, default='data/coco.data', help='*.data file path') 17 | parser.add_argument('--weights', type=str, default='weights/last.pt', help='sparse model weights') 18 | parser.add_argument('--percent', type=float, default=0.8, help='channel prune percent') 19 | opt = parser.parse_args() 20 | print(opt) 21 | 22 | 23 | #%% 24 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 25 | model = Darknet(opt.cfg).to(device) 26 | if opt.weights.endswith('.pt'): 27 | model.load_state_dict(torch.load(opt.weights)['model']) 28 | else: 29 | load_darknet_weights(model, opt.weights) 30 | print('\nloaded weights from ',opt.weights) 31 | 32 | eval_model = lambda model:test(opt.cfg, opt.data, 33 | weights=opt.weights, 34 | batch_size=16, 35 | img_size=416, 36 | iou_thres=0.5, 37 | conf_thres=0.001, 38 | nms_thres=0.5, 39 | save_json=False, 40 | model=model) 41 | obtain_num_parameters = lambda model:sum([param.nelement() for param in model.parameters()]) 42 | 43 | print("\nlet's test the original model first:") 44 | origin_model_metric = eval_model(model) 45 | 46 | origin_nparameters = obtain_num_parameters(model) 47 | 48 | CBL_idx, Conv_idx, prune_idx= parse_module_defs(model.module_defs) 49 | 50 | bn_weights = gather_bn_weights(model.module_list, prune_idx) 51 | 52 | sorted_bn = torch.sort(bn_weights)[0] 53 | 54 | # 避免剪掉所有channel的最高阈值(每个BN层的gamma的最大值的最小值即为阈值上限) 55 | highest_thre = [] 56 | for idx in prune_idx: 57 | highest_thre.append(model.module_list[idx][1].weight.data.abs().max().item()) 58 | highest_thre = min(highest_thre) 59 | 60 | # 找到highest_thre对应的下标对应的百分比 61 | percent_limit = (sorted_bn==highest_thre).nonzero().item()/len(bn_weights) 62 | 63 | print(f'Gamma threshold should be less than {highest_thre:.4f}.') 64 | print(f'The corresponding prune ratio is {percent_limit:.3f}.') 65 | 66 | #%% 67 | def prune_and_eval(model, sorted_bn, percent=.0): 68 | model_copy = deepcopy(model) 69 | thre_index = int(len(sorted_bn) * percent) 70 | thre = sorted_bn[thre_index] 71 | 72 | print(f'Gamma value that less than {thre:.4f} are set to zero!') 73 | 74 | remain_num = 0 75 | for idx in prune_idx: 76 | 77 | bn_module = model_copy.module_list[idx][1] 78 | 79 | mask = obtain_bn_mask(bn_module, thre) 80 | 81 | remain_num += int(mask.sum()) 82 | bn_module.weight.data.mul_(mask) 83 | print("let's test the current model!") 84 | mAP = eval_model(model_copy)[0][2] 85 | 86 | 87 | print(f'Number of channels has been reduced from {len(sorted_bn)} to {remain_num}') 88 | print(f'Prune ratio: {1-remain_num/len(sorted_bn):.3f}') 89 | print(f"mAP of the 'pruned' model is {mAP:.4f}") 90 | 91 | return thre 92 | 93 | percent = opt.percent 94 | print('the required prune percent is', percent) 95 | threshold = prune_and_eval(model, sorted_bn, percent) 96 | #%% 97 | def obtain_filters_mask(model, thre, CBL_idx, prune_idx): 98 | 99 | pruned = 0 100 | total = 0 101 | num_filters = [] 102 | filters_mask = [] 103 | for idx in CBL_idx: 104 | bn_module = model.module_list[idx][1] 105 | if idx in prune_idx: 106 | 107 | mask = obtain_bn_mask(bn_module, thre).cpu().numpy() 108 | remain = int(mask.sum()) 109 | pruned = pruned + mask.shape[0] - remain 110 | 111 | if remain == 0: 112 | print("Channels would be all pruned!") 113 | raise Exception 114 | 115 | print(f'layer index: {idx:>3d} \t total channel: {mask.shape[0]:>4d} \t ' 116 | f'remaining channel: {remain:>4d}') 117 | else: 118 | mask = np.ones(bn_module.weight.data.shape) 119 | remain = mask.shape[0] 120 | 121 | total += mask.shape[0] 122 | num_filters.append(remain) 123 | filters_mask.append(mask.copy()) 124 | 125 | prune_ratio = pruned / total 126 | print(f'Prune channels: {pruned}\tPrune ratio: {prune_ratio:.3f}') 127 | 128 | return num_filters, filters_mask 129 | 130 | num_filters, filters_mask = obtain_filters_mask(model, threshold, CBL_idx, prune_idx) 131 | 132 | #%% 133 | CBLidx2mask = {idx: mask for idx, mask in zip(CBL_idx, filters_mask)} 134 | 135 | pruned_model = prune_model_keep_size(model, prune_idx, CBL_idx, CBLidx2mask) 136 | 137 | print("\nnow prune the model but keep size,(actually add offset of BN beta to next layer), let's see how the mAP goes") 138 | eval_model(pruned_model) 139 | 140 | 141 | #%% 142 | compact_module_defs = deepcopy(model.module_defs) 143 | for idx, num in zip(CBL_idx, num_filters): 144 | assert compact_module_defs[idx]['type'] == 'convolutional' 145 | compact_module_defs[idx]['filters'] = str(num) 146 | 147 | #%% 148 | compact_model = Darknet([model.hyperparams.copy()] + compact_module_defs).to(device) 149 | compact_nparameters = obtain_num_parameters(compact_model) 150 | 151 | init_weights_from_loose_model(compact_model, pruned_model, CBL_idx, Conv_idx, CBLidx2mask) 152 | 153 | #%% 154 | random_input = torch.rand((1, 3, 416, 416)).to(device) 155 | 156 | def obtain_avg_forward_time(input, model, repeat=200): 157 | 158 | model.eval() 159 | start = time.time() 160 | with torch.no_grad(): 161 | for i in range(repeat): 162 | output = model(input)[0] 163 | avg_infer_time = (time.time() - start) / repeat 164 | 165 | return avg_infer_time, output 166 | 167 | print('\ntesting avg forward time...') 168 | pruned_forward_time, pruned_output = obtain_avg_forward_time(random_input, pruned_model) 169 | compact_forward_time, compact_output = obtain_avg_forward_time(random_input, compact_model) 170 | 171 | diff = (pruned_output-compact_output).abs().gt(0.001).sum().item() 172 | if diff > 0: 173 | print('Something wrong with the pruned model!') 174 | 175 | #%% 176 | # 在测试集上测试剪枝后的模型, 并统计模型的参数数量 177 | print('testing the mAP of final pruned model') 178 | compact_model_metric = eval_model(compact_model) 179 | 180 | 181 | #%% 182 | # 比较剪枝前后参数数量的变化、指标性能的变化 183 | metric_table = [ 184 | ["Metric", "Before", "After"], 185 | ["mAP", f'{origin_model_metric[0][2]:.6f}', f'{compact_model_metric[0][2]:.6f}'], 186 | ["Parameters", f"{origin_nparameters}", f"{compact_nparameters}"], 187 | ["Inference", f'{pruned_forward_time:.4f}', f'{compact_forward_time:.4f}'] 188 | ] 189 | print(AsciiTable(metric_table).table) 190 | 191 | #%% 192 | # 生成剪枝后的cfg文件并保存模型 193 | pruned_cfg_name = opt.cfg.replace('/', f'/prune_{percent}_') 194 | pruned_cfg_file = write_cfg(pruned_cfg_name, [model.hyperparams.copy()] + compact_module_defs) 195 | print(f'Config file has been saved: {pruned_cfg_file}') 196 | 197 | compact_model_name = opt.weights.replace('/', f'/prune_{percent}_') 198 | if compact_model_name.endswith('.pt'): 199 | compact_model_name = compact_model_name.replace('.pt', '.weights') 200 | save_weights(compact_model, compact_model_name) 201 | print(f'Compact model has been saved: {compact_model_name}') 202 | 203 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | # pip3 install -U -r requirements.txt 2 | numpy 3 | opencv-python 4 | torch >= 1.2 5 | matplotlib 6 | pycocotools 7 | tqdm 8 | tb-nightly 9 | future 10 | Pillow 11 | tensorboard >=1.13 12 | # Equivalent conda commands ---------------------------------------------------- 13 | # conda update -n base -c defaults conda 14 | # conda install -yc anaconda future numpy opencv matplotlib tqdm pillow 15 | # conda install -yc conda-forge scikit-image tensorboard pycocotools 16 | # conda install -yc spyder-ide spyder-line-profiler 17 | # conda install -yc pytorch pytorch torchvision 18 | -------------------------------------------------------------------------------- /shortcut_prune.py: -------------------------------------------------------------------------------- 1 | from models import * 2 | from utils.utils import * 3 | import numpy as np 4 | from copy import deepcopy 5 | from test import test 6 | from terminaltables import AsciiTable 7 | import time 8 | from utils.prune_utils import * 9 | import argparse 10 | 11 | 12 | 13 | if __name__ == '__main__': 14 | parser = argparse.ArgumentParser() 15 | parser.add_argument('--cfg', type=str, default='cfg/yolov3.cfg', help='cfg file path') 16 | parser.add_argument('--data', type=str, default='data/coco.data', help='*.data file path') 17 | parser.add_argument('--weights', type=str, default='weights/last.pt', help='sparse model weights') 18 | parser.add_argument('--percent', type=float, default=0.8, help='channel prune percent') 19 | opt = parser.parse_args() 20 | print(opt) 21 | 22 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 23 | model = Darknet(opt.cfg).to(device) 24 | 25 | if opt.weights.endswith(".pt"): 26 | model.load_state_dict(torch.load(opt.weights, map_location=device)['model']) 27 | else: 28 | _ = load_darknet_weights(model, opt.weights) 29 | print('\nloaded weights from ',opt.weights) 30 | 31 | 32 | eval_model = lambda model:test(model=model,cfg=opt.cfg, data=opt.data) 33 | obtain_num_parameters = lambda model:sum([param.nelement() for param in model.parameters()]) 34 | 35 | print("\nlet's test the original model first:") 36 | origin_model_metric = eval_model(model) 37 | origin_nparameters = obtain_num_parameters(model) 38 | 39 | CBL_idx, Conv_idx, prune_idx,shortcut_idx,shortcut_all= parse_module_defs2(model.module_defs) 40 | 41 | 42 | sort_prune_idx=[idx for idx in prune_idx if idx not in shortcut_idx] 43 | 44 | #将所有要剪枝的BN层的α参数，拷贝到bn_weights列表 45 | bn_weights = gather_bn_weights(model.module_list, sort_prune_idx) 46 | 47 | #torch.sort返回二维列表，第一维是排序后的值列表，第二维是排序后的值列表对应的索引 48 | sorted_bn = torch.sort(bn_weights)[0] 49 | 50 | 51 | #避免剪掉所有channel的最高阈值(每个BN层的gamma的最大值的最小值即为阈值上限) 52 | highest_thre = [] 53 | for idx in sort_prune_idx: 54 | #.item()可以得到张量里的元素值 55 | highest_thre.append(model.module_list[idx][1].weight.data.abs().max().item()) 56 | highest_thre = min(highest_thre) 57 | 58 | # 找到highest_thre对应的下标对应的百分比 59 | percent_limit = (sorted_bn==highest_thre).nonzero().item()/len(bn_weights) 60 | 61 | print(f'Threshold should be less than {highest_thre:.4f}.') 62 | print(f'The corresponding prune ratio is {percent_limit:.3f}.') 63 | 64 | 65 | def prune_and_eval(model, sorted_bn, percent=.0): 66 | model_copy = deepcopy(model) 67 | thre_index = int(len(sorted_bn) * percent) 68 | #获得α参数的阈值，小于该值的α参数对应的通道，全部裁剪掉 69 | thre1 = sorted_bn[thre_index] 70 | 71 | print(f'Channels with Gamma value less than {thre1:.6f} are pruned!') 72 | 73 | remain_num = 0 74 | idx_new=dict() 75 | for idx in prune_idx: 76 | 77 | if idx not in shortcut_idx: 78 | 79 | bn_module = model_copy.module_list[idx][1] 80 | 81 | mask = obtain_bn_mask(bn_module, thre1) 82 | #记录剪枝后，每一层卷积层对应的mask 83 | # idx_new[idx]=mask.cpu().numpy() 84 | idx_new[idx]=mask 85 | remain_num += int(mask.sum()) 86 | bn_module.weight.data.mul_(mask) 87 | #bn_module.bias.data.mul_(mask*0.0001) 88 | else: 89 | 90 | bn_module = model_copy.module_list[idx][1] 91 | 92 | 93 | mask=idx_new[shortcut_idx[idx]] 94 | idx_new[idx]=mask 95 | 96 | 97 | remain_num += int(mask.sum()) 98 | bn_module.weight.data.mul_(mask) 99 | 100 | #print(int(mask.sum())) 101 | 102 | with torch.no_grad(): 103 | mAP = eval_model(model_copy)[0][2] 104 | 105 | print(f'Number of channels has been reduced from {len(sorted_bn)} to {remain_num}') 106 | print(f'Prune ratio: {1-remain_num/len(sorted_bn):.3f}') 107 | print(f'mAP of the pruned model is {mAP:.4f}') 108 | 109 | return thre1 110 | 111 | percent = opt.percent 112 | threshold = prune_and_eval(model, sorted_bn, percent) 113 | 114 | 115 | 116 | #**************************************************************** 117 | #虽然上面已经能看到剪枝后的效果，但是没有生成剪枝后的模型结构，因此下面的代码是为了生成新的模型结构并拷贝旧模型参数到新模型 118 | 119 | 120 | 121 | #%% 122 | def obtain_filters_mask(model, thre, CBL_idx, prune_idx): 123 | 124 | pruned = 0 125 | total = 0 126 | num_filters = [] 127 | filters_mask = [] 128 | idx_new=dict() 129 | #CBL_idx存储的是所有带BN的卷积层（YOLO层的前一层卷积层是不带BN的） 130 | for idx in CBL_idx: 131 | bn_module = model.module_list[idx][1] 132 | if idx in prune_idx: 133 | if idx not in shortcut_idx: 134 | 135 | mask = obtain_bn_mask(bn_module, thre).cpu().numpy() 136 | idx_new[idx]=mask 137 | remain = int(mask.sum()) 138 | pruned = pruned + mask.shape[0] - remain 139 | 140 | # if remain == 0: 141 | # print("Channels would be all pruned!") 142 | # raise Exception 143 | 144 | # print(f'layer index: {idx:>3d} \t total channel: {mask.shape[0]:>4d} \t ' 145 | # f'remaining channel: {remain:>4d}') 146 | else: 147 | mask=idx_new[shortcut_idx[idx]] 148 | idx_new[idx]=mask 149 | remain= int(mask.sum()) 150 | pruned = pruned + mask.shape[0] - remain 151 | 152 | if remain == 0: 153 | print("Channels would be all pruned!") 154 | raise Exception 155 | 156 | print(f'layer index: {idx:>3d} \t total channel: {mask.shape[0]:>4d} \t ' 157 | f'remaining channel: {remain:>4d}') 158 | else: 159 | mask = np.ones(bn_module.weight.data.shape) 160 | remain = mask.shape[0] 161 | 162 | total += mask.shape[0] 163 | num_filters.append(remain) 164 | filters_mask.append(mask.copy()) 165 | 166 | #因此，这里求出的prune_ratio,需要裁剪的α参数/cbl_idx中所有的α参数 167 | prune_ratio = pruned / total 168 | print(f'Prune channels: {pruned}\tPrune ratio: {prune_ratio:.3f}') 169 | 170 | return num_filters, filters_mask 171 | 172 | num_filters, filters_mask = obtain_filters_mask(model, threshold, CBL_idx, prune_idx) 173 | 174 | 175 | #CBLidx2mask存储CBL_idx中，每一层BN层对应的mask 176 | CBLidx2mask = {idx: mask for idx, mask in zip(CBL_idx, filters_mask)} 177 | 178 | 179 | pruned_model = prune_model_keep_size2(model, prune_idx, CBL_idx, CBLidx2mask) 180 | print("\nnow prune the model but keep size,(actually add offset of BN beta to next layer), let's see how the mAP goes") 181 | 182 | 183 | eval_model(pruned_model) 184 | 185 | 186 | 187 | #获得原始模型的module_defs，并修改该defs中的卷积核数量 188 | compact_module_defs = deepcopy(model.module_defs) 189 | for idx, num in zip(CBL_idx, num_filters): 190 | assert compact_module_defs[idx]['type'] == 'convolutional' 191 | compact_module_defs[idx]['filters'] = str(num) 192 | 193 | 194 | compact_model = Darknet([model.hyperparams.copy()] + compact_module_defs).to(device) 195 | compact_nparameters = obtain_num_parameters(compact_model) 196 | 197 | init_weights_from_loose_model(compact_model, pruned_model, CBL_idx, Conv_idx, CBLidx2mask) 198 | 199 | 200 | random_input = torch.rand((1, 3, 416, 416)).to(device) 201 | 202 | def obtain_avg_forward_time(input, model, repeat=200): 203 | 204 | model.eval() 205 | start = time.time() 206 | with torch.no_grad(): 207 | for i in range(repeat): 208 | output = model(input) 209 | avg_infer_time = (time.time() - start) / repeat 210 | 211 | return avg_infer_time, output 212 | 213 | pruned_forward_time, pruned_output = obtain_avg_forward_time(random_input, pruned_model) 214 | compact_forward_time, compact_output = obtain_avg_forward_time(random_input, compact_model) 215 | 216 | 217 | # 在测试集上测试剪枝后的模型, 并统计模型的参数数量 218 | compact_model_metric = eval_model(compact_model) 219 | 220 | 221 | # 比较剪枝前后参数数量的变化、指标性能的变化 222 | metric_table = [ 223 | ["Metric", "Before", "After"], 224 | ["mAP", f'{origin_model_metric[0][2]:.6f}', f'{compact_model_metric[0][2]:.6f}'], 225 | ["Parameters", f"{origin_nparameters}", f"{compact_nparameters}"], 226 | ["Inference", f'{pruned_forward_time:.4f}', f'{compact_forward_time:.4f}'] 227 | ] 228 | print(AsciiTable(metric_table).table) 229 | 230 | 231 | # 生成剪枝后的cfg文件并保存模型 232 | pruned_cfg_name = opt.cfg.replace('/', f'/prune_{percent}_') 233 | pruned_cfg_file = write_cfg(pruned_cfg_name, [model.hyperparams.copy()] + compact_module_defs) 234 | print(f'Config file has been saved: {pruned_cfg_file}') 235 | 236 | compact_model_name = opt.weights.replace('/', f'/prune_{percent}_') 237 | if compact_model_name.endswith('.pt'): 238 | compact_model_name = compact_model_name.replace('.pt', '.weights') 239 | save_weights(compact_model, path=compact_model_name) 240 | print(f'Compact model has been saved: {compact_model_name}') 241 | 242 | -------------------------------------------------------------------------------- /test.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import json 3 | 4 | from torch.utils.data import DataLoader 5 | 6 | from models import * 7 | from utils.datasets import * 8 | from utils.utils import * 9 | 10 | 11 | def test(cfg, 12 | data, 13 | weights=None, 14 | batch_size=16, 15 | img_size=416, 16 | iou_thres=0.5, 17 | conf_thres=0.001, 18 | nms_thres=0.5, 19 | save_json=False, 20 | model=None): 21 | with torch.no_grad(): 22 | # Initialize/load model and set device 23 | if model is None: 24 | device = torch_utils.select_device(opt.device) 25 | verbose = True 26 | 27 | # Initialize model 28 | model = Darknet(cfg, img_size).to(device) 29 | 30 | # Load weights 31 | attempt_download(weights) 32 | if weights.endswith('.pt'): # pytorch format 33 | model.load_state_dict(torch.load(weights, map_location=device)['model']) 34 | else: # darknet format 35 | _ = load_darknet_weights(model, weights) 36 | 37 | if torch.cuda.device_count() > 1: 38 | model = nn.DataParallel(model) 39 | else: 40 | device = next(model.parameters()).device # get model device 41 | verbose = False 42 | 43 | # Configure run 44 | data = parse_data_cfg(data) 45 | nc = int(data['classes']) # number of classes 46 | test_path = data['valid'] # path to test images 47 | names = load_classes(data['names']) # class names 48 | 49 | # Dataloader 50 | dataset = LoadImagesAndLabels(test_path, img_size, batch_size) 51 | dataloader = DataLoader(dataset, 52 | batch_size=batch_size, 53 | num_workers=min([os.cpu_count(), batch_size, 16]), 54 | pin_memory=True, 55 | collate_fn=dataset.collate_fn) 56 | 57 | seen = 0 58 | model.eval() 59 | coco91class = coco80_to_coco91_class() 60 | s = ('%20s' + '%10s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', 'mAP', 'F1') 61 | p, r, f1, mp, mr, map, mf1 = 0., 0., 0., 0., 0., 0., 0. 62 | loss = torch.zeros(3) 63 | jdict, stats, ap, ap_class = [], [], [], [] 64 | for batch_i, (imgs, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): 65 | targets = targets.to(device) 66 | imgs = imgs.to(device) 67 | _, _, height, width = imgs.shape # batch size, channels, height, width 68 | 69 | # Plot images with bounding boxes 70 | if batch_i == 0 and not os.path.exists('test_batch0.jpg'): 71 | plot_images(imgs=imgs, targets=targets, paths=paths, fname='test_batch0.jpg') 72 | 73 | # Run model 74 | inf_out, train_out = model(imgs) # inference and training outputs 75 | 76 | # Compute loss 77 | if hasattr(model, 'hyp'): # if model has loss hyperparameters 78 | loss += compute_loss(train_out, targets, model)[1][:3].cpu() # GIoU, obj, cls 79 | 80 | # Run NMS 81 | output = non_max_suppression(inf_out, conf_thres=conf_thres, nms_thres=nms_thres) 82 | 83 | # Statistics per image 84 | for si, pred in enumerate(output): 85 | labels = targets[targets[:, 0] == si, 1:] 86 | nl = len(labels) 87 | tcls = labels[:, 0].tolist() if nl else [] # target class 88 | seen += 1 89 | 90 | if pred is None: 91 | if nl: 92 | stats.append(([], torch.Tensor(), torch.Tensor(), tcls)) 93 | continue 94 | 95 | # Append to text file 96 | # with open('test.txt', 'a') as file: 97 | # [file.write('%11.5g' * 7 % tuple(x) + '\n') for x in pred] 98 | 99 | # Append to pycocotools JSON dictionary 100 | if save_json: 101 | # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... 102 | image_id = int(Path(paths[si]).stem.split('_')[-1]) 103 | box = pred[:, :4].clone() # xyxy 104 | scale_coords(imgs[si].shape[1:], box, shapes[si]) # to original shape 105 | box = xyxy2xywh(box) # xywh 106 | box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner 107 | for di, d in enumerate(pred): 108 | jdict.append({'image_id': image_id, 109 | 'category_id': coco91class[int(d[6])], 110 | 'bbox': [floatn(x, 3) for x in box[di]], 111 | 'score': floatn(d[4], 5)}) 112 | 113 | # Clip boxes to image bounds 114 | clip_coords(pred, (height, width)) 115 | 116 | # Assign all predictions as incorrect 117 | correct = [0] * len(pred) 118 | if nl: 119 | detected = [] 120 | tcls_tensor = labels[:, 0] 121 | 122 | # target boxes 123 | tbox = xywh2xyxy(labels[:, 1:5]) 124 | tbox[:, [0, 2]] *= width 125 | tbox[:, [1, 3]] *= height 126 | 127 | # Search for correct predictions 128 | for i, (*pbox, pconf, pcls_conf, pcls) in enumerate(pred): 129 | 130 | # Break if all targets already located in image 131 | if len(detected) == nl: 132 | break 133 | 134 | # Continue if predicted class not among image classes 135 | if pcls.item() not in tcls: 136 | continue 137 | 138 | # Best iou, index between pred and targets 139 | m = (pcls == tcls_tensor).nonzero().view(-1) 140 | iou, bi = bbox_iou(pbox, tbox[m]).max(0) 141 | 142 | # If iou > threshold and class is correct mark as correct 143 | if iou > iou_thres and m[bi] not in detected: # and pcls == tcls[bi]: 144 | correct[i] = 1 145 | detected.append(m[bi]) 146 | 147 | # Append statistics (correct, conf, pcls, tcls) 148 | stats.append((correct, pred[:, 4].cpu(), pred[:, 6].cpu(), tcls)) 149 | 150 | # Compute statistics 151 | stats = [np.concatenate(x, 0) for x in list(zip(*stats))] # to numpy 152 | if len(stats): 153 | p, r, ap, f1, ap_class = ap_per_class(*stats) 154 | mp, mr, map, mf1 = p.mean(), r.mean(), ap.mean(), f1.mean() 155 | nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class 156 | else: 157 | nt = torch.zeros(1) 158 | 159 | # Print results 160 | pf = '%20s' + '%10.3g' * 6 # print format 161 | print(pf % ('all', seen, nt.sum(), mp, mr, map, mf1)) 162 | 163 | # Print results per class 164 | if verbose and nc > 1 and len(stats): 165 | for i, c in enumerate(ap_class): 166 | print(pf % (names[c], seen, nt[c], p[i], r[i], ap[i], f1[i])) 167 | 168 | # Save JSON 169 | if save_json and map and len(jdict): 170 | try: 171 | imgIds = [int(Path(x).stem.split('_')[-1]) for x in dataset.img_files] 172 | with open('results.json', 'w') as file: 173 | json.dump(jdict, file) 174 | 175 | from pycocotools.coco import COCO 176 | from pycocotools.cocoeval import COCOeval 177 | 178 | # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb 179 | cocoGt = COCO('../coco/annotations/instances_val2014.json') # initialize COCO ground truth api 180 | cocoDt = cocoGt.loadRes('results.json') # initialize COCO pred api 181 | 182 | cocoEval = COCOeval(cocoGt, cocoDt, 'bbox') 183 | cocoEval.params.imgIds = imgIds # [:32] # only evaluate these images 184 | cocoEval.evaluate() 185 | cocoEval.accumulate() 186 | cocoEval.summarize() 187 | map = cocoEval.stats[1] # update mAP to pycocotools mAP 188 | except: 189 | print('WARNING: missing dependency pycocotools from requirements.txt. Can not compute official COCO mAP.') 190 | 191 | # Return results 192 | maps = np.zeros(nc) + map 193 | for i, c in enumerate(ap_class): 194 | maps[c] = ap[i] 195 | return (mp, mr, map, mf1, *(loss / len(dataloader)).tolist()), maps 196 | 197 | 198 | if __name__ == '__main__': 199 | parser = argparse.ArgumentParser(prog='test.py') 200 | parser.add_argument('--cfg', type=str, default='cfg/yolov3-spp.cfg', help='cfg file path') 201 | parser.add_argument('--data', type=str, default='data/coco.data', help='coco.data file path') 202 | parser.add_argument('--weights', type=str, default='weights/yolov3-spp.weights', help='path to weights file') 203 | parser.add_argument('--batch-size', type=int, default=16, help='size of each image batch') 204 | parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)') 205 | parser.add_argument('--iou-thres', type=float, default=0.5, help='iou threshold required to qualify as detected') 206 | parser.add_argument('--conf-thres', type=float, default=0.001, help='object confidence threshold') 207 | parser.add_argument('--nms-thres', type=float, default=0.5, help='iou threshold for non-maximum suppression') 208 | parser.add_argument('--save-json', action='store_true', help='save a cocoapi-compatible JSON results file') 209 | parser.add_argument('--device', default='', help='device id (i.e. 0 or 0,1) or cpu') 210 | opt = parser.parse_args() 211 | print(opt) 212 | 213 | test(opt.cfg, 214 | opt.data, 215 | opt.weights, 216 | opt.batch_size, 217 | opt.img_size, 218 | opt.iou_thres, 219 | opt.conf_thres, 220 | opt.nms_thres, 221 | opt.save_json) 222 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | 3 | import torch.distributed as dist 4 | import torch.optim as optim 5 | import torch.optim.lr_scheduler as lr_scheduler 6 | 7 | import test # import test.py to get mAP after each epoch 8 | from models import * 9 | from utils.datasets import * 10 | from utils.utils import * 11 | from utils.prune_utils import * 12 | 13 | 14 | mixed_precision = True 15 | try: # Mixed precision training https://github.com/NVIDIA/apex 16 | from apex import amp 17 | except: 18 | mixed_precision = False # not installed 19 | 20 | wdir = 'weights' + os.sep # weights dir 21 | last = wdir + 'last.pt' 22 | best = wdir + 'best.pt' 23 | results_file = 'results.txt' 24 | 25 | # Hyperparameters (j-series, 50.5 mAP yolov3-320) evolved by @ktian08 https://github.com/ultralytics/yolov3/issues/310 26 | hyp = {'giou': 1.582, # giou loss gain 27 | 'cls': 27.76, # cls loss gain (CE=~1.0, uCE=~20) 28 | 'cls_pw': 1.446, # cls BCELoss positive_weight 29 | 'obj': 21.35, # obj loss gain (*=80 for uBCE with 80 classes) 30 | 'obj_pw': 3.941, # obj BCELoss positive_weight 31 | 'iou_t': 0.2635, # iou training threshold 32 | 'lr0': 0.002324, # initial learning rate (SGD=1E-3, Adam=9E-5) 33 | 'lrf': -4., # final LambdaLR learning rate = lr0 * (10 ** lrf) 34 | 'momentum': 0.97, # SGD momentum 35 | 'weight_decay': 0.0004569, # optimizer weight decay 36 | 'fl_gamma': 0.5, # focal loss gamma 37 | 'hsv_h': 0.10, # image HSV-Hue augmentation (fraction) 38 | 'hsv_s': 0.5703, # image HSV-Saturation augmentation (fraction) 39 | 'hsv_v': 0.3174, # image HSV-Value augmentation (fraction) 40 | 'degrees': 1.113, # image rotation (+/- deg) 41 | 'translate': 0.06797, # image translation (+/- fraction) 42 | 'scale': 0.1059, # image scale (+/- gain) 43 | 'shear': 0.5768} # image shear (+/- deg) 44 | 45 | # Overwrite hyp with hyp*.txt (optional) 46 | f = glob.glob('hyp*.txt') 47 | if f: 48 | for k, v in zip(hyp.keys(), np.loadtxt(f[0])): 49 | hyp[k] = v 50 | 51 | 52 | def train(): 53 | cfg = opt.cfg 54 | data = opt.data 55 | img_size = opt.img_size 56 | epochs = 1 if opt.prebias else opt.epochs # 500200 batches at bs 64, 117263 images = 273 epochs 57 | batch_size = opt.batch_size 58 | accumulate = opt.accumulate # effective bs = batch_size * accumulate = 16 * 4 = 64 59 | weights = opt.weights # initial training weights 60 | 61 | if 'pw' not in opt.arc: # remove BCELoss positive weights 62 | hyp['cls_pw'] = 1. 63 | hyp['obj_pw'] = 1. 64 | 65 | # Initialize 66 | init_seeds() 67 | multi_scale = opt.multi_scale 68 | 69 | if multi_scale: 70 | img_sz_min = round(img_size / 32 / 1.5) + 1 71 | img_sz_max = round(img_size / 32 * 1.5) - 1 72 | img_size = img_sz_max * 32 # initiate with maximum multi_scale size 73 | print('Using multi-scale %g - %g' % (img_sz_min * 32, img_size)) 74 | 75 | # Configure run 76 | data_dict = parse_data_cfg(data) 77 | train_path = data_dict['train'] 78 | nc = int(data_dict['classes']) # number of classes 79 | 80 | # Remove previous results 81 | for f in glob.glob('*_batch*.jpg') + glob.glob(results_file): 82 | os.remove(f) 83 | 84 | # Initialize model 85 | model = Darknet(cfg, arc=opt.arc).to(device) 86 | 87 | # Optimizer 88 | pg0, pg1 = [], [] # optimizer parameter groups 89 | for k, v in dict(model.named_parameters()).items(): 90 | if 'Conv2d.weight' in k: 91 | pg1 += [v] # parameter group 1 (apply weight_decay) 92 | else: 93 | pg0 += [v] # parameter group 0 94 | 95 | if opt.adam: 96 | optimizer = optim.Adam(pg0, lr=hyp['lr0']) 97 | # optimizer = AdaBound(pg0, lr=hyp['lr0'], final_lr=0.1) 98 | else: 99 | optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) 100 | optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay 101 | del pg0, pg1 102 | 103 | cutoff = -1 # backbone reaches to cutoff layer 104 | start_epoch = 0 105 | best_fitness = 0. 106 | attempt_download(weights) 107 | if weights.endswith('.pt'): # pytorch format 108 | # possible weights are 'last.pt', 'yolov3-spp.pt', 'yolov3-tiny.pt' etc. 109 | if opt.bucket: 110 | os.system('gsutil cp gs://%s/last.pt %s' % (opt.bucket, last)) # download from bucket 111 | chkpt = torch.load(weights, map_location=device) 112 | 113 | # load model 114 | # if opt.transfer: 115 | chkpt['model'] = {k: v for k, v in chkpt['model'].items() if model.state_dict()[k].numel() == v.numel()} 116 | model.load_state_dict(chkpt['model'], strict=False) 117 | print('loaded weights from', weights) 118 | # else: 119 | # model.load_state_dict(chkpt['model']) 120 | 121 | # load optimizer 122 | if chkpt['optimizer'] is not None: 123 | optimizer.load_state_dict(chkpt['optimizer']) 124 | best_fitness = chkpt['best_fitness'] 125 | 126 | # load results 127 | if chkpt.get('training_results') is not None: 128 | with open(results_file, 'w') as file: 129 | file.write(chkpt['training_results']) # write results.txt 130 | 131 | start_epoch = chkpt['epoch'] + 1 132 | del chkpt 133 | 134 | # elif weights.endswith('.pth'): 135 | # model.load_state_dict(torch.load(weights)) 136 | 137 | elif len(weights) > 0: # darknet format 138 | # possible weights are 'yolov3.weights', 'yolov3-tiny.conv.15', 'darknet53.conv.74' etc. 139 | cutoff = load_darknet_weights(model, weights) 140 | print('loaded weights from', weights) 141 | 142 | if opt.prune==1: 143 | print('shortcut sparse training') 144 | _, _, prune_idx, shortcut_idx, _=parse_module_defs2(model.module_defs) 145 | prune_idx = [idx for idx in prune_idx if idx not in shortcut_idx] 146 | elif opt.prune==0: 147 | print('normal sparse training ') 148 | _, _, prune_idx= parse_module_defs(model.module_defs) 149 | 150 | 151 | if opt.transfer or opt.prebias: # transfer learning edge (yolo) layers 152 | nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters']) # yolo layer size (i.e. 255) 153 | 154 | if opt.prebias: 155 | for p in optimizer.param_groups: 156 | # lower param count allows more aggressive training settings: i.e. SGD ~0.1 lr0, ~0.9 momentum 157 | p['lr'] *= 100 # lr gain 158 | if p.get('momentum') is not None: # for SGD but not Adam 159 | p['momentum'] *= 0.9 160 | 161 | for p in model.parameters(): 162 | if opt.prebias and p.numel() == nf: # train (yolo biases) 163 | p.requires_grad = True 164 | elif opt.transfer and p.shape[0] == nf: # train (yolo biases+weights) 165 | p.requires_grad = True 166 | else: # freeze layer 167 | p.requires_grad = False 168 | 169 | # Scheduler https://github.com/ultralytics/yolov3/issues/238 170 | # lf = lambda x: 1 - x / epochs # linear ramp to zero 171 | # lf = lambda x: 10 ** (hyp['lrf'] * x / epochs) # exp ramp 172 | # lf = lambda x: 1 - 10 ** (hyp['lrf'] * (1 - x / epochs)) # inverse exp ramp 173 | # scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) 174 | # scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=range(59, 70, 1), gamma=0.8) # gradual fall to 0.1*lr0 175 | if opt.sr: 176 | scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[round(opt.epochs * x) for x in [0.5, 0.6]], gamma=0.1) 177 | else: 178 | scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[round(opt.epochs * x) for x in [0.8, 0.9]], gamma=0.1) 179 | scheduler.last_epoch = start_epoch - 1 180 | 181 | # # Plot lr schedule 182 | # y = [] 183 | # for _ in range(epochs): 184 | # scheduler.step() 185 | # y.append(optimizer.param_groups[0]['lr']) 186 | # plt.plot(y, label='LambdaLR') 187 | # plt.xlabel('epoch') 188 | # plt.ylabel('LR') 189 | # plt.tight_layout() 190 | # plt.savefig('LR.png', dpi=300) 191 | 192 | # Mixed precision training https://github.com/NVIDIA/apex 193 | if mixed_precision: 194 | model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0) 195 | 196 | # Initialize distributed training 197 | if torch.cuda.device_count() > 1: 198 | dist.init_process_group(backend='nccl', # 'distributed backend' 199 | init_method='tcp://127.0.0.1:9999', # distributed training init method 200 | world_size=1, # number of nodes for distributed training 201 | rank=0) # distributed training node rank 202 | model = torch.nn.parallel.DistributedDataParallel(model) 203 | model.yolo_layers = model.module.yolo_layers # move yolo layer indices to top level 204 | 205 | # Dataset 206 | dataset = LoadImagesAndLabels(train_path, 207 | img_size, 208 | batch_size, 209 | augment=True, 210 | hyp=hyp, # augmentation hyperparameters 211 | rect=opt.rect, # rectangular training 212 | image_weights=opt.img_weights, 213 | cache_labels=True if epochs > 10 else False, 214 | cache_images=False if opt.prebias else opt.cache_images) 215 | 216 | # Dataloader 217 | dataloader = torch.utils.data.DataLoader(dataset, 218 | batch_size=batch_size, 219 | num_workers=min([os.cpu_count(), batch_size, 16]), 220 | shuffle=not opt.rect, # Shuffle=True unless rectangular training is used 221 | pin_memory=True, 222 | collate_fn=dataset.collate_fn) 223 | 224 | # Start training 225 | model.nc = nc # attach number of classes to model 226 | model.arc = opt.arc # attach yolo architecture 227 | model.hyp = hyp # attach hyperparameters to model 228 | # model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights 229 | torch_utils.model_info(model, report='summary') # 'full' or 'summary' 230 | nb = len(dataloader) 231 | maps = np.zeros(nc) # mAP per class 232 | results = (0, 0, 0, 0, 0, 0, 0) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification' 233 | t0 = time.time() 234 | print('Starting %s for %g epochs...' % ('prebias' if opt.prebias else 'training', epochs)) 235 | for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------ 236 | model.train() 237 | print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size')) 238 | 239 | # Freeze backbone at epoch 0, unfreeze at epoch 1 (optional) 240 | freeze_backbone = False 241 | if freeze_backbone and epoch < 2: 242 | for name, p in model.named_parameters(): 243 | if int(name.split('.')[1]) < cutoff: # if layer < 75 244 | p.requires_grad = False if epoch == 0 else True 245 | 246 | # Update image weights (optional) 247 | if dataset.image_weights: 248 | w = model.class_weights.cpu().numpy() * (1 - maps) ** 2 # class weights 249 | image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w) 250 | dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n) # rand weighted idx 251 | 252 | mloss = torch.zeros(4).to(device) # mean losses 253 | pbar = tqdm(enumerate(dataloader), total=nb) # progress bar 254 | sr_flag = get_sr_flag(epoch, opt.sr) 255 | for i, (imgs, targets, paths, _) in pbar: # batch ------------------------------------------------------------- 256 | ni = i + nb * epoch # number integrated batches (since train start) 257 | imgs = imgs.to(device) 258 | targets = targets.to(device) 259 | 260 | # Multi-Scale training 261 | if multi_scale: 262 | if ni / accumulate % 10 == 0: #  adjust (67% - 150%) every 10 batches 263 | img_size = random.randrange(img_sz_min, img_sz_max + 1) * 32 264 | sf = img_size / max(imgs.shape[2:]) # scale factor 265 | if sf != 1: 266 | ns = [math.ceil(x * sf / 32.) * 32 for x in imgs.shape[2:]] # new shape (stretched to 32-multiple) 267 | imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) 268 | 269 | # Plot images with bounding boxes 270 | if ni == 0: 271 | fname = 'train_batch%g.jpg' % i 272 | plot_images(imgs=imgs, targets=targets, paths=paths, fname=fname) 273 | if tb_writer: 274 | tb_writer.add_image(fname, cv2.imread(fname)[:, :, ::-1], dataformats='HWC') 275 | 276 | # Hyperparameter burn-in 277 | # n_burn = nb - 1 # min(nb // 5 + 1, 1000) # number of burn-in batches 278 | # if ni <= n_burn: 279 | # for m in model.named_modules(): 280 | # if m[0].endswith('BatchNorm2d'): 281 | # m[1].momentum = 1 - i / n_burn * 0.99 # BatchNorm2d momentum falls from 1 - 0.01 282 | # g = (i / n_burn) ** 4 # gain rises from 0 - 1 283 | # for x in optimizer.param_groups: 284 | # x['lr'] = hyp['lr0'] * g 285 | # x['weight_decay'] = hyp['weight_decay'] * g 286 | 287 | # Run model 288 | pred = model(imgs) 289 | 290 | # Compute loss 291 | loss, loss_items = compute_loss(pred, targets, model) 292 | if not torch.isfinite(loss): 293 | print('WARNING: non-finite loss, ending training ', loss_items) 294 | return results 295 | 296 | # Scale loss by nominal batch_size of 64 297 | loss *= batch_size / 64 298 | 299 | # Compute gradient 300 | if mixed_precision: 301 | with amp.scale_loss(loss, optimizer) as scaled_loss: 302 | scaled_loss.backward() 303 | else: 304 | loss.backward() 305 | 306 | BNOptimizer.updateBN(sr_flag, model.module_list, opt.s, prune_idx) 307 | 308 | # Accumulate gradient for x batches before optimizing 309 | if ni % accumulate == 0: 310 | optimizer.step() 311 | optimizer.zero_grad() 312 | 313 | # Print batch results 314 | mloss = (mloss * i + loss_items) / (i + 1) # update mean losses 315 | mem = torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available() else 0 # (GB) 316 | s = ('%10s' * 2 + '%10.3g' * 6) % ( 317 | '%g/%g' % (epoch, epochs - 1), '%.3gG' % mem, *mloss, len(targets), img_size) 318 | pbar.set_description(s) 319 | 320 | # end batch ------------------------------------------------------------------------------------------------ 321 | 322 | # Update scheduler 323 | scheduler.step() 324 | print('learning rate:',optimizer.param_groups[0]['lr']) 325 | 326 | 327 | # Process epoch results 328 | final_epoch = epoch + 1 == epochs 329 | if opt.prebias: 330 | print_model_biases(model) 331 | else: 332 | # Calculate mAP (always test final epoch, skip first 10 if opt.nosave) 333 | if not (opt.notest or (opt.nosave and epoch < 10)) or final_epoch: 334 | with torch.no_grad(): 335 | results, maps = test.test(cfg, 336 | data, 337 | batch_size=batch_size, 338 | img_size=opt.img_size, 339 | model=model, 340 | conf_thres=0.001 if final_epoch and epoch > 0 else 0.1, # 0.1 for speed 341 | save_json=final_epoch and epoch > 0 and 'coco.data' in data) 342 | 343 | # Write epoch results 344 | with open(results_file, 'a') as f: 345 | f.write(s + '%10.3g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls) 346 | 347 | # Write Tensorboard results 348 | if tb_writer: 349 | x = list(mloss) + list(results) 350 | titles = ['GIoU', 'Objectness', 'Classification', 'Train loss', 351 | 'Precision', 'Recall', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'] 352 | for xi, title in zip(x, titles): 353 | tb_writer.add_scalar(title, xi, epoch) 354 | bn_weights = gather_bn_weights(model.module_list, prune_idx) 355 | tb_writer.add_histogram('bn_weights/hist', bn_weights.numpy(), epoch, bins='doane') 356 | 357 | # Update best mAP 358 | fitness = results[2] # mAP 359 | if fitness > best_fitness: 360 | best_fitness = fitness 361 | 362 | # Save training results 363 | save = (not opt.nosave) or (final_epoch and not opt.evolve) or opt.prebias 364 | if save: 365 | with open(results_file, 'r') as f: 366 | # Create checkpoint 367 | chkpt = {'epoch': epoch, 368 | 'best_fitness': best_fitness, 369 | 'training_results': f.read(), 370 | 'model': model.module.state_dict() if type( 371 | model) is nn.parallel.DistributedDataParallel else model.state_dict(), 372 | 'optimizer': None if final_epoch else optimizer.state_dict()} 373 | 374 | # Save last checkpoint 375 | torch.save(chkpt, last) 376 | if opt.bucket and not opt.prebias: 377 | os.system('gsutil cp %s gs://%s' % (last, opt.bucket)) # upload to bucket 378 | 379 | # Save best checkpoint 380 | if best_fitness == fitness: 381 | torch.save(chkpt, best) 382 | 383 | # Save backup every 10 epochs (optional) 384 | if epoch > 0 and epoch % 5 == 0: 385 | torch.save(chkpt, wdir + 'backup%g.pt' % epoch) 386 | 387 | # Delete checkpoint 388 | del chkpt 389 | 390 | 391 | # end epoch ---------------------------------------------------------------------------------------------------- 392 | 393 | # end training 394 | if len(opt.name): 395 | os.rename('results.txt', 'results_%s.txt' % opt.name) 396 | plot_results() # save as results.png 397 | print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) 398 | dist.destroy_process_group() if torch.cuda.device_count() > 1 else None 399 | torch.cuda.empty_cache() 400 | return results 401 | 402 | 403 | def prebias(): 404 | # trains output bias layers for 1 epoch and creates new backbone 405 | if opt.prebias: 406 | train() # transfer-learn yolo biases for 1 epoch 407 | create_backbone(last) # saved results as backbone.pt 408 | opt.weights = wdir + 'backbone.pt' # assign backbone 409 | opt.prebias = False # disable prebias 410 | 411 | 412 | if __name__ == '__main__': 413 | parser = argparse.ArgumentParser() 414 | parser.add_argument('--epochs', type=int, default=273) # 500200 batches at bs 16, 117263 images = 273 epochs 415 | parser.add_argument('--batch-size', type=int, default=32) # effective bs = batch_size * accumulate = 16 * 4 = 64 416 | parser.add_argument('--accumulate', type=int, default=2, help='batches to accumulate before optimizing') 417 | parser.add_argument('--cfg', type=str, default='cfg/yolov3-spp.cfg', help='cfg file path') 418 | parser.add_argument('--data', type=str, default='data/coco.data', help='*.data file path') 419 | parser.add_argument('--multi-scale', action='store_true', help='adjust (67% - 150%) img_size every 10 batches') 420 | parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)') 421 | parser.add_argument('--rect', action='store_true', help='rectangular training') 422 | parser.add_argument('--resume', action='store_true', help='resume training from last.pt') 423 | parser.add_argument('--transfer', action='store_true', help='transfer learning') 424 | parser.add_argument('--nosave', action='store_true', help='only save final checkpoint') 425 | parser.add_argument('--notest', action='store_true', help='only test final epoch') 426 | parser.add_argument('--evolve', action='store_true', help='evolve hyperparameters') 427 | parser.add_argument('--bucket', type=str, default='', help='gsutil bucket') 428 | parser.add_argument('--img-weights', action='store_true', help='select training images by weight') 429 | parser.add_argument('--cache-images', action='store_true', help='cache images for faster training') 430 | parser.add_argument('--weights', type=str, default='', help='initial weights') # i.e. weights/darknet.53.conv.74 431 | parser.add_argument('--arc', type=str, default='defaultpw', help='yolo architecture') # defaultpw, uCE, uBCE 432 | parser.add_argument('--prebias', action='store_true', help='transfer-learn yolo biases prior to training') 433 | parser.add_argument('--name', default='', help='renames results.txt to results_name.txt if supplied') 434 | parser.add_argument('--device', default='', help='device id (i.e. 0 or 0,1) or cpu') 435 | parser.add_argument('--adam', action='store_true', help='use adam optimizer') 436 | parser.add_argument('--var', type=float, help='debug variable') 437 | parser.add_argument('--sparsity-regularization', '-sr', dest='sr', action='store_true', 438 | help='train with channel sparsity regularization') 439 | parser.add_argument('--s', type=float, default=0.001, help='scale sparse rate') 440 | parser.add_argument('--prune', type=int, default=0, help='0:nomal prune 1:shortcut prune ') 441 | opt = parser.parse_args() 442 | opt.weights = last if opt.resume else opt.weights 443 | print(opt) 444 | device = torch_utils.select_device(opt.device, apex=mixed_precision) 445 | 446 | tb_writer = None 447 | if not opt.evolve: # Train normally 448 | # try: 449 | # Start Tensorboard with "tensorboard --logdir=runs", view at http://localhost:6006/ 450 | from torch.utils.tensorboard import SummaryWriter 451 | 452 | tb_writer = SummaryWriter() 453 | # except: 454 | # pass 455 | 456 | prebias() # optional 457 | train() # train normally 458 | 459 | else: # Evolve hyperparameters (optional) 460 | opt.notest = True # only test final epoch 461 | opt.nosave = True # only save final checkpoint 462 | if opt.bucket: 463 | os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket) # download evolve.txt if exists 464 | 465 | for _ in range(1): # generations to evolve 466 | if os.path.exists('evolve.txt'): # if evolve.txt exists: select best hyps and mutate 467 | # Select parent(s) 468 | x = np.loadtxt('evolve.txt', ndmin=2) 469 | parent = 'weighted' # parent selection method: 'single' or 'weighted' 470 | if parent == 'single' or len(x) == 1: 471 | x = x[fitness(x).argmax()] 472 | elif parent == 'weighted': # weighted combination 473 | n = min(10, x.shape[0]) # number to merge 474 | x = x[np.argsort(-fitness(x))][:n] # top n mutations 475 | w = fitness(x) - fitness(x).min() # weights 476 | x = (x[:n] * w.reshape(n, 1)).sum(0) / w.sum() # new parent 477 | for i, k in enumerate(hyp.keys()): 478 | hyp[k] = x[i + 7] 479 | 480 | # Mutate 481 | np.random.seed(int(time.time())) 482 | s = [.2, .2, .2, .2, .2, .2, .2, .0, .02, .2, .2, .2, .2, .2, .2, .2, .2] # sigmas 483 | for i, k in enumerate(hyp.keys()): 484 | x = (np.random.randn(1) * s[i] + 1) ** 2.0 # plt.hist(x.ravel(), 300) 485 | hyp[k] *= float(x) # vary by sigmas 486 | 487 | # Clip to limits 488 | keys = ['lr0', 'iou_t', 'momentum', 'weight_decay', 'hsv_s', 'hsv_v', 'translate', 'scale', 'fl_gamma'] 489 | limits = [(1e-5, 1e-2), (0.00, 0.70), (0.60, 0.98), (0, 0.001), (0, .9), (0, .9), (0, .9), (0, .9), (0, 3)] 490 | for k, v in zip(keys, limits): 491 | hyp[k] = np.clip(hyp[k], v[0], v[1]) 492 | 493 | # Train mutation 494 | prebias() 495 | results = train() 496 | 497 | # Write mutation results 498 | print_mutation(hyp, results, opt.bucket) 499 | 500 | # Plot results 501 | # plot_evolution_results(hyp) 502 | -------------------------------------------------------------------------------- /utils/adabound.py: -------------------------------------------------------------------------------- 1 | import math 2 | 3 | import torch 4 | from torch.optim import Optimizer 5 | 6 | 7 | class AdaBound(Optimizer): 8 | """Implements AdaBound algorithm. 9 | It has been proposed in `Adaptive Gradient Methods with Dynamic Bound of Learning Rate`_. 10 | Arguments: 11 | params (iterable): iterable of parameters to optimize or dicts defining 12 | parameter groups 13 | lr (float, optional): Adam learning rate (default: 1e-3) 14 | betas (Tuple[float, float], optional): coefficients used for computing 15 | running averages of gradient and its square (default: (0.9, 0.999)) 16 | final_lr (float, optional): final (SGD) learning rate (default: 0.1) 17 | gamma (float, optional): convergence speed of the bound functions (default: 1e-3) 18 | eps (float, optional): term added to the denominator to improve 19 | numerical stability (default: 1e-8) 20 | weight_decay (float, optional): weight decay (L2 penalty) (default: 0) 21 | amsbound (boolean, optional): whether to use the AMSBound variant of this algorithm 22 | .. Adaptive Gradient Methods with Dynamic Bound of Learning Rate: 23 | https://openreview.net/forum?id=Bkg3g2R9FX 24 | """ 25 | 26 | def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), final_lr=0.1, gamma=1e-3, 27 | eps=1e-8, weight_decay=0, amsbound=False): 28 | if not 0.0 <= lr: 29 | raise ValueError("Invalid learning rate: {}".format(lr)) 30 | if not 0.0 <= eps: 31 | raise ValueError("Invalid epsilon value: {}".format(eps)) 32 | if not 0.0 <= betas[0] < 1.0: 33 | raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) 34 | if not 0.0 <= betas[1] < 1.0: 35 | raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) 36 | if not 0.0 <= final_lr: 37 | raise ValueError("Invalid final learning rate: {}".format(final_lr)) 38 | if not 0.0 <= gamma < 1.0: 39 | raise ValueError("Invalid gamma parameter: {}".format(gamma)) 40 | defaults = dict(lr=lr, betas=betas, final_lr=final_lr, gamma=gamma, eps=eps, 41 | weight_decay=weight_decay, amsbound=amsbound) 42 | super(AdaBound, self).__init__(params, defaults) 43 | 44 | self.base_lrs = list(map(lambda group: group['lr'], self.param_groups)) 45 | 46 | def __setstate__(self, state): 47 | super(AdaBound, self).__setstate__(state) 48 | for group in self.param_groups: 49 | group.setdefault('amsbound', False) 50 | 51 | def step(self, closure=None): 52 | """Performs a single optimization step. 53 | Arguments: 54 | closure (callable, optional): A closure that reevaluates the model 55 | and returns the loss. 56 | """ 57 | loss = None 58 | if closure is not None: 59 | loss = closure() 60 | 61 | for group, base_lr in zip(self.param_groups, self.base_lrs): 62 | for p in group['params']: 63 | if p.grad is None: 64 | continue 65 | grad = p.grad.data 66 | if grad.is_sparse: 67 | raise RuntimeError( 68 | 'Adam does not support sparse gradients, please consider SparseAdam instead') 69 | amsbound = group['amsbound'] 70 | 71 | state = self.state[p] 72 | 73 | # State initialization 74 | if len(state) == 0: 75 | state['step'] = 0 76 | # Exponential moving average of gradient values 77 | state['exp_avg'] = torch.zeros_like(p.data) 78 | # Exponential moving average of squared gradient values 79 | state['exp_avg_sq'] = torch.zeros_like(p.data) 80 | if amsbound: 81 | # Maintains max of all exp. moving avg. of sq. grad. values 82 | state['max_exp_avg_sq'] = torch.zeros_like(p.data) 83 | 84 | exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq'] 85 | if amsbound: 86 | max_exp_avg_sq = state['max_exp_avg_sq'] 87 | beta1, beta2 = group['betas'] 88 | 89 | state['step'] += 1 90 | 91 | if group['weight_decay'] != 0: 92 | grad = grad.add(group['weight_decay'], p.data) 93 | 94 | # Decay the first and second moment running average coefficient 95 | exp_avg.mul_(beta1).add_(1 - beta1, grad) 96 | exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad) 97 | if amsbound: 98 | # Maintains the maximum of all 2nd moment running avg. till now 99 | torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq) 100 | # Use the max. for normalizing running avg. of gradient 101 | denom = max_exp_avg_sq.sqrt().add_(group['eps']) 102 | else: 103 | denom = exp_avg_sq.sqrt().add_(group['eps']) 104 | 105 | bias_correction1 = 1 - beta1 ** state['step'] 106 | bias_correction2 = 1 - beta2 ** state['step'] 107 | step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1 108 | 109 | # Applies bounds on actual learning rate 110 | # lr_scheduler cannot affect final_lr, this is a workaround to apply lr decay 111 | final_lr = group['final_lr'] * group['lr'] / base_lr 112 | lower_bound = final_lr * (1 - 1 / (group['gamma'] * state['step'] + 1)) 113 | upper_bound = final_lr * (1 + 1 / (group['gamma'] * state['step'])) 114 | step_size = torch.full_like(denom, step_size) 115 | step_size.div_(denom).clamp_(lower_bound, upper_bound).mul_(exp_avg) 116 | 117 | p.data.add_(-step_size) 118 | 119 | return loss 120 | 121 | 122 | class AdaBoundW(Optimizer): 123 | """Implements AdaBound algorithm with Decoupled Weight Decay (arxiv.org/abs/1711.05101) 124 | It has been proposed in `Adaptive Gradient Methods with Dynamic Bound of Learning Rate`_. 125 | Arguments: 126 | params (iterable): iterable of parameters to optimize or dicts defining 127 | parameter groups 128 | lr (float, optional): Adam learning rate (default: 1e-3) 129 | betas (Tuple[float, float], optional): coefficients used for computing 130 | running averages of gradient and its square (default: (0.9, 0.999)) 131 | final_lr (float, optional): final (SGD) learning rate (default: 0.1) 132 | gamma (float, optional): convergence speed of the bound functions (default: 1e-3) 133 | eps (float, optional): term added to the denominator to improve 134 | numerical stability (default: 1e-8) 135 | weight_decay (float, optional): weight decay (L2 penalty) (default: 0) 136 | amsbound (boolean, optional): whether to use the AMSBound variant of this algorithm 137 | .. Adaptive Gradient Methods with Dynamic Bound of Learning Rate: 138 | https://openreview.net/forum?id=Bkg3g2R9FX 139 | """ 140 | 141 | def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), final_lr=0.1, gamma=1e-3, 142 | eps=1e-8, weight_decay=0, amsbound=False): 143 | if not 0.0 <= lr: 144 | raise ValueError("Invalid learning rate: {}".format(lr)) 145 | if not 0.0 <= eps: 146 | raise ValueError("Invalid epsilon value: {}".format(eps)) 147 | if not 0.0 <= betas[0] < 1.0: 148 | raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) 149 | if not 0.0 <= betas[1] < 1.0: 150 | raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) 151 | if not 0.0 <= final_lr: 152 | raise ValueError("Invalid final learning rate: {}".format(final_lr)) 153 | if not 0.0 <= gamma < 1.0: 154 | raise ValueError("Invalid gamma parameter: {}".format(gamma)) 155 | defaults = dict(lr=lr, betas=betas, final_lr=final_lr, gamma=gamma, eps=eps, 156 | weight_decay=weight_decay, amsbound=amsbound) 157 | super(AdaBoundW, self).__init__(params, defaults) 158 | 159 | self.base_lrs = list(map(lambda group: group['lr'], self.param_groups)) 160 | 161 | def __setstate__(self, state): 162 | super(AdaBoundW, self).__setstate__(state) 163 | for group in self.param_groups: 164 | group.setdefault('amsbound', False) 165 | 166 | def step(self, closure=None): 167 | """Performs a single optimization step. 168 | Arguments: 169 | closure (callable, optional): A closure that reevaluates the model 170 | and returns the loss. 171 | """ 172 | loss = None 173 | if closure is not None: 174 | loss = closure() 175 | 176 | for group, base_lr in zip(self.param_groups, self.base_lrs): 177 | for p in group['params']: 178 | if p.grad is None: 179 | continue 180 | grad = p.grad.data 181 | if grad.is_sparse: 182 | raise RuntimeError( 183 | 'Adam does not support sparse gradients, please consider SparseAdam instead') 184 | amsbound = group['amsbound'] 185 | 186 | state = self.state[p] 187 | 188 | # State initialization 189 | if len(state) == 0: 190 | state['step'] = 0 191 | # Exponential moving average of gradient values 192 | state['exp_avg'] = torch.zeros_like(p.data) 193 | # Exponential moving average of squared gradient values 194 | state['exp_avg_sq'] = torch.zeros_like(p.data) 195 | if amsbound: 196 | # Maintains max of all exp. moving avg. of sq. grad. values 197 | state['max_exp_avg_sq'] = torch.zeros_like(p.data) 198 | 199 | exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq'] 200 | if amsbound: 201 | max_exp_avg_sq = state['max_exp_avg_sq'] 202 | beta1, beta2 = group['betas'] 203 | 204 | state['step'] += 1 205 | 206 | # Decay the first and second moment running average coefficient 207 | exp_avg.mul_(beta1).add_(1 - beta1, grad) 208 | exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad) 209 | if amsbound: 210 | # Maintains the maximum of all 2nd moment running avg. till now 211 | torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq) 212 | # Use the max. for normalizing running avg. of gradient 213 | denom = max_exp_avg_sq.sqrt().add_(group['eps']) 214 | else: 215 | denom = exp_avg_sq.sqrt().add_(group['eps']) 216 | 217 | bias_correction1 = 1 - beta1 ** state['step'] 218 | bias_correction2 = 1 - beta2 ** state['step'] 219 | step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1 220 | 221 | # Applies bounds on actual learning rate 222 | # lr_scheduler cannot affect final_lr, this is a workaround to apply lr decay 223 | final_lr = group['final_lr'] * group['lr'] / base_lr 224 | lower_bound = final_lr * (1 - 1 / (group['gamma'] * state['step'] + 1)) 225 | upper_bound = final_lr * (1 + 1 / (group['gamma'] * state['step'])) 226 | step_size = torch.full_like(denom, step_size) 227 | step_size.div_(denom).clamp_(lower_bound, upper_bound).mul_(exp_avg) 228 | 229 | if group['weight_decay'] != 0: 230 | decayed_weights = torch.mul(p.data, group['weight_decay']) 231 | p.data.add_(-step_size) 232 | p.data.sub_(decayed_weights) 233 | else: 234 | p.data.add_(-step_size) 235 | 236 | return loss 237 | -------------------------------------------------------------------------------- /utils/google_utils.py: -------------------------------------------------------------------------------- 1 | # This file contains google utils: https://cloud.google.com/storage/docs/reference/libraries 2 | # pip install --upgrade google-cloud-storage 3 | 4 | import os 5 | import time 6 | 7 | 8 | # from google.cloud import storage 9 | 10 | 11 | def gdrive_download(id='1HaXkef9z6y5l4vUnCYgdmEAj61c6bfWO', name='coco.zip'): 12 | # https://gist.github.com/tanaikech/f0f2d122e05bf5f971611258c22c110f 13 | # Downloads a file from Google Drive, accepting presented query 14 | # from utils.google_utils import *; gdrive_download() 15 | t = time.time() 16 | 17 | print('Downloading https://drive.google.com/uc?export=download&id=%s as %s... ' % (id, name), end='') 18 | if os.path.exists(name): # remove existing 19 | os.remove(name) 20 | 21 | # Attempt large file download 22 | s = ["curl -c ./cookie -s -L \"https://drive.google.com/uc?export=download&id=%s\" > /dev/null" % id, 23 | "curl -Lb ./cookie -s \"https://drive.google.com/uc?export=download&confirm=`awk '/download/ {print $NF}' ./cookie`&id=%s\" -o %s" % ( 24 | id, name), 25 | 'rm ./cookie'] 26 | [os.system(x) for x in s] # run commands 27 | 28 | # Attempt small file download 29 | if not os.path.exists(name): # file size < 40MB 30 | s = 'curl -f -L -o %s https://drive.google.com/uc?export=download&id=%s' % (name, id) 31 | os.system(s) 32 | 33 | # Unzip if archive 34 | if name.endswith('.zip'): 35 | print('unzipping... ', end='') 36 | os.system('unzip -q %s' % name) # unzip 37 | os.remove(name) # remove zip to free space 38 | 39 | print('Done (%.1fs)' % (time.time() - t)) 40 | 41 | 42 | def upload_blob(bucket_name, source_file_name, destination_blob_name): 43 | # Uploads a file to a bucket 44 | # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python 45 | 46 | storage_client = storage.Client() 47 | bucket = storage_client.get_bucket(bucket_name) 48 | blob = bucket.blob(destination_blob_name) 49 | 50 | blob.upload_from_filename(source_file_name) 51 | 52 | print('File {} uploaded to {}.'.format( 53 | source_file_name, 54 | destination_blob_name)) 55 | 56 | 57 | def download_blob(bucket_name, source_blob_name, destination_file_name): 58 | # Uploads a blob from a bucket 59 | storage_client = storage.Client() 60 | bucket = storage_client.get_bucket(bucket_name) 61 | blob = bucket.blob(source_blob_name) 62 | 63 | blob.download_to_filename(destination_file_name) 64 | 65 | print('Blob {} downloaded to {}.'.format( 66 | source_blob_name, 67 | destination_file_name)) 68 | -------------------------------------------------------------------------------- /utils/parse_config.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | 4 | def parse_model_cfg(path): 5 | # Parses the yolo-v3 layer configuration file and returns module definitions 6 | file = open(path, 'r') 7 | lines = file.read().split('\n') 8 | lines = [x for x in lines if x and not x.startswith('#')] 9 | lines = [x.rstrip().lstrip() for x in lines] # get rid of fringe whitespaces 10 | mdefs = [] # module definitions 11 | for line in lines: 12 | if line.startswith('['): # This marks the start of a new block 13 | mdefs.append({}) 14 | mdefs[-1]['type'] = line[1:-1].rstrip() 15 | if mdefs[-1]['type'] == 'convolutional': 16 | mdefs[-1]['batch_normalize'] = 0 # pre-populate with zeros (may be overwritten later) 17 | else: 18 | key, val = line.split("=") 19 | key = key.rstrip() 20 | 21 | if 'anchors' in key: 22 | mdefs[-1][key] = np.array([float(x) for x in val.split(',')]).reshape((-1, 2)) # np anchors 23 | else: 24 | mdefs[-1][key] = val.strip() 25 | 26 | return mdefs 27 | 28 | 29 | def parse_data_cfg(path): 30 | # Parses the data configuration file 31 | options = dict() 32 | with open(path, 'r') as fp: 33 | lines = fp.readlines() 34 | 35 | for line in lines: 36 | line = line.strip() 37 | if line == '' or line.startswith('#'): 38 | continue 39 | key, val = line.split('=') 40 | options[key.strip()] = val.strip() 41 | 42 | return options 43 | 44 | -------------------------------------------------------------------------------- /utils/prune_utils.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from terminaltables import AsciiTable 3 | from copy import deepcopy 4 | import numpy as np 5 | import torch.nn.functional as F 6 | 7 | 8 | def get_sr_flag(epoch, sr): 9 | # return epoch >= 5 and sr 10 | return sr 11 | 12 | 13 | def parse_module_defs(module_defs): 14 | 15 | CBL_idx = [] 16 | Conv_idx = [] 17 | for i, module_def in enumerate(module_defs): 18 | if module_def['type'] == 'convolutional': 19 | if module_def['batch_normalize'] == '1': 20 | CBL_idx.append(i) 21 | else: 22 | Conv_idx.append(i) 23 | 24 | ignore_idx = set() 25 | for i, module_def in enumerate(module_defs): 26 | if module_def['type'] == 'shortcut': 27 | ignore_idx.add(i-1) 28 | identity_idx = (i + int(module_def['from'])) 29 | if module_defs[identity_idx]['type'] == 'convolutional': 30 | ignore_idx.add(identity_idx) 31 | elif module_defs[identity_idx]['type'] == 'shortcut': 32 | ignore_idx.add(identity_idx - 1) 33 | 34 | ignore_idx.add(84) 35 | ignore_idx.add(96) 36 | 37 | prune_idx = [idx for idx in CBL_idx if idx not in ignore_idx] 38 | 39 | return CBL_idx, Conv_idx, prune_idx 40 | 41 | 42 | def parse_module_defs2(module_defs): 43 | 44 | CBL_idx = [] 45 | Conv_idx = [] 46 | shortcut_idx=dict() 47 | shortcut_all=set() 48 | for i, module_def in enumerate(module_defs): 49 | if module_def['type'] == 'convolutional': 50 | if module_def['batch_normalize'] == '1': 51 | CBL_idx.append(i) 52 | else: 53 | Conv_idx.append(i) 54 | 55 | ignore_idx = set() 56 | for i, module_def in enumerate(module_defs): 57 | if module_def['type'] == 'shortcut': 58 | identity_idx = (i + int(module_def['from'])) 59 | if module_defs[identity_idx]['type'] == 'convolutional': 60 | 61 | #ignore_idx.add(identity_idx) 62 | shortcut_idx[i-1]=identity_idx 63 | shortcut_all.add(identity_idx) 64 | elif module_defs[identity_idx]['type'] == 'shortcut': 65 | 66 | #ignore_idx.add(identity_idx - 1) 67 | shortcut_idx[i-1]=identity_idx-1 68 | shortcut_all.add(identity_idx-1) 69 | shortcut_all.add(i-1) 70 | #上采样层前的卷积层不裁剪 71 | ignore_idx.add(84) 72 | ignore_idx.add(96) 73 | 74 | prune_idx = [idx for idx in CBL_idx if idx not in ignore_idx] 75 | 76 | return CBL_idx, Conv_idx, prune_idx,shortcut_idx,shortcut_all 77 | 78 | 79 | 80 | def gather_bn_weights(module_list, prune_idx): 81 | 82 | size_list = [module_list[idx][1].weight.data.shape[0] for idx in prune_idx] 83 | 84 | bn_weights = torch.zeros(sum(size_list)) 85 | index = 0 86 | for idx, size in zip(prune_idx, size_list): 87 | bn_weights[index:(index + size)] = module_list[idx][1].weight.data.abs().clone() 88 | index += size 89 | 90 | return bn_weights 91 | 92 | 93 | def write_cfg(cfg_file, module_defs): 94 | 95 | with open(cfg_file, 'w') as f: 96 | for module_def in module_defs: 97 | f.write(f"[{module_def['type']}]\n") 98 | for key, value in module_def.items(): 99 | if key == 'batch_normalize' and value == 0: 100 | continue 101 | 102 | if key != 'type': 103 | if key == 'anchors': 104 | value = ', '.join(','.join(str(int(i)) for i in j) for j in value) 105 | f.write(f"{key}={value}\n") 106 | f.write("\n") 107 | return cfg_file 108 | 109 | 110 | class BNOptimizer(): 111 | 112 | @staticmethod 113 | def updateBN(sr_flag, module_list, s, prune_idx): 114 | if sr_flag: 115 | for idx in prune_idx: 116 | # Squential(Conv, BN, Lrelu) 117 | bn_module = module_list[idx][1] 118 | bn_module.weight.grad.data.add_(s * torch.sign(bn_module.weight.data)) # L1 119 | 120 | 121 | def obtain_quantiles(bn_weights, num_quantile=5): 122 | 123 | sorted_bn_weights, i = torch.sort(bn_weights) 124 | total = sorted_bn_weights.shape[0] 125 | quantiles = sorted_bn_weights.tolist()[-1::-total//num_quantile][::-1] 126 | print("\nBN weights quantile:") 127 | quantile_table = [ 128 | [f'{i}/{num_quantile}' for i in range(1, num_quantile+1)], 129 | ["%.3f" % quantile for quantile in quantiles] 130 | ] 131 | print(AsciiTable(quantile_table).table) 132 | 133 | return quantiles 134 | 135 | 136 | def get_input_mask(module_defs, idx, CBLidx2mask): 137 | 138 | if idx == 0: 139 | return np.ones(3) 140 | 141 | if module_defs[idx - 1]['type'] == 'convolutional': 142 | return CBLidx2mask[idx - 1] 143 | elif module_defs[idx - 1]['type'] == 'shortcut': 144 | return CBLidx2mask[idx - 2] 145 | elif module_defs[idx - 1]['type'] == 'route': 146 | route_in_idxs = [] 147 | for layer_i in module_defs[idx - 1]['layers'].split(","): 148 | if int(layer_i) < 0: 149 | route_in_idxs.append(idx - 1 + int(layer_i)) 150 | else: 151 | route_in_idxs.append(int(layer_i)) 152 | if len(route_in_idxs) == 1: 153 | return CBLidx2mask[route_in_idxs[0]] 154 | elif len(route_in_idxs) == 2: 155 | return np.concatenate([CBLidx2mask[in_idx - 1] for in_idx in route_in_idxs]) 156 | else: 157 | print("Something wrong with route module!") 158 | raise Exception 159 | 160 | 161 | def init_weights_from_loose_model(compact_model, loose_model, CBL_idx, Conv_idx, CBLidx2mask): 162 | 163 | for idx in CBL_idx: 164 | compact_CBL = compact_model.module_list[idx] 165 | loose_CBL = loose_model.module_list[idx] 166 | out_channel_idx = np.argwhere(CBLidx2mask[idx])[:, 0].tolist() 167 | 168 | compact_bn, loose_bn = compact_CBL[1], loose_CBL[1] 169 | compact_bn.weight.data = loose_bn.weight.data[out_channel_idx].clone() 170 | compact_bn.bias.data = loose_bn.bias.data[out_channel_idx].clone() 171 | compact_bn.running_mean.data = loose_bn.running_mean.data[out_channel_idx].clone() 172 | compact_bn.running_var.data = loose_bn.running_var.data[out_channel_idx].clone() 173 | 174 | input_mask = get_input_mask(loose_model.module_defs, idx, CBLidx2mask) 175 | in_channel_idx = np.argwhere(input_mask)[:, 0].tolist() 176 | compact_conv, loose_conv = compact_CBL[0], loose_CBL[0] 177 | tmp = loose_conv.weight.data[:, in_channel_idx, :, :].clone() 178 | compact_conv.weight.data = tmp[out_channel_idx, :, :, :].clone() 179 | 180 | for idx in Conv_idx: 181 | compact_conv = compact_model.module_list[idx][0] 182 | loose_conv = loose_model.module_list[idx][0] 183 | 184 | input_mask = get_input_mask(loose_model.module_defs, idx, CBLidx2mask) 185 | in_channel_idx = np.argwhere(input_mask)[:, 0].tolist() 186 | compact_conv.weight.data = loose_conv.weight.data[:, in_channel_idx, :, :].clone() 187 | compact_conv.bias.data = loose_conv.bias.data.clone() 188 | 189 | 190 | def prune_model_keep_size(model, prune_idx, CBL_idx, CBLidx2mask): 191 | 192 | pruned_model = deepcopy(model) 193 | for idx in prune_idx: 194 | mask = torch.from_numpy(CBLidx2mask[idx]).cuda() 195 | bn_module = pruned_model.module_list[idx][1] 196 | 197 | bn_module.weight.data.mul_(mask) 198 | 199 | activation = F.leaky_relu((1 - mask) * bn_module.bias.data, 0.1) 200 | 201 | # 两个上采样层前的卷积层 202 | next_idx_list = [idx + 1] 203 | if idx == 79: 204 | next_idx_list.append(84) 205 | elif idx == 91: 206 | next_idx_list.append(96) 207 | 208 | for next_idx in next_idx_list: 209 | next_conv = pruned_model.module_list[next_idx][0] 210 | conv_sum = next_conv.weight.data.sum(dim=(2, 3)) 211 | offset = conv_sum.matmul(activation.reshape(-1, 1)).reshape(-1) 212 | if next_idx in CBL_idx: 213 | next_bn = pruned_model.module_list[next_idx][1] 214 | next_bn.running_mean.data.sub_(offset) 215 | else: 216 | next_conv.bias.data.add_(offset) 217 | 218 | bn_module.bias.data.mul_(mask) 219 | 220 | return pruned_model 221 | 222 | 223 | def obtain_bn_mask(bn_module, thre): 224 | 225 | thre = thre.cuda() 226 | mask = bn_module.weight.data.abs().ge(thre).float() 227 | 228 | return mask 229 | 230 | 231 | 232 | def update_activation(i, pruned_model, activation, CBL_idx): 233 | next_idx = i + 1 234 | if pruned_model.module_defs[next_idx]['type'] == 'convolutional': 235 | next_conv = pruned_model.module_list[next_idx][0] 236 | conv_sum = next_conv.weight.data.sum(dim=(2, 3)) 237 | offset = conv_sum.matmul(activation.reshape(-1, 1)).reshape(-1) 238 | if next_idx in CBL_idx: 239 | next_bn = pruned_model.module_list[next_idx][1] 240 | next_bn.running_mean.data.sub_(offset) 241 | else: 242 | next_conv.bias.data.add_(offset) 243 | 244 | 245 | 246 | def prune_model_keep_size2(model, prune_idx, CBL_idx, CBLidx2mask): 247 | 248 | pruned_model = deepcopy(model) 249 | activations = [] 250 | for i, model_def in enumerate(model.module_defs): 251 | 252 | if model_def['type'] == 'convolutional': 253 | activation = None 254 | if i in prune_idx: 255 | mask = torch.from_numpy(CBLidx2mask[i]).cuda() 256 | bn_module = pruned_model.module_list[i][1] 257 | bn_module.weight.data.mul_(mask) 258 | activation = F.leaky_relu((1 - mask) * bn_module.bias.data, 0.1) 259 | update_activation(i, pruned_model, activation, CBL_idx) 260 | bn_module.bias.data.mul_(mask) 261 | activations.append(activation) 262 | 263 | if model_def['type'] == 'shortcut': 264 | actv1 = activations[i - 1] 265 | from_layer = int(model_def['from']) 266 | actv2 = activations[i + from_layer] 267 | activation = actv1 + actv2 268 | update_activation(i, pruned_model, activation, CBL_idx) 269 | activations.append(activation) 270 | 271 | 272 | 273 | if model_def['type'] == 'route': 274 | from_layers = [int(s) for s in model_def['layers'].split(',')] 275 | if len(from_layers) == 1: 276 | activation = activations[i + from_layers[0]] 277 | update_activation(i, pruned_model, activation, CBL_idx) 278 | else: 279 | actv1 = activations[i + from_layers[0]] 280 | actv2 = activations[from_layers[1]] 281 | activation = torch.cat((actv1, actv2)) 282 | update_activation(i, pruned_model, activation, CBL_idx) 283 | activations.append(activation) 284 | 285 | if model_def['type'] == 'upsample': 286 | activation = torch.zeros(int(model.module_defs[i - 1]['filters'])).cuda() 287 | activations.append(activation) 288 | 289 | if model_def['type'] == 'yolo': 290 | activations.append(None) 291 | 292 | return pruned_model 293 | -------------------------------------------------------------------------------- /utils/torch_utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | import torch 4 | 5 | 6 | def init_seeds(seed=0): 7 | torch.manual_seed(seed) 8 | torch.cuda.manual_seed(seed) 9 | torch.cuda.manual_seed_all(seed) 10 | 11 | # Remove randomness (may be slower on Tesla GPUs) # https://pytorch.org/docs/stable/notes/randomness.html 12 | if seed == 0: 13 | torch.backends.cudnn.deterministic = True 14 | torch.backends.cudnn.benchmark = False 15 | 16 | 17 | def select_device(device='', apex=False): 18 | # device = 'cpu' or '0' or '0,1,2,3' 19 | cpu_request = device.lower() == 'cpu' 20 | if device and not cpu_request: # if device requested other than 'cpu' 21 | os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable 22 | assert torch.cuda.is_available(), 'CUDA unavailable, invalid device %s requested' % device # check availablity 23 | 24 | cuda = False if cpu_request else torch.cuda.is_available() 25 | if cuda: 26 | c = 1024 ** 2 # bytes to MB 27 | ng = torch.cuda.device_count() 28 | x = [torch.cuda.get_device_properties(i) for i in range(ng)] 29 | cuda_str = 'Using CUDA ' + ('Apex ' if apex else '') # apex for mixed precision https://github.com/NVIDIA/apex 30 | for i in range(0, ng): 31 | if i == 1: 32 | cuda_str = ' ' * len(cuda_str) 33 | print("%sdevice%g _CudaDeviceProperties(name='%s', total_memory=%dMB)" % 34 | (cuda_str, i, x[i].name, x[i].total_memory / c)) 35 | else: 36 | print('Using CPU') 37 | 38 | print('') # skip a line 39 | return torch.device('cuda:0' if cuda else 'cpu') 40 | 41 | 42 | def fuse_conv_and_bn(conv, bn): 43 | # https://tehnokv.com/posts/fusing-batchnorm-and-conv/ 44 | with torch.no_grad(): 45 | # init 46 | fusedconv = torch.nn.Conv2d(conv.in_channels, 47 | conv.out_channels, 48 | kernel_size=conv.kernel_size, 49 | stride=conv.stride, 50 | padding=conv.padding, 51 | bias=True) 52 | 53 | # prepare filters 54 | w_conv = conv.weight.clone().view(conv.out_channels, -1) 55 | w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) 56 | fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.size())) 57 | 58 | # prepare spatial bias 59 | if conv.bias is not None: 60 | b_conv = conv.bias 61 | else: 62 | b_conv = torch.zeros(conv.weight.size(0)) 63 | b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps)) 64 | fusedconv.bias.copy_(b_conv + b_bn) 65 | 66 | return fusedconv 67 | 68 | 69 | def model_info(model, report='summary'): 70 | # Plots a line-by-line description of a PyTorch model 71 | n_p = sum(x.numel() for x in model.parameters()) # number parameters 72 | n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients 73 | if report is 'full': 74 | print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma')) 75 | for i, (name, p) in enumerate(model.named_parameters()): 76 | name = name.replace('module_list.', '') 77 | print('%5g %40s %9s %12g %20s %10.3g %10.3g' % 78 | (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std())) 79 | print('Model Summary: %g layers, %g parameters, %g gradients' % (len(list(model.parameters())), n_p, n_g)) 80 | -------------------------------------------------------------------------------- /weights/download_yolov3_weights.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | # make '/weights' directory if it does not exist and cd into it 4 | mkdir -p weights && cd weights 5 | 6 | # copy darknet weight files, continue '-c' if partially downloaded 7 | wget -c https://pjreddie.com/media/files/yolov3.weights 8 | wget -c https://pjreddie.com/media/files/yolov3-tiny.weights 9 | wget -c https://pjreddie.com/media/files/yolov3-spp.weights 10 | 11 | # yolov3 pytorch weights 12 | # download from Google Drive: https://drive.google.com/drive/folders/1uxgUBemJVw9wZsdpboYbzUN4bcRhsuAI 13 | 14 | # darknet53 weights (first 75 layers only) 15 | wget -c https://pjreddie.com/media/files/darknet53.conv.74 16 | 17 | # yolov3-tiny weights from darknet (first 16 layers only) 18 | # ./darknet partial cfg/yolov3-tiny.cfg yolov3-tiny.weights yolov3-tiny.conv.15 15 19 | # mv yolov3-tiny.conv.15 ../ 20 | 21 | --------------------------------------------------------------------------------