├── .dockerignore ├── .gitignore ├── Dockerfile ├── README.md ├── READMEYolov.md ├── allbn.txt ├── data ├── GlobalWheat2020.yaml ├── VisDrone.yaml ├── argoverse_hd.yaml ├── coco.yaml ├── coco128.yaml ├── hyp.finetune.yaml ├── hyp.finetune_objects365.yaml ├── hyp.scratch.yaml ├── images │ ├── bus.jpg │ └── zidane.jpg ├── objects365.yaml ├── scripts │ ├── get_argoverse_hd.sh │ ├── get_coco.sh │ ├── get_coco128.sh │ └── get_voc.sh └── voc.yaml ├── detect.py ├── detect_prune.py ├── finetune_prune_conv.py ├── finetune_pruned.py ├── finetune_pruned2.py ├── getweight.py ├── hubconf.py ├── img ├── Screenshot from 2021-05-23 20-19-08.png ├── Screenshot from 2021-05-23 20-19-30.png ├── Screenshot from 2021-05-24 22-17-16.png ├── Screenshot from 2021-05-25 00-26-23.png ├── Screenshot from 2021-05-25 00-26-45.png ├── Screenshot from 2021-05-25 00-28-15.png ├── Screenshot from 2021-05-25 00-28-52.png ├── Screenshot from 2021-05-27 22-20-33.png ├── Screenshot from 2021-05-28 08-33-25.png ├── Screenshot from 2021-05-31 22-29-12.png ├── Screenshot from 2021-05-31 22-30-21.png ├── Screenshot from 2021-06-05 00-06-27.png └── Selection_007.png ├── map.txt ├── model_change.txt ├── modelparse.py ├── models ├── __init__.py ├── common.py ├── experimental.py ├── export.py ├── hub │ ├── anchors.yaml │ ├── yolov3-spp.yaml │ ├── yolov3-tiny.yaml │ ├── yolov3.yaml │ ├── yolov5-fpn.yaml │ ├── yolov5-p2.yaml │ ├── yolov5-p6.yaml │ ├── yolov5-p7.yaml │ ├── yolov5-panet.yaml │ ├── yolov5l6.yaml │ ├── yolov5m6.yaml │ ├── yolov5s-transformer.yaml │ ├── yolov5s6.yaml │ └── yolov5x6.yaml ├── modul.txt ├── pruned_common.py ├── yolo.py ├── yolov5l.yaml ├── yolov5m.yaml ├── yolov5s-opt.param ├── yolov5s.yaml ├── yolov5slite.yaml ├── yolov5sprune.yaml └── yolov5x.yaml ├── prune.py ├── prune2.py ├── prune_conv.py ├── prune_convbn.py ├── prune_utils.py ├── reprune.py ├── requirements.txt ├── showbn.py ├── test.py ├── testprune.py ├── train.py ├── train_prune_sparsity.py ├── train_sparsity.py ├── train_sparsity2.py ├── train_sparsity3.py ├── train_sparsity4.py ├── tutorial.ipynb ├── utils ├── __init__.py ├── activations.py ├── autoanchor.py ├── aws │ ├── __init__.py │ ├── mime.sh │ ├── resume.py │ └── userdata.sh ├── datasets.py ├── flask_rest_api │ ├── README.md │ ├── example_request.py │ └── restapi.py ├── general.py ├── google_app_engine │ ├── Dockerfile │ ├── additional_requirements.txt │ └── app.yaml ├── google_utils.py ├── loss.py ├── metrics.py ├── plots.py ├── torch_utils.py └── wandb_logging │ ├── __init__.py │ ├── log_dataset.py │ └── wandb_utils.py ├── weights └── download_weights.sh └── yolov5prune.md /.dockerignore: -------------------------------------------------------------------------------- 1 | # Repo-specific DockerIgnore ------------------------------------------------------------------------------------------- 2 | #.git 3 | .cache 4 | .idea 5 | runs 6 | output 7 | coco 8 | storage.googleapis.com 9 | 10 | data/samples/* 11 | **/results*.txt 12 | *.jpg 13 | 14 | # Neural Network weights ----------------------------------------------------------------------------------------------- 15 | **/*.weights 16 | **/*.pt 17 | **/*.pth 18 | **/*.onnx 19 | **/*.mlmodel 20 | **/*.torchscript 21 | 22 | 23 | # Below Copied From .gitignore ----------------------------------------------------------------------------------------- 24 | # Below Copied From .gitignore ----------------------------------------------------------------------------------------- 25 | 26 | 27 | # GitHub Python GitIgnore ---------------------------------------------------------------------------------------------- 28 | # Byte-compiled / optimized / DLL files 29 | __pycache__/ 30 | *.py[cod] 31 | *$py.class 32 | 33 | # C extensions 34 | *.so 35 | 36 | # Distribution / packaging 37 | .Python 38 | env/ 39 | build/ 40 | develop-eggs/ 41 | dist/ 42 | downloads/ 43 | eggs/ 44 | .eggs/ 45 | lib/ 46 | lib64/ 47 | parts/ 48 | sdist/ 49 | var/ 50 | wheels/ 51 | *.egg-info/ 52 | wandb/ 53 | .installed.cfg 54 | *.egg 55 | 56 | # PyInstaller 57 | # Usually these files are written by a python script from a template 58 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 59 | *.manifest 60 | *.spec 61 | 62 | # Installer logs 63 | pip-log.txt 64 | pip-delete-this-directory.txt 65 | 66 | # Unit test / coverage reports 67 | htmlcov/ 68 | .tox/ 69 | .coverage 70 | .coverage.* 71 | .cache 72 | nosetests.xml 73 | coverage.xml 74 | *.cover 75 | .hypothesis/ 76 | 77 | # Translations 78 | *.mo 79 | *.pot 80 | 81 | # Django stuff: 82 | *.log 83 | local_settings.py 84 | 85 | # Flask stuff: 86 | instance/ 87 | .webassets-cache 88 | 89 | # Scrapy stuff: 90 | .scrapy 91 | 92 | # Sphinx documentation 93 | docs/_build/ 94 | 95 | # PyBuilder 96 | target/ 97 | 98 | # Jupyter Notebook 99 | .ipynb_checkpoints 100 | 101 | # pyenv 102 | .python-version 103 | 104 | # celery beat schedule file 105 | celerybeat-schedule 106 | 107 | # SageMath parsed files 108 | *.sage.py 109 | 110 | # dotenv 111 | .env 112 | 113 | # virtualenv 114 | .venv* 115 | venv*/ 116 | ENV*/ 117 | 118 | # Spyder project settings 119 | .spyderproject 120 | .spyproject 121 | 122 | # Rope project settings 123 | .ropeproject 124 | 125 | # mkdocs documentation 126 | /site 127 | 128 | # mypy 129 | .mypy_cache/ 130 | 131 | 132 | # https://github.com/github/gitignore/blob/master/Global/macOS.gitignore ----------------------------------------------- 133 | 134 | # General 135 | .DS_Store 136 | .AppleDouble 137 | .LSOverride 138 | 139 | # Icon must end with two \r 140 | Icon 141 | Icon? 142 | 143 | # Thumbnails 144 | ._* 145 | 146 | # Files that might appear in the root of a volume 147 | .DocumentRevisions-V100 148 | .fseventsd 149 | .Spotlight-V100 150 | .TemporaryItems 151 | .Trashes 152 | .VolumeIcon.icns 153 | .com.apple.timemachine.donotpresent 154 | 155 | # Directories potentially created on remote AFP share 156 | .AppleDB 157 | .AppleDesktop 158 | Network Trash Folder 159 | Temporary Items 160 | .apdisk 161 | 162 | 163 | # https://github.com/github/gitignore/blob/master/Global/JetBrains.gitignore 164 | # Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio and WebStorm 165 | # Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839 166 | 167 | # User-specific stuff: 168 | .idea/* 169 | .idea/**/workspace.xml 170 | .idea/**/tasks.xml 171 | .idea/dictionaries 172 | .html # Bokeh Plots 173 | .pg # TensorFlow Frozen Graphs 174 | .avi # videos 175 | 176 | # Sensitive or high-churn files: 177 | .idea/**/dataSources/ 178 | .idea/**/dataSources.ids 179 | .idea/**/dataSources.local.xml 180 | .idea/**/sqlDataSources.xml 181 | .idea/**/dynamic.xml 182 | .idea/**/uiDesigner.xml 183 | 184 | # Gradle: 185 | .idea/**/gradle.xml 186 | .idea/**/libraries 187 | 188 | # CMake 189 | cmake-build-debug/ 190 | cmake-build-release/ 191 | 192 | # Mongo Explorer plugin: 193 | .idea/**/mongoSettings.xml 194 | 195 | ## File-based project format: 196 | *.iws 197 | 198 | ## Plugin-specific files: 199 | 200 | # IntelliJ 201 | out/ 202 | 203 | # mpeltonen/sbt-idea plugin 204 | .idea_modules/ 205 | 206 | # JIRA plugin 207 | atlassian-ide-plugin.xml 208 | 209 | # Cursive Clojure plugin 210 | .idea/replstate.xml 211 | 212 | # Crashlytics plugin (for Android Studio and IntelliJ) 213 | com_crashlytics_export_strings.xml 214 | crashlytics.properties 215 | crashlytics-build.properties 216 | fabric.properties 217 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Repo-specific GitIgnore ---------------------------------------------------------------------------------------------- 2 | *.jpg 3 | *.jpeg 4 | *.bmp 5 | *.tif 6 | *.tiff 7 | *.heic 8 | *.JPG 9 | *.JPEG 10 | *.PNG 11 | *.BMP 12 | *.TIF 13 | *.TIFF 14 | *.HEIC 15 | *.mp4 16 | *.mov 17 | *.MOV 18 | *.avi 19 | *.data 20 | *.json 21 | *.pt 22 | *.pth 23 | data/Mini 24 | runs/* 25 | data/ztdiva.names 26 | data/mini.yaml 27 | runs 28 | runs/* 29 | data/test 30 | data/test/* 31 | *.pybk 32 | 33 | *.cfg 34 | !cfg/yolov3*.cfg 35 | 36 | storage.googleapis.com 37 | runs/* 38 | 39 | !data/images/zidane.jpg 40 | !data/images/bus.jpg 41 | !data/coco.names 42 | !data/coco_paper.names 43 | !data/coco.data 44 | !data/coco_*.data 45 | !data/coco_*.txt 46 | !data/trainvalno5k.shapes 47 | !data/*.sh 48 | 49 | pycocotools/* 50 | results*.txt 51 | gcp_test*.sh 52 | 53 | # Datasets ------------------------------------------------------------------------------------------------------------- 54 | coco/ 55 | coco128/ 56 | VOC/ 57 | 58 | # MATLAB GitIgnore ----------------------------------------------------------------------------------------------------- 59 | *.m~ 60 | *.mat 61 | !targets*.mat 62 | 63 | # Neural Network weights ----------------------------------------------------------------------------------------------- 64 | *.weights 65 | *.pt 66 | *.onnx 67 | *.mlmodel 68 | *.torchscript 69 | darknet53.conv.74 70 | yolov3-tiny.conv.15 71 | 72 | # GitHub Python GitIgnore ---------------------------------------------------------------------------------------------- 73 | # Byte-compiled / optimized / DLL files 74 | __pycache__/ 75 | *.py[cod] 76 | *$py.class 77 | 78 | # C extensions 79 | *.so 80 | 81 | # Distribution / packaging 82 | .Python 83 | env/ 84 | build/ 85 | develop-eggs/ 86 | dist/ 87 | downloads/ 88 | eggs/ 89 | .eggs/ 90 | lib/ 91 | lib64/ 92 | parts/ 93 | sdist/ 94 | var/ 95 | wheels/ 96 | *.egg-info/ 97 | wandb/ 98 | .installed.cfg 99 | *.egg 100 | 101 | 102 | # PyInstaller 103 | # Usually these files are written by a python script from a template 104 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 105 | *.manifest 106 | *.spec 107 | 108 | # Installer logs 109 | pip-log.txt 110 | pip-delete-this-directory.txt 111 | 112 | # Unit test / coverage reports 113 | htmlcov/ 114 | .tox/ 115 | .coverage 116 | .coverage.* 117 | .cache 118 | nosetests.xml 119 | coverage.xml 120 | *.cover 121 | .hypothesis/ 122 | 123 | # Translations 124 | *.mo 125 | *.pot 126 | 127 | # Django stuff: 128 | *.log 129 | local_settings.py 130 | 131 | # Flask stuff: 132 | instance/ 133 | .webassets-cache 134 | 135 | # Scrapy stuff: 136 | .scrapy 137 | 138 | # Sphinx documentation 139 | docs/_build/ 140 | 141 | # PyBuilder 142 | target/ 143 | 144 | # Jupyter Notebook 145 | .ipynb_checkpoints 146 | 147 | # pyenv 148 | .python-version 149 | 150 | # celery beat schedule file 151 | celerybeat-schedule 152 | 153 | # SageMath parsed files 154 | *.sage.py 155 | 156 | # dotenv 157 | .env 158 | 159 | # virtualenv 160 | .venv* 161 | venv*/ 162 | ENV*/ 163 | 164 | # Spyder project settings 165 | .spyderproject 166 | .spyproject 167 | 168 | # Rope project settings 169 | .ropeproject 170 | 171 | # mkdocs documentation 172 | /site 173 | 174 | # mypy 175 | .mypy_cache/ 176 | 177 | 178 | # https://github.com/github/gitignore/blob/master/Global/macOS.gitignore ----------------------------------------------- 179 | 180 | # General 181 | .DS_Store 182 | .AppleDouble 183 | .LSOverride 184 | 185 | # Icon must end with two \r 186 | Icon 187 | Icon? 188 | 189 | # Thumbnails 190 | ._* 191 | 192 | # Files that might appear in the root of a volume 193 | .DocumentRevisions-V100 194 | .fseventsd 195 | .Spotlight-V100 196 | .TemporaryItems 197 | .Trashes 198 | .VolumeIcon.icns 199 | .com.apple.timemachine.donotpresent 200 | 201 | # Directories potentially created on remote AFP share 202 | .AppleDB 203 | .AppleDesktop 204 | Network Trash Folder 205 | Temporary Items 206 | .apdisk 207 | 208 | 209 | # https://github.com/github/gitignore/blob/master/Global/JetBrains.gitignore 210 | # Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio and WebStorm 211 | # Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839 212 | 213 | # User-specific stuff: 214 | .idea/* 215 | .idea/**/workspace.xml 216 | .idea/**/tasks.xml 217 | .idea/dictionaries 218 | .html # Bokeh Plots 219 | .pg # TensorFlow Frozen Graphs 220 | .avi # videos 221 | 222 | # Sensitive or high-churn files: 223 | .idea/**/dataSources/ 224 | .idea/**/dataSources.ids 225 | .idea/**/dataSources.local.xml 226 | .idea/**/sqlDataSources.xml 227 | .idea/**/dynamic.xml 228 | .idea/**/uiDesigner.xml 229 | 230 | # Gradle: 231 | .idea/**/gradle.xml 232 | .idea/**/libraries 233 | 234 | # CMake 235 | cmake-build-debug/ 236 | cmake-build-release/ 237 | 238 | # Mongo Explorer plugin: 239 | .idea/**/mongoSettings.xml 240 | 241 | ## File-based project format: 242 | *.iws 243 | 244 | ## Plugin-specific files: 245 | 246 | # IntelliJ 247 | out/ 248 | 249 | # mpeltonen/sbt-idea plugin 250 | .idea_modules/ 251 | 252 | # JIRA plugin 253 | atlassian-ide-plugin.xml 254 | 255 | # Cursive Clojure plugin 256 | .idea/replstate.xml 257 | 258 | # Crashlytics plugin (for Android Studio and IntelliJ) 259 | com_crashlytics_export_strings.xml 260 | crashlytics.properties 261 | crashlytics-build.properties 262 | fabric.properties 263 | -------------------------------------------------------------------------------- /Dockerfile: -------------------------------------------------------------------------------- 1 | # Start FROM Nvidia PyTorch image https://ngc.nvidia.com/catalog/containers/nvidia:pytorch 2 | FROM nvcr.io/nvidia/pytorch:21.03-py3 3 | 4 | # Install linux packages 5 | RUN apt update && apt install -y zip htop screen libgl1-mesa-glx 6 | 7 | # Install python dependencies 8 | COPY requirements.txt . 9 | RUN python -m pip install --upgrade pip 10 | RUN pip uninstall -y nvidia-tensorboard nvidia-tensorboard-plugin-dlprof 11 | RUN pip install --no-cache -r requirements.txt coremltools onnx gsutil notebook 12 | 13 | # Create working directory 14 | RUN mkdir -p /usr/src/app 15 | WORKDIR /usr/src/app 16 | 17 | # Copy contents 18 | COPY . /usr/src/app 19 | 20 | # Set environment variables 21 | ENV HOME=/usr/src/app 22 | 23 | 24 | # --------------------------------------------------- Extras Below --------------------------------------------------- 25 | 26 | # Build and Push 27 | # t=ultralytics/yolov5:latest && sudo docker build -t $t . && sudo docker push $t 28 | # for v in {300..303}; do t=ultralytics/coco:v$v && sudo docker build -t $t . && sudo docker push $t; done 29 | 30 | # Pull and Run 31 | # t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all $t 32 | 33 | # Pull and Run with local directory access 34 | # t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all -v "$(pwd)"/coco:/usr/src/coco $t 35 | 36 | # Kill all 37 | # sudo docker kill $(sudo docker ps -q) 38 | 39 | # Kill all image-based 40 | # sudo docker kill $(sudo docker ps -qa --filter ancestor=ultralytics/yolov5:latest) 41 | 42 | # Bash into running container 43 | # sudo docker exec -it 5a9b5863d93d bash 44 | 45 | # Bash into stopped container 46 | # id=$(sudo docker ps -qa) && sudo docker start $id && sudo docker exec -it $id bash 47 | 48 | # Send weights to GCP 49 | # python -c "from utils.general import *; strip_optimizer('runs/train/exp0_*/weights/best.pt', 'tmp.pt')" && gsutil cp tmp.pt gs://*.pt 50 | 51 | # Clean up 52 | # docker system prune -a --volumes 53 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # yolov5模型剪枝 2 | 3 | **`2022-1-4`**: 已更新v5.0版本m/l/x模型剪枝,理论上yolov5l6等模型也支持. 4 | 5 | **`2022-1-1`**: 已更新v6.0版本剪枝:https://github.com/midasklr/yolov5prune/tree/v6.0 6 | 7 | **`2021-12-14`**:近期会更新v6.0版本剪枝和蒸馏. 8 | 9 | 10 | 基于yolov5最新v5.0进行剪枝，采用yolov5s模型,目前仅支持s模型。 11 | 12 | 相关原理： 13 | 14 | Learning Efficient Convolutional Networks Through Network Slimming（https://arxiv.org/abs/1708.06519） 15 | 16 | Pruning Filters for Efficient ConvNets（https://arxiv.org/abs/1608.08710） 17 | 18 | 相关原理见https://blog.csdn.net/IEEE_FELLOW/article/details/117236025 19 | 20 | 这里实验了三种剪枝方式 21 | 22 | ## 剪枝方法1 23 | 24 | 基于BN层系数gamma剪枝。 25 | 26 | 在一个卷积-BN-激活模块中，BN层可以实现通道的缩放。如下： 27 | 28 |

29 | 30 |

31 | 32 | BN层的具体操作有两部分： 33 | 34 |

35 | 36 |

37 | 38 | 在归一化后会进行线性变换，那么当系数gamma很小时候，对应的激活（Zout）会相应很小。这些响应很小的输出可以裁剪掉，这样就实现了bn层的通道剪枝。 39 | 40 | 通过在loss函数中添加gamma的L1正则约束，可以实现gamma的稀疏化。 41 | 42 |

43 | 44 |

45 | 46 | 47 | 48 | 上面损失函数L右边第一项是原始的损失函数，第二项是约束，其中g(s) = |s|，λ是正则系数，根据数据集调整 49 | 50 | 实际训练的时候，就是在优化L最小，依据梯度下降算法： 51 | 52 | ​ 𝐿′=∑𝑙′+𝜆∑𝑔′(𝛾)=∑𝑙′+𝜆∑|𝛾|′=∑𝑙′+𝜆∑𝛾∗𝑠𝑖𝑔𝑛(𝛾) 53 | 54 | 所以只需要在BP传播时候，在BN层权重乘以权重的符号函数输出和系数即可，对应添加如下代码: 55 | 56 | ```python 57 | # Backward 58 | loss.backward() 59 | # scaler.scale(loss).backward() 60 | # # ============================= sparsity training ========================== # 61 | srtmp = opt.sr*(1 - 0.9*epoch/epochs) 62 | if opt.st: 63 | ignore_bn_list = [] 64 | for k, m in model.named_modules(): 65 | if isinstance(m, Bottleneck): 66 | if m.add: 67 | ignore_bn_list.append(k.rsplit(".", 2)[0] + ".cv1.bn") 68 | ignore_bn_list.append(k + '.cv1.bn') 69 | ignore_bn_list.append(k + '.cv2.bn') 70 | if isinstance(m, nn.BatchNorm2d) and (k not in ignore_bn_list): 71 | m.weight.grad.data.add_(srtmp * torch.sign(m.weight.data)) # L1 72 | m.bias.grad.data.add_(opt.sr*10 * torch.sign(m.bias.data)) # L1 73 | # # ============================= sparsity training ========================== # 74 | 75 | optimizer.step() 76 | # scaler.step(optimizer) # optimizer.step 77 | # scaler.update() 78 | optimizer.zero_grad() 79 | ``` 80 | 81 | 这里并未对所有BN层gamma进行约束，详情见yolov5s每个模块 https://blog.csdn.net/IEEE_FELLOW/article/details/117536808 82 | 分析，这里对C3结构中的Bottleneck结构中有shortcut的层不进行剪枝，主要是为了保持tensor维度可以加： 83 | 84 |

85 | 86 |

87 | 88 | 实际上，在yolov5中，只有backbone中的Bottleneck是有shortcut的，Head中全部没有shortcut. 89 | 90 | 如果不加L1正则约束，训练结束后的BN层gamma分布近似正太分布： 91 | 92 |

93 | 94 |

95 | 96 | 是无法进行剪枝的。 97 | 98 | 稀疏训练后的分布： 99 | 100 |

101 | 102 |

103 | 104 | 可以看到，随着训练epoch进行，越来越多的gamma逼近0. 105 | 106 | 训练完成后可以进行剪枝，一个基本的原则是阈值不能大于任何通道bn的最大gamma。然后根据设定的裁剪比例剪枝。 107 | 108 | 剪掉一个BN层，需要将对应上一层的卷积核裁剪掉，同时将下一层卷积核对应的通道减掉。 109 | 110 | 这里在某个数据集上实验。 111 | 112 | 首先使用train.py进行正常训练： 113 | 114 | ``` 115 | python train.py --weights yolov5s.pt --adam --epochs 100 116 | ``` 117 | 118 | 然后稀疏训练： 119 | 120 | ``` 121 | python train_sparsity.py --st --sr 0.0001 --weights yolov5s.pt --adam --epochs 100 122 | ``` 123 | 124 | sr的选择需要根据数据集调整，可以通过观察tensorboard的map，gamma变化直方图等选择。 125 | 在run/train/exp*/目录下: 126 | ``` 127 | tensorboard --logdir . 128 | ``` 129 | 然后点击出现的链接观察训练中的各项指标. 130 | 131 | 训练完成后进行剪枝： 132 | 133 | ``` 134 | python prune.py --weights runs/train/exp1/weights/last.pt --percent 0.5 --cfg models/yolov5s.yaml 135 | ``` 136 | 137 | 裁剪比例percent根据效果调整，可以从小到大试。注意cfg的模型文件需要和weights对应上,否则会出现[运行prune 过程中出现键值不对应的问题](https://github.com/midasklr/yolov5prune/issues/65),裁剪完成会保存对应的模型pruned_model.pt。 138 | 139 | 微调： 140 | 141 | ``` 142 | python finetune_pruned.py --weights pruned_model.pt --adam --epochs 100 143 | ``` 144 | 145 | 在VOC2007数据集上实验,训练集为VOC07 trainval, 测试集为VOC07 test.作为对比,这里列举了faster rcnn和SSD512在相同数据集上的实验结果, yolov5输入大小为512.为了节省时间,这里使用AdamW训练100 epoch. 146 | 147 | | model | optim&epoch | sparity | mAP@.5 | mode size | forward time | 148 | | ----------------- | ----------- | ------- | ----------- | --------- | ------------ | 149 | | faster rcnn | | - | 69.9(paper) | | | 150 | | SSD512 | | - | 71.6(paper) | | | 151 | | yolov5s | sgd 300 | 0 | 67.4 | | | 152 | | yolov5s | adamw 100 | 0 | 66.3 | | | 153 | | yolov5s | adamw 100 | 0.0001 | 69.2 | | | 154 | | yolov5s | sgd 300 | 0.001 | Inf. error | | | 155 | | yolov5s | adamw 100 | 0.001 | 65.7 | 28.7 | 7.32 ms | 156 | | 55% prune yolov5s | | | 64.1 | 8.6 | 7.30 ms | 157 | | fine-tune above | | | 67.3 | | 7.21 ms | 158 | | yolov5l | adamw 100 | 0 | 70.1 | | | 159 | | yolov5l | adamw 100 | 0.001 | 0.659 | | 12.95 ms | 160 | 161 | 162 | 163 | 在自己数据集上的实验结果: 164 | 165 | | model | sparity | map | mode size | 166 | | --------------------- | ------- | ----- | --------- | 167 | | yolov5s | 0 | 0.322 | 28.7 M | 168 | | sparity train yolov5s | 0.001 | 0.325 | 28.7 M | 169 | | 65% pruned yolov5s | 0.001 | 0.318 | 6.8 M | 170 | | fine-tune | 0 | 0.325 | 6.8 M | 171 | 172 | ## 剪枝方法2 173 | 174 | 对于Bottleneck结构： 175 | 176 |

177 | 178 |

179 | 180 | 如果有右边的参差很小，那么就只剩下左边shortcut连接，相当于整个模块都裁剪掉。可以进行约束让参差逼近0.见train_sparsity2.py。 181 | 182 | backbone一共有3个bottleneck，裁剪全部bottleneck： 183 | 184 | | model | sparity | map | model size | 185 | | --------------------------- | ------- | ----- | ---------- | 186 | | yolov5s-prune all bottlenet | 0.001 | 0.167 | 28.7 M | 187 | | 85%+Bottlenet | | 0.151 | 1.1 M | 188 | | finetune | | 0.148 | | 189 | 190 | | 裁剪Bottleneck数 | map | 191 | | ----------------- | ----- | 192 | | 所有bottle res | 0.167 | 193 | | 第2,3的bottle res | 0.174 | 194 | | 第3的bottle res | 0.198 | 195 | 196 | 可以看到实际效果并不好，从bn层分布也可以看到，浅层特征很少被裁减掉。 197 | 198 | ## 剪枝方法3 199 | 200 | 卷积核剪枝，那些权重很小的卷积核对应输出也较小，那么对kernel进行约束，是可以对卷积核进行裁剪的。 201 | 202 | 裁剪卷积核需要将下一层BN层对应裁剪，同时裁剪下一层卷积层的输出通道。见train_sparsity3.py 203 | 204 | | | s | model size | map | 205 | | ---------------- | ---- | ---------- | ----- | 206 | | sparity train | 1e-5 | 28.7 M | 0.335 | 207 | | 50% kernel prune | | 8.4 M | 0.151 | 208 | | finetune | | 8.4 M | 0.332 | 209 | 210 | ## 剪枝方法4 211 | 212 | 混合1和3，见train_sparsity4.py 213 | 214 | | | map | model size | 215 | | --------------------------- | ----- | ---------- | 216 | | conv+bn sparity train | 0.284 | 28.7 M | 217 | | 85% bn prune | 0.284 | 3.7 M | 218 | | 78% conv prune | 0.284 | 3.9 M | 219 | | 85% bn prune+78% conv prune | 0.284 | 3.7 M | 220 | 221 | 222 | ## 替换backbone 223 | 224 | | model | size | mAPval 0.5:0.95 | mAPval 0.5 | 225 | | --------------------------- | ----- | ---------- | ------- | 226 | | yolov5s | 640 | 0.357 | 0.558 | 227 | | mobilenetv3small 0.75 | 640 | 0.24 | 0.421 | 228 | 229 | 230 | 231 | ## 调参 232 | 1. 浅层尽量少剪,从训练完成后gamma每一层的分布也可以看出来. 233 | 2. 系数λ的选择需要平衡map和剪枝力度.首先通过train.py训练一个正常情况下的baseline.然后在稀疏训练过程中观察MAP和gamma直方图变化,MAP掉点严重和gamma稀疏过快等情况下,可以适当降低λ.反之如果你想压缩一个尽量小的模型,可以适当调整λ. 234 | 3. 稀疏训练=>剪枝=>微调 可以反复迭代这个过程多次剪枝. 235 | 4. 使用yolov5默认的一些参数通常效果能获得不错的效果，比如使用SGD训练300 epoch，lr 0.01->0.001等，这里实验为了快速选用adamw训练了100 epoch。 236 | 5. 看到许多小伙伴提出了很多问题，有的我也没碰到，能解答的尽量解答。 237 | 6. 剪枝多少参数，有的是时候和数据集关系很大，我分别在简单任务（5k images,40+ class）和复杂数据集（20w+ images， 120+ class）实验过，简单任务可以将模型剪到很小（小模型也相对不够鲁棒）；复杂的任务最终参数较难稀疏，能剪的参数很少（<20%）。 238 | 7. yolov5的s,m,l,x四个模型结构是一样的，只是深度和宽度两个维度的缩放系数不同，所以本代码应该也适用m,l,x模型。 239 | 8. 可以试试用大模型开始剪枝，比如用yolov5l,可能比直接用yolov5s开始剪枝效果更好？大模型的搜索空间通常更大。 240 | 9. 在自己的数据集上,设置合理的输入往往很重要, 公开数据集VOC和COCO等通常做了处理,例如VOC长边都是500, COCO长边都是640, 这也是SSD设置输入300和512, yolov5设置输入640的一个重要原因.如果要在自己数据集上获得较好的性能,可以试试调整输入. 241 | 242 | ## 常见问题 243 | 1. 稀疏训练是非常种重要的,也是调参的重点,多观察bn直方图变化,过快或者过慢都不适合,所以需要平衡你的sr, lr等.一般情况下,稀疏训练的结果和正常训练map是比较接近的. 244 | 2. 剪枝时候多试试不同的ratio,一个基本的准则是每层bn层至少保留一个channel,所以有时候稀疏训练不到位,而ratio设置的很大,会看到remaining channel里面会有0出现,这时候要么设置更小的ratio,要么重新稀疏训练,获得更稀疏的参数. 245 | 3. 如果想要移植到移动端，可以使用ncnn加速，另外剪枝时控制剩余channel为2^n能有效提升推理速度；GPU可以使用TensorRT加速。 246 | -------------------------------------------------------------------------------- /data/GlobalWheat2020.yaml: -------------------------------------------------------------------------------- 1 | # Global Wheat 2020 dataset http://www.global-wheat.com/ 2 | # Train command: python train.py --data GlobalWheat2020.yaml 3 | # Default dataset location is next to YOLOv5: 4 | # /parent_folder 5 | # /datasets/GlobalWheat2020 6 | # /yolov5 7 | 8 | 9 | # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] 10 | train: # 3422 images 11 | - ../datasets/GlobalWheat2020/images/arvalis_1 12 | - ../datasets/GlobalWheat2020/images/arvalis_2 13 | - ../datasets/GlobalWheat2020/images/arvalis_3 14 | - ../datasets/GlobalWheat2020/images/ethz_1 15 | - ../datasets/GlobalWheat2020/images/rres_1 16 | - ../datasets/GlobalWheat2020/images/inrae_1 17 | - ../datasets/GlobalWheat2020/images/usask_1 18 | 19 | val: # 748 images (WARNING: train set contains ethz_1) 20 | - ../datasets/GlobalWheat2020/images/ethz_1 21 | 22 | test: # 1276 23 | - ../datasets/GlobalWheat2020/images/utokyo_1 24 | - ../datasets/GlobalWheat2020/images/utokyo_2 25 | - ../datasets/GlobalWheat2020/images/nau_1 26 | - ../datasets/GlobalWheat2020/images/uq_1 27 | 28 | # number of classes 29 | nc: 1 30 | 31 | # class names 32 | names: [ 'wheat_head' ] 33 | 34 | 35 | # download command/URL (optional) -------------------------------------------------------------------------------------- 36 | download: | 37 | from utils.general import download, Path 38 | 39 | # Download 40 | dir = Path('../datasets/GlobalWheat2020') # dataset directory 41 | urls = ['https://zenodo.org/record/4298502/files/global-wheat-codalab-official.zip', 42 | 'https://github.com/ultralytics/yolov5/releases/download/v1.0/GlobalWheat2020_labels.zip'] 43 | download(urls, dir=dir) 44 | 45 | # Make Directories 46 | for p in 'annotations', 'images', 'labels': 47 | (dir / p).mkdir(parents=True, exist_ok=True) 48 | 49 | # Move 50 | for p in 'arvalis_1', 'arvalis_2', 'arvalis_3', 'ethz_1', 'rres_1', 'inrae_1', 'usask_1', \ 51 | 'utokyo_1', 'utokyo_2', 'nau_1', 'uq_1': 52 | (dir / p).rename(dir / 'images' / p) # move to /images 53 | f = (dir / p).with_suffix('.json') # json file 54 | if f.exists(): 55 | f.rename((dir / 'annotations' / p).with_suffix('.json')) # move to /annotations 56 | -------------------------------------------------------------------------------- /data/VisDrone.yaml: -------------------------------------------------------------------------------- 1 | # VisDrone2019-DET dataset https://github.com/VisDrone/VisDrone-Dataset 2 | # Train command: python train.py --data VisDrone.yaml 3 | # Default dataset location is next to YOLOv5: 4 | # /parent_folder 5 | # /VisDrone 6 | # /yolov5 7 | 8 | 9 | # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] 10 | train: ../VisDrone/VisDrone2019-DET-train/images # 6471 images 11 | val: ../VisDrone/VisDrone2019-DET-val/images # 548 images 12 | test: ../VisDrone/VisDrone2019-DET-test-dev/images # 1610 images 13 | 14 | # number of classes 15 | nc: 10 16 | 17 | # class names 18 | names: [ 'pedestrian', 'people', 'bicycle', 'car', 'van', 'truck', 'tricycle', 'awning-tricycle', 'bus', 'motor' ] 19 | 20 | 21 | # download command/URL (optional) -------------------------------------------------------------------------------------- 22 | download: | 23 | from utils.general import download, os, Path 24 | 25 | def visdrone2yolo(dir): 26 | from PIL import Image 27 | from tqdm import tqdm 28 | 29 | def convert_box(size, box): 30 | # Convert VisDrone box to YOLO xywh box 31 | dw = 1. / size[0] 32 | dh = 1. / size[1] 33 | return (box[0] + box[2] / 2) * dw, (box[1] + box[3] / 2) * dh, box[2] * dw, box[3] * dh 34 | 35 | (dir / 'labels').mkdir(parents=True, exist_ok=True) # make labels directory 36 | pbar = tqdm((dir / 'annotations').glob('*.txt'), desc=f'Converting {dir}') 37 | for f in pbar: 38 | img_size = Image.open((dir / 'images' / f.name).with_suffix('.jpg')).size 39 | lines = [] 40 | with open(f, 'r') as file: # read annotation.txt 41 | for row in [x.split(',') for x in file.read().strip().splitlines()]: 42 | if row[4] == '0': # VisDrone 'ignored regions' class 0 43 | continue 44 | cls = int(row[5]) - 1 45 | box = convert_box(img_size, tuple(map(int, row[:4]))) 46 | lines.append(f"{cls} {' '.join(f'{x:.6f}' for x in box)}\n") 47 | with open(str(f).replace(os.sep + 'annotations' + os.sep, os.sep + 'labels' + os.sep), 'w') as fl: 48 | fl.writelines(lines) # write label.txt 49 | 50 | 51 | # Download 52 | dir = Path('../VisDrone') # dataset directory 53 | urls = ['https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-train.zip', 54 | 'https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-val.zip', 55 | 'https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-test-dev.zip', 56 | 'https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-test-challenge.zip'] 57 | download(urls, dir=dir) 58 | 59 | # Convert 60 | for d in 'VisDrone2019-DET-train', 'VisDrone2019-DET-val', 'VisDrone2019-DET-test-dev': 61 | visdrone2yolo(dir / d) # convert VisDrone annotations to YOLO labels 62 | -------------------------------------------------------------------------------- /data/argoverse_hd.yaml: -------------------------------------------------------------------------------- 1 | # Argoverse-HD dataset (ring-front-center camera) http://www.cs.cmu.edu/~mengtial/proj/streaming/ 2 | # Train command: python train.py --data argoverse_hd.yaml 3 | # Default dataset location is next to YOLOv5: 4 | # /parent_folder 5 | # /argoverse 6 | # /yolov5 7 | 8 | 9 | # download command/URL (optional) 10 | download: bash data/scripts/get_argoverse_hd.sh 11 | 12 | # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] 13 | train: ../argoverse/Argoverse-1.1/images/train/ # 39384 images 14 | val: ../argoverse/Argoverse-1.1/images/val/ # 15062 iamges 15 | test: ../argoverse/Argoverse-1.1/images/test/ # Submit to: https://eval.ai/web/challenges/challenge-page/800/overview 16 | 17 | # number of classes 18 | nc: 8 19 | 20 | # class names 21 | names: [ 'person', 'bicycle', 'car', 'motorcycle', 'bus', 'truck', 'traffic_light', 'stop_sign' ] 22 | -------------------------------------------------------------------------------- /data/coco.yaml: -------------------------------------------------------------------------------- 1 | # COCO 2017 dataset http://cocodataset.org 2 | # Train command: python train.py --data coco.yaml 3 | # Default dataset location is next to YOLOv5: 4 | # /parent_folder 5 | # /coco 6 | # /yolov5 7 | 8 | 9 | # download command/URL (optional) 10 | download: bash data/scripts/get_coco.sh 11 | 12 | # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] 13 | train: ../coco/train2017.txt # 118287 images 14 | val: ../coco/val2017.txt # 5000 images 15 | test: ../coco/test-dev2017.txt # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794 16 | 17 | # number of classes 18 | nc: 80 19 | 20 | # class names 21 | names: [ 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 22 | 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 23 | 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 24 | 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 25 | 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 26 | 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 27 | 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 28 | 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 29 | 'hair drier', 'toothbrush' ] 30 | 31 | # Print classes 32 | # with open('data/coco.yaml') as f: 33 | # d = yaml.safe_load(f) # dict 34 | # for i, x in enumerate(d['names']): 35 | # print(i, x) 36 | -------------------------------------------------------------------------------- /data/coco128.yaml: -------------------------------------------------------------------------------- 1 | # COCO 2017 dataset http://cocodataset.org - first 128 training images 2 | # Train command: python train.py --data coco128.yaml 3 | # Default dataset location is next to YOLOv5: 4 | # /parent_folder 5 | # /coco128 6 | # /yolov5 7 | 8 | 9 | # download command/URL (optional) 10 | download: https://github.com/ultralytics/yolov5/releases/download/v1.0/coco128.zip 11 | 12 | # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] 13 | train: ../coco128/images/train2017/ # 128 images 14 | val: ../coco128/images/train2017/ # 128 images 15 | 16 | # number of classes 17 | nc: 80 18 | 19 | # class names 20 | names: [ 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 21 | 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 22 | 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 23 | 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 24 | 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 25 | 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 26 | 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 27 | 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 28 | 'hair drier', 'toothbrush' ] 29 | -------------------------------------------------------------------------------- /data/hyp.finetune.yaml: -------------------------------------------------------------------------------- 1 | # Hyperparameters for VOC finetuning 2 | # python train.py --batch 64 --weights yolov5m.pt --data voc.yaml --img 512 --epochs 50 3 | # See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials 4 | 5 | 6 | # Hyperparameter Evolution Results 7 | # Generations: 306 8 | # P R mAP.5 mAP.5:.95 box obj cls 9 | # Metrics: 0.6 0.936 0.896 0.684 0.0115 0.00805 0.00146 10 | 11 | lr0: 0.001 12 | lrf: 0.2 13 | momentum: 0.843 14 | weight_decay: 0.00036 15 | warmup_epochs: 2.0 16 | warmup_momentum: 0.5 17 | warmup_bias_lr: 0.05 18 | box: 0.0296 19 | cls: 0.243 20 | cls_pw: 0.631 21 | obj: 0.301 22 | obj_pw: 0.911 23 | iou_t: 0.2 24 | anchor_t: 2.91 25 | # anchors: 3.63 26 | fl_gamma: 0.0 27 | hsv_h: 0.0138 28 | hsv_s: 0.664 29 | hsv_v: 0.464 30 | degrees: 0.373 31 | translate: 0.245 32 | scale: 0.898 33 | shear: 0.602 34 | perspective: 0.0 35 | flipud: 0.00856 36 | fliplr: 0.5 37 | mosaic: 1.0 38 | mixup: 0.243 39 | -------------------------------------------------------------------------------- /data/hyp.finetune_objects365.yaml: -------------------------------------------------------------------------------- 1 | lr0: 0.00258 2 | lrf: 0.17 3 | momentum: 0.779 4 | weight_decay: 0.00058 5 | warmup_epochs: 1.33 6 | warmup_momentum: 0.86 7 | warmup_bias_lr: 0.0711 8 | box: 0.0539 9 | cls: 0.299 10 | cls_pw: 0.825 11 | obj: 0.632 12 | obj_pw: 1.0 13 | iou_t: 0.2 14 | anchor_t: 3.44 15 | anchors: 3.2 16 | fl_gamma: 0.0 17 | hsv_h: 0.0188 18 | hsv_s: 0.704 19 | hsv_v: 0.36 20 | degrees: 0.0 21 | translate: 0.0902 22 | scale: 0.491 23 | shear: 0.0 24 | perspective: 0.0 25 | flipud: 0.0 26 | fliplr: 0.5 27 | mosaic: 1.0 28 | mixup: 0.0 29 | -------------------------------------------------------------------------------- /data/hyp.scratch.yaml: -------------------------------------------------------------------------------- 1 | # Hyperparameters for COCO training from scratch 2 | # python train.py --batch 40 --cfg yolov5m.yaml --weights '' --data coco.yaml --img 640 --epochs 300 3 | # See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials 4 | 5 | 6 | lr0: 0.001 # initial learning rate (SGD=1E-2, Adam=1E-3) 7 | lrf: 0.1 # final OneCycleLR learning rate (lr0 * lrf) 8 | momentum: 0.937 # SGD momentum/Adam beta1 9 | weight_decay: 0.0005 # optimizer weight decay 5e-4 10 | warmup_epochs: 3.0 # warmup epochs (fractions ok) 11 | warmup_momentum: 0.8 # warmup initial momentum 12 | warmup_bias_lr: 0.1 # warmup initial bias lr 13 | box: 0.05 # box loss gain 14 | cls: 0.5 # cls loss gain 15 | cls_pw: 1.0 # cls BCELoss positive_weight 16 | obj: 1.0 # obj loss gain (scale with pixels) 17 | obj_pw: 1.0 # obj BCELoss positive_weight 18 | iou_t: 0.20 # IoU training threshold 19 | anchor_t: 4.0 # anchor-multiple threshold 20 | # anchors: 3 # anchors per output layer (0 to ignore) 21 | fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) 22 | hsv_h: 0.015 # image HSV-Hue augmentation (fraction) 23 | hsv_s: 0.7 # image HSV-Saturation augmentation (fraction) 24 | hsv_v: 0.4 # image HSV-Value augmentation (fraction) 25 | degrees: 0.0 # image rotation (+/- deg) 26 | translate: 0.1 # image translation (+/- fraction) 27 | scale: 0.5 # image scale (+/- gain) 28 | shear: 0.0 # image shear (+/- deg) 29 | perspective: 0.0 # image perspective (+/- fraction), range 0-0.001 30 | flipud: 0.0 # image flip up-down (probability) 31 | fliplr: 0.5 # image flip left-right (probability) 32 | mosaic: 1.0 # image mosaic (probability) 33 | mixup: 0.0 # image mixup (probability) 34 | -------------------------------------------------------------------------------- /data/images/bus.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/data/images/bus.jpg -------------------------------------------------------------------------------- /data/images/zidane.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/data/images/zidane.jpg -------------------------------------------------------------------------------- /data/objects365.yaml: -------------------------------------------------------------------------------- 1 | # Objects365 dataset https://www.objects365.org/ 2 | # Train command: python train.py --data objects365.yaml 3 | # Default dataset location is next to YOLOv5: 4 | # /parent_folder 5 | # /datasets/objects365 6 | # /yolov5 7 | 8 | # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] 9 | train: ../datasets/objects365/images/train # 1742289 images 10 | val: ../datasets/objects365/images/val # 5570 images 11 | 12 | # number of classes 13 | nc: 365 14 | 15 | # class names 16 | names: [ 'Person', 'Sneakers', 'Chair', 'Other Shoes', 'Hat', 'Car', 'Lamp', 'Glasses', 'Bottle', 'Desk', 'Cup', 17 | 'Street Lights', 'Cabinet/shelf', 'Handbag/Satchel', 'Bracelet', 'Plate', 'Picture/Frame', 'Helmet', 'Book', 18 | 'Gloves', 'Storage box', 'Boat', 'Leather Shoes', 'Flower', 'Bench', 'Potted Plant', 'Bowl/Basin', 'Flag', 19 | 'Pillow', 'Boots', 'Vase', 'Microphone', 'Necklace', 'Ring', 'SUV', 'Wine Glass', 'Belt', 'Monitor/TV', 20 | 'Backpack', 'Umbrella', 'Traffic Light', 'Speaker', 'Watch', 'Tie', 'Trash bin Can', 'Slippers', 'Bicycle', 21 | 'Stool', 'Barrel/bucket', 'Van', 'Couch', 'Sandals', 'Basket', 'Drum', 'Pen/Pencil', 'Bus', 'Wild Bird', 22 | 'High Heels', 'Motorcycle', 'Guitar', 'Carpet', 'Cell Phone', 'Bread', 'Camera', 'Canned', 'Truck', 23 | 'Traffic cone', 'Cymbal', 'Lifesaver', 'Towel', 'Stuffed Toy', 'Candle', 'Sailboat', 'Laptop', 'Awning', 24 | 'Bed', 'Faucet', 'Tent', 'Horse', 'Mirror', 'Power outlet', 'Sink', 'Apple', 'Air Conditioner', 'Knife', 25 | 'Hockey Stick', 'Paddle', 'Pickup Truck', 'Fork', 'Traffic Sign', 'Balloon', 'Tripod', 'Dog', 'Spoon', 'Clock', 26 | 'Pot', 'Cow', 'Cake', 'Dinning Table', 'Sheep', 'Hanger', 'Blackboard/Whiteboard', 'Napkin', 'Other Fish', 27 | 'Orange/Tangerine', 'Toiletry', 'Keyboard', 'Tomato', 'Lantern', 'Machinery Vehicle', 'Fan', 28 | 'Green Vegetables', 'Banana', 'Baseball Glove', 'Airplane', 'Mouse', 'Train', 'Pumpkin', 'Soccer', 'Skiboard', 29 | 'Luggage', 'Nightstand', 'Tea pot', 'Telephone', 'Trolley', 'Head Phone', 'Sports Car', 'Stop Sign', 30 | 'Dessert', 'Scooter', 'Stroller', 'Crane', 'Remote', 'Refrigerator', 'Oven', 'Lemon', 'Duck', 'Baseball Bat', 31 | 'Surveillance Camera', 'Cat', 'Jug', 'Broccoli', 'Piano', 'Pizza', 'Elephant', 'Skateboard', 'Surfboard', 32 | 'Gun', 'Skating and Skiing shoes', 'Gas stove', 'Donut', 'Bow Tie', 'Carrot', 'Toilet', 'Kite', 'Strawberry', 33 | 'Other Balls', 'Shovel', 'Pepper', 'Computer Box', 'Toilet Paper', 'Cleaning Products', 'Chopsticks', 34 | 'Microwave', 'Pigeon', 'Baseball', 'Cutting/chopping Board', 'Coffee Table', 'Side Table', 'Scissors', 35 | 'Marker', 'Pie', 'Ladder', 'Snowboard', 'Cookies', 'Radiator', 'Fire Hydrant', 'Basketball', 'Zebra', 'Grape', 36 | 'Giraffe', 'Potato', 'Sausage', 'Tricycle', 'Violin', 'Egg', 'Fire Extinguisher', 'Candy', 'Fire Truck', 37 | 'Billiards', 'Converter', 'Bathtub', 'Wheelchair', 'Golf Club', 'Briefcase', 'Cucumber', 'Cigar/Cigarette', 38 | 'Paint Brush', 'Pear', 'Heavy Truck', 'Hamburger', 'Extractor', 'Extension Cord', 'Tong', 'Tennis Racket', 39 | 'Folder', 'American Football', 'earphone', 'Mask', 'Kettle', 'Tennis', 'Ship', 'Swing', 'Coffee Machine', 40 | 'Slide', 'Carriage', 'Onion', 'Green beans', 'Projector', 'Frisbee', 'Washing Machine/Drying Machine', 41 | 'Chicken', 'Printer', 'Watermelon', 'Saxophone', 'Tissue', 'Toothbrush', 'Ice cream', 'Hot-air balloon', 42 | 'Cello', 'French Fries', 'Scale', 'Trophy', 'Cabbage', 'Hot dog', 'Blender', 'Peach', 'Rice', 'Wallet/Purse', 43 | 'Volleyball', 'Deer', 'Goose', 'Tape', 'Tablet', 'Cosmetics', 'Trumpet', 'Pineapple', 'Golf Ball', 44 | 'Ambulance', 'Parking meter', 'Mango', 'Key', 'Hurdle', 'Fishing Rod', 'Medal', 'Flute', 'Brush', 'Penguin', 45 | 'Megaphone', 'Corn', 'Lettuce', 'Garlic', 'Swan', 'Helicopter', 'Green Onion', 'Sandwich', 'Nuts', 46 | 'Speed Limit Sign', 'Induction Cooker', 'Broom', 'Trombone', 'Plum', 'Rickshaw', 'Goldfish', 'Kiwi fruit', 47 | 'Router/modem', 'Poker Card', 'Toaster', 'Shrimp', 'Sushi', 'Cheese', 'Notepaper', 'Cherry', 'Pliers', 'CD', 48 | 'Pasta', 'Hammer', 'Cue', 'Avocado', 'Hamimelon', 'Flask', 'Mushroom', 'Screwdriver', 'Soap', 'Recorder', 49 | 'Bear', 'Eggplant', 'Board Eraser', 'Coconut', 'Tape Measure/Ruler', 'Pig', 'Showerhead', 'Globe', 'Chips', 50 | 'Steak', 'Crosswalk Sign', 'Stapler', 'Camel', 'Formula 1', 'Pomegranate', 'Dishwasher', 'Crab', 51 | 'Hoverboard', 'Meat ball', 'Rice Cooker', 'Tuba', 'Calculator', 'Papaya', 'Antelope', 'Parrot', 'Seal', 52 | 'Butterfly', 'Dumbbell', 'Donkey', 'Lion', 'Urinal', 'Dolphin', 'Electric Drill', 'Hair Dryer', 'Egg tart', 53 | 'Jellyfish', 'Treadmill', 'Lighter', 'Grapefruit', 'Game board', 'Mop', 'Radish', 'Baozi', 'Target', 'French', 54 | 'Spring Rolls', 'Monkey', 'Rabbit', 'Pencil Case', 'Yak', 'Red Cabbage', 'Binoculars', 'Asparagus', 'Barbell', 55 | 'Scallop', 'Noddles', 'Comb', 'Dumpling', 'Oyster', 'Table Tennis paddle', 'Cosmetics Brush/Eyeliner Pencil', 56 | 'Chainsaw', 'Eraser', 'Lobster', 'Durian', 'Okra', 'Lipstick', 'Cosmetics Mirror', 'Curling', 'Table Tennis' ] 57 | 58 | 59 | # download command/URL (optional) -------------------------------------------------------------------------------------- 60 | download: | 61 | from pycocotools.coco import COCO 62 | from tqdm import tqdm 63 | 64 | from utils.general import download, Path 65 | 66 | # Make Directories 67 | dir = Path('../datasets/objects365') # dataset directory 68 | for p in 'images', 'labels': 69 | (dir / p).mkdir(parents=True, exist_ok=True) 70 | for q in 'train', 'val': 71 | (dir / p / q).mkdir(parents=True, exist_ok=True) 72 | 73 | # Download 74 | url = "https://dorc.ks3-cn-beijing.ksyun.com/data-set/2020Objects365%E6%95%B0%E6%8D%AE%E9%9B%86/train/" 75 | download([url + 'zhiyuan_objv2_train.tar.gz'], dir=dir, delete=False) # annotations json 76 | download([url + f for f in [f'patch{i}.tar.gz' for i in range(51)]], dir=dir / 'images' / 'train', 77 | curl=True, delete=False, threads=8) 78 | 79 | # Move 80 | train = dir / 'images' / 'train' 81 | for f in tqdm(train.rglob('*.jpg'), desc=f'Moving images'): 82 | f.rename(train / f.name) # move to /images/train 83 | 84 | # Labels 85 | coco = COCO(dir / 'zhiyuan_objv2_train.json') 86 | names = [x["name"] for x in coco.loadCats(coco.getCatIds())] 87 | for cid, cat in enumerate(names): 88 | catIds = coco.getCatIds(catNms=[cat]) 89 | imgIds = coco.getImgIds(catIds=catIds) 90 | for im in tqdm(coco.loadImgs(imgIds), desc=f'Class {cid + 1}/{len(names)} {cat}'): 91 | width, height = im["width"], im["height"] 92 | path = Path(im["file_name"]) # image filename 93 | try: 94 | with open(dir / 'labels' / 'train' / path.with_suffix('.txt').name, 'a') as file: 95 | annIds = coco.getAnnIds(imgIds=im["id"], catIds=catIds, iscrowd=None) 96 | for a in coco.loadAnns(annIds): 97 | x, y, w, h = a['bbox'] # bounding box in xywh (xy top-left corner) 98 | x, y = x + w / 2, y + h / 2 # xy to center 99 | file.write(f"{cid} {x / width:.5f} {y / height:.5f} {w / width:.5f} {h / height:.5f}\n") 100 | 101 | except Exception as e: 102 | print(e) 103 | -------------------------------------------------------------------------------- /data/scripts/get_argoverse_hd.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Argoverse-HD dataset (ring-front-center camera) http://www.cs.cmu.edu/~mengtial/proj/streaming/ 3 | # Download command: bash data/scripts/get_argoverse_hd.sh 4 | # Train command: python train.py --data argoverse_hd.yaml 5 | # Default dataset location is next to YOLOv5: 6 | # /parent_folder 7 | # /argoverse 8 | # /yolov5 9 | 10 | # Download/unzip images 11 | d='../argoverse/' # unzip directory 12 | mkdir $d 13 | url=https://argoverse-hd.s3.us-east-2.amazonaws.com/ 14 | f=Argoverse-HD-Full.zip 15 | curl -L $url$f -o $f && unzip -q $f -d $d && rm $f &# download, unzip, remove in background 16 | wait # finish background tasks 17 | 18 | cd ../argoverse/Argoverse-1.1/ 19 | ln -s tracking images 20 | 21 | cd ../Argoverse-HD/annotations/ 22 | 23 | python3 - "$@" <train.txt 84 | cat 2007_train.txt 2007_val.txt 2007_test.txt 2012_train.txt 2012_val.txt >train.all.txt 85 | 86 | mkdir ../VOC ../VOC/images ../VOC/images/train ../VOC/images/val 87 | mkdir ../VOC/labels ../VOC/labels/train ../VOC/labels/val 88 | 89 | python3 - "$@" <= 1 87 | p, s, im0, frame = path[i], '%g: ' % i, im0s[i].copy(), dataset.count 88 | else: 89 | p, s, im0, frame = path, '', im0s.copy(), getattr(dataset, 'frame', 0) 90 | 91 | p = Path(p) # to Path 92 | save_path = str(save_dir / p.name) # img.jpg 93 | txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}') # img.txt 94 | s += '%gx%g ' % img.shape[2:] # print string 95 | gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh 96 | if len(det): 97 | # Rescale boxes from img_size to im0 size 98 | det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round() 99 | 100 | # Print results 101 | for c in det[:, -1].unique(): 102 | n = (det[:, -1] == c).sum() # detections per class 103 | s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string 104 | 105 | # Write results 106 | for *xyxy, conf, cls in reversed(det): 107 | if save_txt: # Write to file 108 | xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh 109 | line = (cls, *xywh, conf) if opt.save_conf else (cls, *xywh) # label format 110 | with open(txt_path + '.txt', 'a') as f: 111 | f.write(('%g ' * len(line)).rstrip() % line + '\n') 112 | 113 | if save_img or opt.save_crop or view_img: # Add bbox to image 114 | c = int(cls) # integer class 115 | label = None if opt.hide_labels else (names[c] if opt.hide_conf else f'{names[c]} {conf:.2f}') 116 | 117 | plot_one_box(xyxy, im0, label=label, color=colors(c, True), line_thickness=opt.line_thickness) 118 | if opt.save_crop: 119 | save_one_box(xyxy, im0s, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True) 120 | 121 | # Print time (inference + NMS) 122 | print(f'{s}Done. ({t2 - t1:.3f}s)') 123 | 124 | # Stream results 125 | if view_img: 126 | cv2.imshow(str(p), im0) 127 | cv2.waitKey(1) # 1 millisecond 128 | 129 | # Save results (image with detections) 130 | if save_img: 131 | if dataset.mode == 'image': 132 | cv2.imwrite(save_path, im0) 133 | else: # 'video' or 'stream' 134 | if vid_path != save_path: # new video 135 | vid_path = save_path 136 | if isinstance(vid_writer, cv2.VideoWriter): 137 | vid_writer.release() # release previous video writer 138 | if vid_cap: # video 139 | fps = vid_cap.get(cv2.CAP_PROP_FPS) 140 | w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) 141 | h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) 142 | else: # stream 143 | fps, w, h = 30, im0.shape[1], im0.shape[0] 144 | save_path += '.mp4' 145 | vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h)) 146 | vid_writer.write(im0) 147 | 148 | if save_txt or save_img: 149 | s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' 150 | print(f"Results saved to {save_dir}{s}") 151 | 152 | print(f'Done. ({time.time() - t0:.3f}s) average inference time : {totaltime/len(dataset)} s') 153 | 154 | 155 | if __name__ == '__main__': 156 | parser = argparse.ArgumentParser() 157 | parser.add_argument('--weights', nargs='+', type=str, default='runs/train/exp20/weights/last.pt', help='model.pt path(s)') 158 | parser.add_argument('--source', type=str, default='VOC/images/test', help='source') # file/folder, 0 for webcam 159 | parser.add_argument('--img-size', type=int, default=512, help='inference size (pixels)') 160 | parser.add_argument('--conf-thres', type=float, default=0.25, help='object confidence threshold') 161 | parser.add_argument('--iou-thres', type=float, default=0.45, help='IOU threshold for NMS') 162 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 163 | parser.add_argument('--view-img', action='store_true', help='display results') 164 | parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') 165 | parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') 166 | parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes') 167 | parser.add_argument('--nosave', action='store_true', help='do not save images/videos') 168 | parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --class 0, or --class 0 2 3') 169 | parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') 170 | parser.add_argument('--augment', action='store_true', help='augmented inference') 171 | parser.add_argument('--update', action='store_true', help='update all models') 172 | parser.add_argument('--project', default='runs/detect', help='save results to project/name') 173 | parser.add_argument('--name', default='exp', help='save results to project/name') 174 | parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') 175 | parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)') 176 | parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels') 177 | parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences') 178 | opt = parser.parse_args() 179 | print(opt) 180 | check_requirements(exclude=('tensorboard', 'pycocotools', 'thop')) 181 | 182 | with torch.no_grad(): 183 | if opt.update: # update all models (to fix SourceChangeWarning) 184 | for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']: 185 | detect(opt=opt) 186 | strip_optimizer(opt.weights) 187 | else: 188 | detect(opt=opt) 189 | -------------------------------------------------------------------------------- /detect_prune.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import time 3 | from pathlib import Path 4 | 5 | import cv2 6 | import torch 7 | import torch.backends.cudnn as cudnn 8 | from numpy import random 9 | 10 | from models.experimental import attempt_load 11 | from utils.datasets import LoadStreams, LoadImages 12 | from utils.general import check_img_size, check_requirements, check_imshow, non_max_suppression, apply_classifier, \ 13 | scale_coords, xyxy2xywh, strip_optimizer, set_logging, increment_path, save_one_box 14 | from utils.plots import colors, plot_one_box 15 | from utils.torch_utils import select_device, load_classifier, time_synchronized 16 | import time 17 | 18 | def detect(opt): 19 | source, weights, view_img, save_txt, imgsz = opt.source, opt.weights, opt.view_img, opt.save_txt, opt.img_size 20 | save_img = not opt.nosave and not source.endswith('.txt') # save inference images 21 | webcam = source.isnumeric() or source.endswith('.txt') or source.lower().startswith( 22 | ('rtsp://', 'rtmp://', 'http://', 'https://')) 23 | 24 | # Directories 25 | save_dir = increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok) # increment run 26 | (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir 27 | 28 | # Initialize 29 | set_logging() 30 | device = select_device(opt.device) 31 | half = device.type != 'cpu' # half precision only supported on CUDA 32 | 33 | # Load model 34 | model = attempt_load(weights, map_location=device) # load FP32 model 35 | stride = int(model.stride.max()) # model stride 36 | imgsz = check_img_size(imgsz, s=stride) # check img_size 37 | names = model.module.names if hasattr(model, 'module') else model.names # get class names 38 | if half: 39 | model.half() # to FP16 40 | 41 | # Second-stage classifier 42 | classify = False 43 | if classify: 44 | modelc = load_classifier(name='resnet101', n=2) # initialize 45 | modelc.load_state_dict(torch.load('weights/resnet101.pt', map_location=device)['model']).to(device).eval() 46 | 47 | # Set Dataloader 48 | vid_path, vid_writer = None, None 49 | if webcam: 50 | view_img = check_imshow() 51 | cudnn.benchmark = True # set True to speed up constant image size inference 52 | dataset = LoadStreams(source, img_size=imgsz, stride=stride) 53 | else: 54 | dataset = LoadImages(source, img_size=imgsz, stride=stride) 55 | 56 | # Run inference 57 | if device.type != 'cpu': 58 | model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters()))) # run once 59 | t0 = time.time() 60 | totaltime = 0 61 | for path, img, im0s, vid_cap in dataset: 62 | img = torch.from_numpy(img).to(device) 63 | img = img.half() if half else img.float() # uint8 to fp16/32 64 | img /= 255.0 # 0 - 255 to 0.0 - 1.0 65 | if img.ndimension() == 3: 66 | img = img.unsqueeze(0) 67 | 68 | # Inference 69 | t1 = time_synchronized() 70 | start = time.time() 71 | pred = model(img, augment=opt.augment)[0] 72 | end = time.time() 73 | infertime = end - start 74 | totaltime += infertime 75 | # Apply NMS 76 | pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=opt.agnostic_nms) 77 | t2 = time_synchronized() 78 | 79 | # Apply Classifier 80 | if classify: 81 | pred = apply_classifier(pred, modelc, img, im0s) 82 | 83 | # Process detections 84 | for i, det in enumerate(pred): # detections per image 85 | if webcam: # batch_size >= 1 86 | p, s, im0, frame = path[i], '%g: ' % i, im0s[i].copy(), dataset.count 87 | else: 88 | p, s, im0, frame = path, '', im0s.copy(), getattr(dataset, 'frame', 0) 89 | 90 | p = Path(p) # to Path 91 | save_path = str(save_dir / p.name) # img.jpg 92 | txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}') # img.txt 93 | s += '%gx%g ' % img.shape[2:] # print string 94 | gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh 95 | if len(det): 96 | # Rescale boxes from img_size to im0 size 97 | det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round() 98 | 99 | # Print results 100 | for c in det[:, -1].unique(): 101 | n = (det[:, -1] == c).sum() # detections per class 102 | s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string 103 | 104 | # Write results 105 | for *xyxy, conf, cls in reversed(det): 106 | if save_txt: # Write to file 107 | xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh 108 | line = (cls, *xywh, conf) if opt.save_conf else (cls, *xywh) # label format 109 | with open(txt_path + '.txt', 'a') as f: 110 | f.write(('%g ' * len(line)).rstrip() % line + '\n') 111 | 112 | if save_img or opt.save_crop or view_img: # Add bbox to image 113 | c = int(cls) # integer class 114 | label = None if opt.hide_labels else (names[c] if opt.hide_conf else f'{names[c]} {conf:.2f}') 115 | 116 | plot_one_box(xyxy, im0, label=label, color=colors(c, True), line_thickness=opt.line_thickness) 117 | if opt.save_crop: 118 | save_one_box(xyxy, im0s, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True) 119 | 120 | # Print time (inference + NMS) 121 | print(f'{s}Done. ({t2 - t1:.3f}s)') 122 | 123 | # Stream results 124 | if view_img: 125 | cv2.imshow(str(p), im0) 126 | cv2.waitKey(1) # 1 millisecond 127 | 128 | # Save results (image with detections) 129 | if save_img: 130 | if dataset.mode == 'image': 131 | cv2.imwrite(save_path, im0) 132 | else: # 'video' or 'stream' 133 | if vid_path != save_path: # new video 134 | vid_path = save_path 135 | if isinstance(vid_writer, cv2.VideoWriter): 136 | vid_writer.release() # release previous video writer 137 | if vid_cap: # video 138 | fps = vid_cap.get(cv2.CAP_PROP_FPS) 139 | w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) 140 | h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) 141 | else: # stream 142 | fps, w, h = 30, im0.shape[1], im0.shape[0] 143 | save_path += '.mp4' 144 | vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h)) 145 | vid_writer.write(im0) 146 | 147 | if save_txt or save_img: 148 | s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' 149 | print(f"Results saved to {save_dir}{s}") 150 | 151 | print(f'Done. ({time.time() - t0:.3f}s) average inference time : {totaltime/len(dataset)} s') 152 | 153 | 154 | if __name__ == '__main__': 155 | parser = argparse.ArgumentParser() 156 | parser.add_argument('--weights', nargs='+', type=str, default='runs/train/exp23/weights/last.pt', help='model.pt path(s)') 157 | parser.add_argument('--source', type=str, default='VOC/images/test', help='source') # file/folder, 0 for webcam 158 | parser.add_argument('--img-size', type=int, default=512, help='inference size (pixels)') 159 | parser.add_argument('--conf-thres', type=float, default=0.25, help='object confidence threshold') 160 | parser.add_argument('--iou-thres', type=float, default=0.45, help='IOU threshold for NMS') 161 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 162 | parser.add_argument('--view-img', action='store_true', help='display results') 163 | parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') 164 | parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') 165 | parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes') 166 | parser.add_argument('--nosave', action='store_true', help='do not save images/videos') 167 | parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --class 0, or --class 0 2 3') 168 | parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') 169 | parser.add_argument('--augment', action='store_true', help='augmented inference') 170 | parser.add_argument('--update', action='store_true', help='update all models') 171 | parser.add_argument('--project', default='runs/detect', help='save results to project/name') 172 | parser.add_argument('--name', default='exp', help='save results to project/name') 173 | parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') 174 | parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)') 175 | parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels') 176 | parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences') 177 | opt = parser.parse_args() 178 | print(opt) 179 | check_requirements(exclude=('tensorboard', 'pycocotools', 'thop')) 180 | 181 | with torch.no_grad(): 182 | if opt.update: # update all models (to fix SourceChangeWarning) 183 | for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']: 184 | detect(opt=opt) 185 | strip_optimizer(opt.weights) 186 | else: 187 | detect(opt=opt) 188 | -------------------------------------------------------------------------------- /getweight.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | with open("map.txt","r") as f: 4 | lines = f.readlines() 5 | 6 | 7 | w = [] 8 | 9 | for line in lines: 10 | w.append(line.split()[-2]) 11 | 12 | w = [float(i) for i in w] 13 | print(w) 14 | -------------------------------------------------------------------------------- /hubconf.py: -------------------------------------------------------------------------------- 1 | """YOLOv5 PyTorch Hub models https://pytorch.org/hub/ultralytics_yolov5/ 2 | 3 | Usage: 4 | import torch 5 | model = torch.hub.load('ultralytics/yolov5', 'yolov5s') 6 | """ 7 | 8 | from pathlib import Path 9 | 10 | import torch 11 | 12 | from models.yolo import Model, attempt_load 13 | from utils.general import check_requirements, set_logging 14 | from utils.google_utils import attempt_download 15 | from utils.torch_utils import select_device 16 | 17 | dependencies = ['torch', 'yaml'] 18 | check_requirements(Path(__file__).parent / 'requirements.txt', exclude=('tensorboard', 'pycocotools', 'thop')) 19 | 20 | 21 | def create(name, pretrained=True, channels=3, classes=80, autoshape=True, verbose=True): 22 | """Creates a specified YOLOv5 model 23 | 24 | Arguments: 25 | name (str): name of model, i.e. 'yolov5s' 26 | pretrained (bool): load pretrained weights into the model 27 | channels (int): number of input channels 28 | classes (int): number of model classes 29 | autoshape (bool): apply YOLOv5 .autoshape() wrapper to model 30 | verbose (bool): print all information to screen 31 | 32 | Returns: 33 | YOLOv5 pytorch model 34 | """ 35 | set_logging(verbose=verbose) 36 | fname = Path(name).with_suffix('.pt') # checkpoint filename 37 | try: 38 | if pretrained and channels == 3 and classes == 80: 39 | model = attempt_load(fname, map_location=torch.device('cpu')) # download/load FP32 model 40 | else: 41 | cfg = list((Path(__file__).parent / 'models').rglob(f'{name}.yaml'))[0] # model.yaml path 42 | model = Model(cfg, channels, classes) # create model 43 | if pretrained: 44 | attempt_download(fname) # download if not found locally 45 | ckpt = torch.load(fname, map_location=torch.device('cpu')) # load 46 | msd = model.state_dict() # model state_dict 47 | csd = ckpt['model'].float().state_dict() # checkpoint state_dict as FP32 48 | csd = {k: v for k, v in csd.items() if msd[k].shape == v.shape} # filter 49 | model.load_state_dict(csd, strict=False) # load 50 | if len(ckpt['model'].names) == classes: 51 | model.names = ckpt['model'].names # set class names attribute 52 | if autoshape: 53 | model = model.autoshape() # for file/URI/PIL/cv2/np inputs and NMS 54 | device = select_device('0' if torch.cuda.is_available() else 'cpu') # default to GPU if available 55 | return model.to(device) 56 | 57 | except Exception as e: 58 | help_url = 'https://github.com/ultralytics/yolov5/issues/36' 59 | s = 'Cache may be out of date, try `force_reload=True`. See %s for help.' % help_url 60 | raise Exception(s) from e 61 | 62 | 63 | def custom(path='path/to/model.pt', autoshape=True, verbose=True): 64 | # YOLOv5 custom or local model 65 | return create(path, autoshape=autoshape, verbose=verbose) 66 | 67 | 68 | def yolov5s(pretrained=True, channels=3, classes=80, autoshape=True, verbose=True): 69 | # YOLOv5-small model https://github.com/ultralytics/yolov5 70 | return create('yolov5s', pretrained, channels, classes, autoshape, verbose) 71 | 72 | 73 | def yolov5m(pretrained=True, channels=3, classes=80, autoshape=True, verbose=True): 74 | # YOLOv5-medium model https://github.com/ultralytics/yolov5 75 | return create('yolov5m', pretrained, channels, classes, autoshape, verbose) 76 | 77 | 78 | def yolov5l(pretrained=True, channels=3, classes=80, autoshape=True, verbose=True): 79 | # YOLOv5-large model https://github.com/ultralytics/yolov5 80 | return create('yolov5l', pretrained, channels, classes, autoshape, verbose) 81 | 82 | 83 | def yolov5x(pretrained=True, channels=3, classes=80, autoshape=True, verbose=True): 84 | # YOLOv5-xlarge model https://github.com/ultralytics/yolov5 85 | return create('yolov5x', pretrained, channels, classes, autoshape, verbose) 86 | 87 | 88 | def yolov5s6(pretrained=True, channels=3, classes=80, autoshape=True, verbose=True): 89 | # YOLOv5-small-P6 model https://github.com/ultralytics/yolov5 90 | return create('yolov5s6', pretrained, channels, classes, autoshape, verbose) 91 | 92 | 93 | def yolov5m6(pretrained=True, channels=3, classes=80, autoshape=True, verbose=True): 94 | # YOLOv5-medium-P6 model https://github.com/ultralytics/yolov5 95 | return create('yolov5m6', pretrained, channels, classes, autoshape, verbose) 96 | 97 | 98 | def yolov5l6(pretrained=True, channels=3, classes=80, autoshape=True, verbose=True): 99 | # YOLOv5-large-P6 model https://github.com/ultralytics/yolov5 100 | return create('yolov5l6', pretrained, channels, classes, autoshape, verbose) 101 | 102 | 103 | def yolov5x6(pretrained=True, channels=3, classes=80, autoshape=True, verbose=True): 104 | # YOLOv5-xlarge-P6 model https://github.com/ultralytics/yolov5 105 | return create('yolov5x6', pretrained, channels, classes, autoshape, verbose) 106 | 107 | 108 | if __name__ == '__main__': 109 | model = create(name='yolov5s', pretrained=True, channels=3, classes=80, autoshape=True, verbose=True) # pretrained 110 | # model = custom(path='path/to/model.pt') # custom 111 | 112 | # Verify inference 113 | import cv2 114 | import numpy as np 115 | from PIL import Image 116 | 117 | imgs = ['data/images/zidane.jpg', # filename 118 | 'https://github.com/ultralytics/yolov5/releases/download/v1.0/zidane.jpg', # URI 119 | cv2.imread('data/images/bus.jpg')[:, :, ::-1], # OpenCV 120 | Image.open('data/images/bus.jpg'), # PIL 121 | np.zeros((320, 640, 3))] # numpy 122 | 123 | results = model(imgs) # batched inference 124 | results.print() 125 | results.save() 126 | -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-23 20-19-08.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-23 20-19-08.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-23 20-19-30.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-23 20-19-30.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-24 22-17-16.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-24 22-17-16.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-25 00-26-23.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-25 00-26-23.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-25 00-26-45.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-25 00-26-45.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-25 00-28-15.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-25 00-28-15.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-25 00-28-52.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-25 00-28-52.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-27 22-20-33.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-27 22-20-33.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-28 08-33-25.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-28 08-33-25.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-31 22-29-12.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-31 22-29-12.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-05-31 22-30-21.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-05-31 22-30-21.png -------------------------------------------------------------------------------- /img/Screenshot from 2021-06-05 00-06-27.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Screenshot from 2021-06-05 00-06-27.png -------------------------------------------------------------------------------- /img/Selection_007.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/img/Selection_007.png -------------------------------------------------------------------------------- /map.txt: -------------------------------------------------------------------------------- 1 | person 999 3500 0.76 0.648 0.698 0.345 2 | chair 999 464 0.601 0.364 0.411 0.183 3 | vehicle 999 2153 0.765 0.786 0.815 0.557 4 | motorbike 999 826 0.624 0.558 0.579 0.292 5 | tablet_ad 999 1311 0.653 0.712 0.714 0.41 6 | trafficsign 999 164 0.568 0.457 0.476 0.293 7 | table 999 96 0.389 0.331 0.207 0.0921 8 | normal_tree 999 678 0.61 0.451 0.473 0.162 9 | rect_stall 999 111 0.543 0.225 0.251 0.0976 10 | tricycle 999 105 0.655 0.343 0.389 0.18 11 | bucket 999 228 0.577 0.487 0.481 0.261 12 | truck 999 89 0.503 0.461 0.471 0.295 13 | light_body 999 128 0.498 0.0776 0.14 0.0418 14 | cluster 999 177 0.34 0.0734 0.092 0.0418 15 | billboard_ad 999 261 0.411 0.264 0.237 0.125 16 | umbrella 999 186 0.804 0.747 0.781 0.46 17 | pedestrain_pile 999 178 0.684 0.292 0.324 0.153 18 | traffic_rail 999 180 0.462 0.434 0.374 0.157 19 | flat_stall 999 323 0.557 0.211 0.269 0.101 20 | vendor_business 999 229 0.59 0.393 0.383 0.128 21 | clothes 999 102 0.411 0.353 0.32 0.126 22 | basket_stall 999 140 0.333 0.243 0.216 0.103 23 | 24 | 25 | 26 | 27 | 28 | Class Images Labels P R mAP@.5 mAP@.5:.95: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 32/32 [00:10<00:00, 2.93it/s] 29 | all 999 11629 0.549 0.392 0.408 0.208 30 | person 999 3500 0.735 0.637 0.687 0.338 31 | chair 999 464 0.615 0.315 0.382 0.18 32 | vehicle 999 2153 0.752 0.787 0.828 0.563 33 | motorbike 999 826 0.635 0.567 0.585 0.293 34 | tablet_ad 999 1311 0.622 0.715 0.71 0.41 35 | trafficsign 999 164 0.549 0.5 0.476 0.295 36 | table 999 96 0.515 0.276 0.267 0.134 37 | normal_tree 999 678 0.574 0.441 0.455 0.165 38 | rect_stall 999 111 0.444 0.261 0.221 0.077 39 | tricycle 999 105 0.586 0.305 0.349 0.167 40 | bucket 999 228 0.618 0.421 0.459 0.248 41 | truck 999 89 0.535 0.551 0.523 0.327 42 | light_body 999 128 0.513 0.0495 0.154 0.0455 43 | cluster 999 177 0.267 0.0621 0.0681 0.0309 44 | billboard_ad 999 261 0.479 0.276 0.266 0.147 45 | umbrella 999 186 0.766 0.737 0.78 0.454 46 | pedestrain_pile 999 178 0.613 0.275 0.333 0.157 47 | traffic_rail 999 180 0.437 0.417 0.348 0.14 48 | flat_stall 999 323 0.563 0.232 0.286 0.0965 49 | vendor_business 999 229 0.521 0.31 0.318 0.114 50 | clothes 999 102 0.442 0.333 0.3 0.108 51 | basket_stall 999 140 0.304 0.157 0.173 0.0799 52 | 53 | 54 | 55 | Class Images Labels P R mAP@.5 mAP@.5:.95: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 250/250 [00:13<00:00, 18.34it/s] 56 | all 999 11629 0.539 0.392 0.401 0.201 57 | person 999 3500 0.743 0.635 0.685 0.339 58 | chair 999 464 0.547 0.351 0.383 0.176 59 | vehicle 999 2153 0.756 0.795 0.824 0.561 60 | motorbike 999 826 0.626 0.541 0.58 0.295 61 | tablet_ad 999 1311 0.619 0.732 0.702 0.406 62 | trafficsign 999 164 0.552 0.488 0.474 0.276 63 | table 999 96 0.302 0.323 0.221 0.103 64 | normal_tree 999 678 0.622 0.42 0.462 0.15 65 | rect_stall 999 111 0.406 0.148 0.171 0.0613 66 | tricycle 999 105 0.607 0.309 0.365 0.176 67 | bucket 999 228 0.666 0.464 0.471 0.246 68 | truck 999 89 0.476 0.506 0.509 0.324 69 | light_body 999 128 0.434 0.0859 0.121 0.0331 70 | cluster 999 177 0.305 0.0847 0.079 0.0277 71 | billboard_ad 999 261 0.382 0.199 0.212 0.109 72 | umbrella 999 186 0.765 0.726 0.757 0.432 73 | pedestrain_pile 999 178 0.708 0.287 0.328 0.152 74 | traffic_rail 999 180 0.396 0.433 0.345 0.153 75 | flat_stall 999 323 0.544 0.207 0.256 0.0846 76 | vendor_business 999 229 0.575 0.301 0.324 0.0996 77 | clothes 999 102 0.44 0.392 0.364 0.142 78 | basket_stall 999 140 0.382 0.186 0.18 0.0757 79 | 80 | 81 | 82 | 83 | 84 | 85 | -------------------------------------------------------------------------------- /model_change.txt: -------------------------------------------------------------------------------- 1 | model.model[0].conv.conv = Conv2d(12, 31, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) 2 | model.model[0].conv.bn = BatchNorm2d(31, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 3 | model.model[1].conv = Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) 4 | model.model[1].bn = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 5 | model.model[2].cv1.conv = Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False) 6 | model.model[2].cv1.bn = BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 7 | model.model[2].cv2.conv = Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False) 8 | model.model[2].cv2.bn = BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 9 | model.model[2].cv3.conv = Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) 10 | model.model[2].cv3.bn = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 11 | model.model[2].m[0].cv1.conv = Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1), bias=False) 12 | model.model[2].m[0].cv1.bn = BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 13 | model.model[3].conv = Conv2d(64, 127, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) 14 | model.model[3].bn = BatchNorm2d(127, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 15 | model.model[4].cv1.conv = Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) 16 | model.model[4].cv1.bn = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 17 | model.model[4].cv2.conv = Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) 18 | model.model[4].cv2.bn = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 19 | model.model[4].cv3.conv = Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False) 20 | model.model[4].cv3.bn = BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 21 | model.model[4].m[0].cv1.conv = Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) 22 | model.model[4].m[0].cv1.bn = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 23 | model.model[4].m[1].cv1.conv = Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) 24 | model.model[4].m[1].cv1.bn = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 25 | model.model[4].m[2].cv1.conv = Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) 26 | model.model[4].m[2].cv1.bn = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 27 | model.model[5].conv = Conv2d(128, 225, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) 28 | model.model[5].bn = BatchNorm2d(225, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 29 | model.model[6].cv1.conv = Conv2d(256, 110, kernel_size=(1, 1), stride=(1, 1), bias=False) 30 | model.model[6].cv1.bn = BatchNorm2d(110, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 31 | model.model[6].cv2.conv = Conv2d(256, 90, kernel_size=(1, 1), stride=(1, 1), bias=False) 32 | model.model[6].cv2.bn = BatchNorm2d(90, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 33 | model.model[6].cv3.conv = Conv2d(256, 195, kernel_size=(1, 1), stride=(1, 1), bias=False) 34 | model.model[6].cv3.bn = BatchNorm2d(195, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 35 | model.model[6].m[0].cv1.conv = Conv2d(128, 102, kernel_size=(1, 1), stride=(1, 1), bias=False) 36 | model.model[6].m[0].cv1.bn = BatchNorm2d(102, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 37 | model.model[6].m[1].cv1.conv = Conv2d(128, 116, kernel_size=(1, 1), stride=(1, 1), bias=False) 38 | model.model[6].m[1].cv1.bn = BatchNorm2d(116, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 39 | model.model[6].m[2].cv1.conv = Conv2d(128, 106, kernel_size=(1, 1), stride=(1, 1), bias=False) 40 | model.model[6].m[2].cv1.bn = BatchNorm2d(106, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 41 | model.model[7].conv = Conv2d(256, 127, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) 42 | model.model[7].bn = BatchNorm2d(127, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 43 | model.model[8].cv1.conv = Conv2d(512, 118, kernel_size=(1, 1), stride=(1, 1), bias=False) 44 | model.model[8].cv1.bn = BatchNorm2d(118, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 45 | model.model[8].cv2.conv = Conv2d(1024, 53, kernel_size=(1, 1), stride=(1, 1), bias=False) 46 | model.model[8].cv2.bn = BatchNorm2d(53, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 47 | model.model[9].cv1.conv = Conv2d(512, 14, kernel_size=(1, 1), stride=(1, 1), bias=False) 48 | model.model[9].cv1.bn = BatchNorm2d(14, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 49 | model.model[9].cv2.conv = Conv2d(512, 27, kernel_size=(1, 1), stride=(1, 1), bias=False) 50 | model.model[9].cv2.bn = BatchNorm2d(27, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 51 | model.model[9].cv3.conv = Conv2d(512, 34, kernel_size=(1, 1), stride=(1, 1), bias=False) 52 | model.model[9].cv3.bn = BatchNorm2d(34, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 53 | model.model[9].m[0].cv1.conv = Conv2d(256, 18, kernel_size=(1, 1), stride=(1, 1), bias=False) 54 | model.model[9].m[0].cv1.bn = BatchNorm2d(18, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 55 | model.model[9].m[0].cv2.conv = Conv2d(256, 26, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) 56 | model.model[9].m[0].cv2.bn = BatchNorm2d(26, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 57 | model.model[10].conv = Conv2d(512, 40, kernel_size=(1, 1), stride=(1, 1), bias=False) 58 | model.model[10].bn = BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 59 | model.model[13].cv1.conv = Conv2d(512, 103, kernel_size=(1, 1), stride=(1, 1), bias=False) 60 | model.model[13].cv1.bn = BatchNorm2d(103, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 61 | model.model[13].cv2.conv = Conv2d(512, 63, kernel_size=(1, 1), stride=(1, 1), bias=False) 62 | model.model[13].cv2.bn = BatchNorm2d(63, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 63 | model.model[13].cv3.conv = Conv2d(256, 131, kernel_size=(1, 1), stride=(1, 1), bias=False) 64 | model.model[13].cv3.bn = BatchNorm2d(131, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 65 | model.model[13].m[0].cv1.conv = Conv2d(128, 100, kernel_size=(1, 1), stride=(1, 1), bias=False) 66 | model.model[13].m[0].cv1.bn = BatchNorm2d(100, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 67 | model.model[13].m[0].cv2.conv = Conv2d(128, 112, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) 68 | model.model[13].m[0].cv2.bn = BatchNorm2d(112, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 69 | model.model[14].conv = Conv2d(256, 93, kernel_size=(1, 1), stride=(1, 1), bias=False) 70 | model.model[14].bn = BatchNorm2d(93, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 71 | model.model[17].cv1.conv = Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False) 72 | model.model[17].cv1.bn = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 73 | model.model[17].cv2.conv = Conv2d(256, 44, kernel_size=(1, 1), stride=(1, 1), bias=False) 74 | model.model[17].cv2.bn = BatchNorm2d(44, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 75 | model.model[17].cv3.conv = Conv2d(128, 105, kernel_size=(1, 1), stride=(1, 1), bias=False) 76 | model.model[17].cv3.bn = BatchNorm2d(105, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 77 | model.model[17].m[0].cv1.conv = Conv2d(64, 58, kernel_size=(1, 1), stride=(1, 1), bias=False) 78 | model.model[17].m[0].cv1.bn = BatchNorm2d(58, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 79 | model.model[17].m[0].cv2.conv = Conv2d(64, 59, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) 80 | model.model[17].m[0].cv2.bn = BatchNorm2d(59, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 81 | model.model[18].conv = Conv2d(128, 56, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) 82 | model.model[18].bn = BatchNorm2d(56, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 83 | model.model[20].cv1.conv = Conv2d(256, 69, kernel_size=(1, 1), stride=(1, 1), bias=False) 84 | model.model[20].cv1.bn = BatchNorm2d(69, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 85 | model.model[20].cv2.conv = Conv2d(256, 47, kernel_size=(1, 1), stride=(1, 1), bias=False) 86 | model.model[20].cv2.bn = BatchNorm2d(47, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 87 | model.model[20].cv3.conv = Conv2d(256, 160, kernel_size=(1, 1), stride=(1, 1), bias=False) 88 | model.model[20].cv3.bn = BatchNorm2d(160, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 89 | model.model[20].m[0].cv1.conv = Conv2d(128, 69, kernel_size=(1, 1), stride=(1, 1), bias=False) 90 | model.model[20].m[0].cv1.bn = BatchNorm2d(69, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 91 | model.model[20].m[0].cv2.conv = Conv2d(128, 87, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) 92 | model.model[20].m[0].cv2.bn = BatchNorm2d(87, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 93 | model.model[21].conv = Conv2d(256, 88, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) 94 | model.model[21].bn = BatchNorm2d(88, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 95 | model.model[23].cv1.conv = Conv2d(512, 40, kernel_size=(1, 1), stride=(1, 1), bias=False) 96 | model.model[23].cv1.bn = BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 97 | model.model[23].cv2.conv = Conv2d(512, 53, kernel_size=(1, 1), stride=(1, 1), bias=False) 98 | model.model[23].cv2.bn = BatchNorm2d(53, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 99 | model.model[23].cv3.conv = Conv2d(512, 146, kernel_size=(1, 1), stride=(1, 1), bias=False) 100 | model.model[23].cv3.bn = BatchNorm2d(146, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 101 | model.model[23].m[0].cv1.conv = Conv2d(256, 35, kernel_size=(1, 1), stride=(1, 1), bias=False) 102 | model.model[23].m[0].cv1.bn = BatchNorm2d(35, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 103 | model.model[23].m[0].cv2.conv = Conv2d(256, 48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) 104 | model.model[23].m[0].cv2.bn = BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) 105 | -------------------------------------------------------------------------------- /modelparse.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import time 3 | from pathlib import Path 4 | 5 | import cv2 6 | import torch 7 | import torch.backends.cudnn as cudnn 8 | from numpy import random 9 | from torch.utils.tensorboard import SummaryWriter 10 | from models.experimental import attempt_load 11 | from utils.datasets import LoadStreams, LoadImages 12 | from utils.general import check_img_size, check_requirements, check_imshow, non_max_suppression, apply_classifier, \ 13 | scale_coords, xyxy2xywh, strip_optimizer, set_logging, increment_path, save_one_box 14 | from utils.plots import colors, plot_one_box 15 | from utils.torch_utils import select_device, load_classifier, time_synchronized 16 | import torch.nn as nn 17 | 18 | tb_writer = SummaryWriter() 19 | 20 | def detect(opt): 21 | source, weights, view_img, save_txt, imgsz = opt.source, opt.weights, opt.view_img, opt.save_txt, opt.img_size 22 | save_img = not opt.nosave and not source.endswith('.txt') # save inference images 23 | webcam = source.isnumeric() or source.endswith('.txt') or source.lower().startswith( 24 | ('rtsp://', 'rtmp://', 'http://', 'https://')) 25 | 26 | # Directories 27 | save_dir = increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok) # increment run 28 | (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir 29 | 30 | # Initialize 31 | set_logging() 32 | device = select_device(opt.device) 33 | half = device.type != 'cpu' # half precision only supported on CUDA 34 | 35 | # Load model 36 | print("weights:",weights) 37 | model = attempt_load(weights, map_location=device) # load FP32 model 38 | stride = int(model.stride.max()) # model stride 39 | imgsz = check_img_size(imgsz, s=stride) # check img_size 40 | names = model.module.names if hasattr(model, 'module') else model.names # get class names 41 | if half: 42 | model.half() # to FP16 43 | 44 | # Second-stage classifier 45 | classify = False 46 | if classify: 47 | modelc = load_classifier(name='resnet101', n=2) # initialize 48 | modelc.load_state_dict(torch.load('weights/resnet101.pt', map_location=device)['model']).to(device).eval() 49 | 50 | # Set Dataloader 51 | vid_path, vid_writer = None, None 52 | if webcam: 53 | view_img = check_imshow() 54 | cudnn.benchmark = True # set True to speed up constant image size inference 55 | dataset = LoadStreams(source, img_size=imgsz, stride=stride) 56 | else: 57 | dataset = LoadImages(source, img_size=imgsz, stride=stride) 58 | 59 | # Run inference 60 | if device.type != 'cpu': 61 | model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters()))) # run once 62 | t0 = time.time() 63 | # =============== show bn weights ===================== # 64 | module_list = [] 65 | for i,layer in model.named_modules(): 66 | if isinstance(layer,nn.BatchNorm2d): 67 | bnw = layer.state_dict()['weight'] 68 | module_list.append(bnw) 69 | # bnw = bnw.sort() 70 | # print(f"{i} : {bnw} : ") 71 | size_list = [idx.data.shape[0] for idx in module_list] 72 | 73 | bn_weights = torch.zeros(sum(size_list)) 74 | index = 0 75 | for idx, size in enumerate(size_list): 76 | bn_weights[index:(index + size)] = module_list[idx].data.abs().clone() 77 | index += size 78 | 79 | print("bn_weights:",bn_weights.sort()) 80 | tb_writer.add_histogram('bn_weights/hist', bn_weights.numpy(), 1, bins='doane') 81 | 82 | for path, img, im0s, vid_cap in dataset: 83 | img = torch.from_numpy(img).to(device) 84 | img = img.half() if half else img.float() # uint8 to fp16/32 85 | img /= 255.0 # 0 - 255 to 0.0 - 1.0 86 | if img.ndimension() == 3: 87 | img = img.unsqueeze(0) 88 | 89 | # Inference 90 | t1 = time_synchronized() 91 | pred = model(img, augment=opt.augment)[0] 92 | 93 | # Apply NMS 94 | pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=opt.agnostic_nms) 95 | t2 = time_synchronized() 96 | 97 | # Apply Classifier 98 | if classify: 99 | pred = apply_classifier(pred, modelc, img, im0s) 100 | 101 | # Process detections 102 | for i, det in enumerate(pred): # detections per image 103 | if webcam: # batch_size >= 1 104 | p, s, im0, frame = path[i], '%g: ' % i, im0s[i].copy(), dataset.count 105 | else: 106 | p, s, im0, frame = path, '', im0s.copy(), getattr(dataset, 'frame', 0) 107 | 108 | p = Path(p) # to Path 109 | save_path = str(save_dir / p.name) # img.jpg 110 | txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}') # img.txt 111 | s += '%gx%g ' % img.shape[2:] # print string 112 | gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh 113 | if len(det): 114 | # Rescale boxes from img_size to im0 size 115 | det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round() 116 | 117 | # Print results 118 | for c in det[:, -1].unique(): 119 | n = (det[:, -1] == c).sum() # detections per class 120 | s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string 121 | 122 | # Write results 123 | for *xyxy, conf, cls in reversed(det): 124 | if save_txt: # Write to file 125 | xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh 126 | line = (cls, *xywh, conf) if opt.save_conf else (cls, *xywh) # label format 127 | with open(txt_path + '.txt', 'a') as f: 128 | f.write(('%g ' * len(line)).rstrip() % line + '\n') 129 | 130 | if save_img or opt.save_crop or view_img: # Add bbox to image 131 | c = int(cls) # integer class 132 | label = None if opt.hide_labels else (names[c] if opt.hide_conf else f'{names[c]} {conf:.2f}') 133 | 134 | plot_one_box(xyxy, im0, label=label, color=colors(c, True), line_thickness=opt.line_thickness) 135 | if opt.save_crop: 136 | save_one_box(xyxy, im0s, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True) 137 | 138 | # Print time (inference + NMS) 139 | print(f'{s}Done. ({t2 - t1:.3f}s)') 140 | 141 | # Stream results 142 | if view_img: 143 | cv2.imshow(str(p), im0) 144 | cv2.waitKey(1) # 1 millisecond 145 | 146 | # Save results (image with detections) 147 | if save_img: 148 | if dataset.mode == 'image': 149 | cv2.imwrite(save_path, im0) 150 | else: # 'video' or 'stream' 151 | if vid_path != save_path: # new video 152 | vid_path = save_path 153 | if isinstance(vid_writer, cv2.VideoWriter): 154 | vid_writer.release() # release previous video writer 155 | if vid_cap: # video 156 | fps = vid_cap.get(cv2.CAP_PROP_FPS) 157 | w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) 158 | h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) 159 | else: # stream 160 | fps, w, h = 30, im0.shape[1], im0.shape[0] 161 | save_path += '.mp4' 162 | vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h)) 163 | vid_writer.write(im0) 164 | 165 | if save_txt or save_img: 166 | s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' 167 | print(f"Results saved to {save_dir}{s}") 168 | 169 | print(f'Done. ({time.time() - t0:.3f}s)') 170 | 171 | 172 | if __name__ == '__main__': 173 | parser = argparse.ArgumentParser() 174 | parser.add_argument('--weights', nargs='+', type=str, default='/home/kong/yolov5/runs/train/exp78/weights/last.pt', help='model.pt path(s)') 175 | parser.add_argument('--source', type=str, default='/home/kong/yolov5/data/test', help='source') # file/folder, 0 for webcam 176 | parser.add_argument('--img-size', type=int, default=320, help='inference size (pixels)') 177 | parser.add_argument('--conf-thres', type=float, default=0.25, help='object confidence threshold') 178 | parser.add_argument('--iou-thres', type=float, default=0.45, help='IOU threshold for NMS') 179 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 180 | parser.add_argument('--view-img', action='store_true', help='display results') 181 | parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') 182 | parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') 183 | parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes') 184 | parser.add_argument('--nosave', action='store_true', help='do not save images/videos') 185 | parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --class 0, or --class 0 2 3') 186 | parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') 187 | parser.add_argument('--augment', action='store_true', help='augmented inference') 188 | parser.add_argument('--update', action='store_true', help='update all models') 189 | parser.add_argument('--project', default='runs/detect', help='save results to project/name') 190 | parser.add_argument('--name', default='exp', help='save results to project/name') 191 | parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') 192 | parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)') 193 | parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels') 194 | parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences') 195 | opt = parser.parse_args() 196 | print(opt) 197 | # check_requirements(exclude=('tensorboard', 'pycocotools', 'thop')) 198 | 199 | with torch.no_grad(): 200 | if opt.update: # update all models (to fix SourceChangeWarning) 201 | for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']: 202 | detect(opt=opt) 203 | strip_optimizer(opt.weights) 204 | else: 205 | detect(opt=opt) 206 | -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/models/__init__.py -------------------------------------------------------------------------------- /models/experimental.py: -------------------------------------------------------------------------------- 1 | # YOLOv5 experimental modules 2 | 3 | import numpy as np 4 | import torch 5 | import torch.nn as nn 6 | 7 | from models.common import Conv, DWConv 8 | from utils.google_utils import attempt_download 9 | 10 | 11 | class CrossConv(nn.Module): 12 | # Cross Convolution Downsample 13 | def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False): 14 | # ch_in, ch_out, kernel, stride, groups, expansion, shortcut 15 | super(CrossConv, self).__init__() 16 | c_ = int(c2 * e) # hidden channels 17 | self.cv1 = Conv(c1, c_, (1, k), (1, s)) 18 | self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g) 19 | self.add = shortcut and c1 == c2 20 | 21 | def forward(self, x): 22 | return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 23 | 24 | 25 | class Sum(nn.Module): 26 | # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070 27 | def __init__(self, n, weight=False): # n: number of inputs 28 | super(Sum, self).__init__() 29 | self.weight = weight # apply weights boolean 30 | self.iter = range(n - 1) # iter object 31 | if weight: 32 | self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True) # layer weights 33 | 34 | def forward(self, x): 35 | y = x[0] # no weight 36 | if self.weight: 37 | w = torch.sigmoid(self.w) * 2 38 | for i in self.iter: 39 | y = y + x[i + 1] * w[i] 40 | else: 41 | for i in self.iter: 42 | y = y + x[i + 1] 43 | return y 44 | 45 | 46 | class GhostConv(nn.Module): 47 | # Ghost Convolution https://github.com/huawei-noah/ghostnet 48 | def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups 49 | super(GhostConv, self).__init__() 50 | c_ = c2 // 2 # hidden channels 51 | self.cv1 = Conv(c1, c_, k, s, None, g, act) 52 | self.cv2 = Conv(c_, c_, 5, 1, None, c_, act) 53 | 54 | def forward(self, x): 55 | y = self.cv1(x) 56 | return torch.cat([y, self.cv2(y)], 1) 57 | 58 | 59 | class GhostBottleneck(nn.Module): 60 | # Ghost Bottleneck https://github.com/huawei-noah/ghostnet 61 | def __init__(self, c1, c2, k=3, s=1): # ch_in, ch_out, kernel, stride 62 | super(GhostBottleneck, self).__init__() 63 | c_ = c2 // 2 64 | self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1), # pw 65 | DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw 66 | GhostConv(c_, c2, 1, 1, act=False)) # pw-linear 67 | self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False), 68 | Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity() 69 | 70 | def forward(self, x): 71 | return self.conv(x) + self.shortcut(x) 72 | 73 | 74 | class MixConv2d(nn.Module): 75 | # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595 76 | def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True): 77 | super(MixConv2d, self).__init__() 78 | groups = len(k) 79 | if equal_ch: # equal c_ per group 80 | i = torch.linspace(0, groups - 1E-6, c2).floor() # c2 indices 81 | c_ = [(i == g).sum() for g in range(groups)] # intermediate channels 82 | else: # equal weight.numel() per group 83 | b = [c2] + [0] * groups 84 | a = np.eye(groups + 1, groups, k=-1) 85 | a -= np.roll(a, 1, axis=1) 86 | a *= np.array(k) ** 2 87 | a[0] = 1 88 | c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b 89 | 90 | self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)]) 91 | self.bn = nn.BatchNorm2d(c2) 92 | self.act = nn.LeakyReLU(0.1, inplace=True) 93 | 94 | def forward(self, x): 95 | return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1))) 96 | 97 | 98 | class Ensemble(nn.ModuleList): 99 | # Ensemble of models 100 | def __init__(self): 101 | super(Ensemble, self).__init__() 102 | 103 | def forward(self, x, augment=False): 104 | y = [] 105 | for module in self: 106 | y.append(module(x, augment)[0]) 107 | # y = torch.stack(y).max(0)[0] # max ensemble 108 | # y = torch.stack(y).mean(0) # mean ensemble 109 | y = torch.cat(y, 1) # nms ensemble 110 | return y, None # inference, train output 111 | 112 | 113 | def attempt_load(weights, map_location=None, inplace=True): 114 | from models.yolo import Detect, Model 115 | 116 | # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a 117 | model = Ensemble() 118 | for w in weights if isinstance(weights, list) else [weights]: 119 | attempt_download(w) 120 | ckpt = torch.load(w, map_location=map_location) # load 121 | # print("ckpt:",ckpt['model']) 122 | model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().eval()) # FP32 model 123 | # model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval()) # FP32 model 124 | 125 | # Compatibility updates 126 | for m in model.modules(): 127 | if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model]: 128 | m.inplace = inplace # pytorch 1.7.0 compatibility 129 | elif type(m) is Conv: 130 | m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility 131 | 132 | if len(model) == 1: 133 | return model[-1] # return model 134 | else: 135 | print('Ensemble created with %s\n' % weights) 136 | for k in ['names', 'stride']: 137 | setattr(model, k, getattr(model[-1], k)) 138 | return model # return ensemble 139 | -------------------------------------------------------------------------------- /models/export.py: -------------------------------------------------------------------------------- 1 | """Exports a YOLOv5 *.pt model to ONNX and TorchScript formats 2 | 3 | Usage: 4 | $ export PYTHONPATH="$PWD" && python models/export.py --weights yolov5s.pt --img 640 --batch 1 5 | """ 6 | 7 | import argparse 8 | import sys 9 | import time 10 | from pathlib import Path 11 | 12 | sys.path.append(Path(__file__).parent.parent.absolute().__str__()) # to run '$ python *.py' files in subdirectories 13 | 14 | import torch 15 | import torch.nn as nn 16 | from torch.utils.mobile_optimizer import optimize_for_mobile 17 | 18 | import models 19 | from models.experimental import attempt_load 20 | from utils.activations import Hardswish, SiLU 21 | from utils.general import colorstr, check_img_size, check_requirements, file_size, set_logging 22 | from utils.torch_utils import select_device 23 | 24 | if __name__ == '__main__': 25 | parser = argparse.ArgumentParser() 26 | parser.add_argument('--weights', type=str, default='/home/kong/yolov5/runs/train/exp181/weights/last.pt', help='weights path') 27 | parser.add_argument('--img-size', nargs='+', type=int, default=[416, 416], help='image size') # height, width 28 | parser.add_argument('--batch-size', type=int, default=1, help='batch size') 29 | parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 30 | parser.add_argument('--half', action='store_true', help='FP16 half-precision export') 31 | parser.add_argument('--inplace', action='store_true', help='set YOLOv5 Detect() inplace=True') 32 | parser.add_argument('--dynamic', action='store_true', help='dynamic ONNX axes') # ONNX-only 33 | parser.add_argument('--simplify', action='store_true', help='simplify ONNX model') # ONNX-only 34 | opt = parser.parse_args() 35 | opt.img_size *= 2 if len(opt.img_size) == 1 else 1 # expand 36 | print(opt) 37 | set_logging() 38 | t = time.time() 39 | 40 | # Load PyTorch model 41 | device = select_device(opt.device) 42 | model = attempt_load(opt.weights, map_location=device) # load FP32 model 43 | labels = model.names 44 | 45 | # Checks 46 | gs = int(max(model.stride)) # grid size (max stride) 47 | opt.img_size = [check_img_size(x, gs) for x in opt.img_size] # verify img_size are gs-multiples 48 | assert not (opt.device.lower() == "cpu" and opt.half), '--half only compatible with GPU export, i.e. use --device 0' 49 | 50 | # Input 51 | img = torch.zeros(opt.batch_size, 3, *opt.img_size).to(device) # image size(1,3,320,192) iDetection 52 | 53 | # Update model 54 | if opt.half: 55 | img, model = img.half(), model.half() # to FP16 56 | for k, m in model.named_modules(): 57 | m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility 58 | if isinstance(m, models.common.Conv): # assign export-friendly activations 59 | if isinstance(m.act, nn.Hardswish): 60 | m.act = Hardswish() 61 | elif isinstance(m.act, nn.SiLU): 62 | m.act = SiLU() 63 | elif isinstance(m, models.yolo.Detect): 64 | m.inplace = opt.inplace 65 | m.onnx_dynamic = opt.dynamic 66 | # m.forward = m.forward_export # assign forward (optional) 67 | 68 | for _ in range(2): 69 | y = model(img) # dry runs 70 | print(f"\n{colorstr('PyTorch:')} starting from {opt.weights} ({file_size(opt.weights):.1f} MB)") 71 | 72 | # TorchScript export ----------------------------------------------------------------------------------------------- 73 | prefix = colorstr('TorchScript:') 74 | try: 75 | print(f'\n{prefix} starting export with torch {torch.__version__}...') 76 | f = opt.weights.replace('.pt', '.torchscript.pt') # filename 77 | ts = torch.jit.trace(model, img, strict=False) 78 | optimize_for_mobile(ts).save(f) # https://pytorch.org/tutorials/recipes/script_optimized.html 79 | print(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') 80 | except Exception as e: 81 | print(f'{prefix} export failure: {e}') 82 | 83 | # ONNX export ------------------------------------------------------------------------------------------------------ 84 | prefix = colorstr('ONNX:') 85 | try: 86 | import onnx 87 | 88 | print(f'{prefix} starting export with onnx {onnx.__version__}...') 89 | f = opt.weights.replace('.pt', '.onnx') # filename 90 | torch.onnx.export(model, img, f, verbose=False, opset_version=12, input_names=['images'], 91 | dynamic_axes={'images': {0: 'batch', 2: 'height', 3: 'width'}, # size(1,3,640,640) 92 | 'output': {0: 'batch', 2: 'y', 3: 'x'}} if opt.dynamic else None) 93 | 94 | # Checks 95 | model_onnx = onnx.load(f) # load onnx model 96 | onnx.checker.check_model(model_onnx) # check onnx model 97 | # print(onnx.helper.printable_graph(model_onnx.graph)) # print 98 | 99 | # Simplify 100 | if opt.simplify: 101 | try: 102 | check_requirements(['onnx-simplifier']) 103 | import onnxsim 104 | 105 | print(f'{prefix} simplifying with onnx-simplifier {onnxsim.__version__}...') 106 | model_onnx, check = onnxsim.simplify(model_onnx, 107 | dynamic_input_shape=opt.dynamic, 108 | input_shapes={'images': list(img.shape)} if opt.dynamic else None) 109 | assert check, 'assert check failed' 110 | onnx.save(model_onnx, f) 111 | except Exception as e: 112 | print(f'{prefix} simplifier failure: {e}') 113 | print(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') 114 | except Exception as e: 115 | print(f'{prefix} export failure: {e}') 116 | 117 | # CoreML export ---------------------------------------------------------------------------------------------------- 118 | prefix = colorstr('CoreML:') 119 | try: 120 | import coremltools as ct 121 | 122 | print(f'{prefix} starting export with coremltools {ct.__version__}...') 123 | # convert model from torchscript and apply pixel scaling as per detect.py 124 | model = ct.convert(ts, inputs=[ct.ImageType(name='image', shape=img.shape, scale=1 / 255.0, bias=[0, 0, 0])]) 125 | f = opt.weights.replace('.pt', '.mlmodel') # filename 126 | model.save(f) 127 | print(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') 128 | except Exception as e: 129 | print(f'{prefix} export failure: {e}') 130 | 131 | # Finish 132 | print(f'\nExport complete ({time.time() - t:.2f}s). Visualize with https://github.com/lutzroeder/netron.') 133 | -------------------------------------------------------------------------------- /models/hub/anchors.yaml: -------------------------------------------------------------------------------- 1 | # Default YOLOv5 anchors for COCO data 2 | 3 | 4 | # P5 ------------------------------------------------------------------------------------------------------------------- 5 | # P5-640: 6 | anchors_p5_640: 7 | - [ 10,13, 16,30, 33,23 ] # P3/8 8 | - [ 30,61, 62,45, 59,119 ] # P4/16 9 | - [ 116,90, 156,198, 373,326 ] # P5/32 10 | 11 | 12 | # P6 ------------------------------------------------------------------------------------------------------------------- 13 | # P6-640: thr=0.25: 0.9964 BPR, 5.54 anchors past thr, n=12, img_size=640, metric_all=0.281/0.716-mean/best, past_thr=0.469-mean: 9,11, 21,19, 17,41, 43,32, 39,70, 86,64, 65,131, 134,130, 120,265, 282,180, 247,354, 512,387 14 | anchors_p6_640: 15 | - [ 9,11, 21,19, 17,41 ] # P3/8 16 | - [ 43,32, 39,70, 86,64 ] # P4/16 17 | - [ 65,131, 134,130, 120,265 ] # P5/32 18 | - [ 282,180, 247,354, 512,387 ] # P6/64 19 | 20 | # P6-1280: thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1280, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 19,27, 44,40, 38,94, 96,68, 86,152, 180,137, 140,301, 303,264, 238,542, 436,615, 739,380, 925,792 21 | anchors_p6_1280: 22 | - [ 19,27, 44,40, 38,94 ] # P3/8 23 | - [ 96,68, 86,152, 180,137 ] # P4/16 24 | - [ 140,301, 303,264, 238,542 ] # P5/32 25 | - [ 436,615, 739,380, 925,792 ] # P6/64 26 | 27 | # P6-1920: thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1920, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 28,41, 67,59, 57,141, 144,103, 129,227, 270,205, 209,452, 455,396, 358,812, 653,922, 1109,570, 1387,1187 28 | anchors_p6_1920: 29 | - [ 28,41, 67,59, 57,141 ] # P3/8 30 | - [ 144,103, 129,227, 270,205 ] # P4/16 31 | - [ 209,452, 455,396, 358,812 ] # P5/32 32 | - [ 653,922, 1109,570, 1387,1187 ] # P6/64 33 | 34 | 35 | # P7 ------------------------------------------------------------------------------------------------------------------- 36 | # P7-640: thr=0.25: 0.9962 BPR, 6.76 anchors past thr, n=15, img_size=640, metric_all=0.275/0.733-mean/best, past_thr=0.466-mean: 11,11, 13,30, 29,20, 30,46, 61,38, 39,92, 78,80, 146,66, 79,163, 149,150, 321,143, 157,303, 257,402, 359,290, 524,372 37 | anchors_p7_640: 38 | - [ 11,11, 13,30, 29,20 ] # P3/8 39 | - [ 30,46, 61,38, 39,92 ] # P4/16 40 | - [ 78,80, 146,66, 79,163 ] # P5/32 41 | - [ 149,150, 321,143, 157,303 ] # P6/64 42 | - [ 257,402, 359,290, 524,372 ] # P7/128 43 | 44 | # P7-1280: thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1280, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 19,22, 54,36, 32,77, 70,83, 138,71, 75,173, 165,159, 148,334, 375,151, 334,317, 251,626, 499,474, 750,326, 534,814, 1079,818 45 | anchors_p7_1280: 46 | - [ 19,22, 54,36, 32,77 ] # P3/8 47 | - [ 70,83, 138,71, 75,173 ] # P4/16 48 | - [ 165,159, 148,334, 375,151 ] # P5/32 49 | - [ 334,317, 251,626, 499,474 ] # P6/64 50 | - [ 750,326, 534,814, 1079,818 ] # P7/128 51 | 52 | # P7-1920: thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1920, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 29,34, 81,55, 47,115, 105,124, 207,107, 113,259, 247,238, 222,500, 563,227, 501,476, 376,939, 749,711, 1126,489, 801,1222, 1618,1227 53 | anchors_p7_1920: 54 | - [ 29,34, 81,55, 47,115 ] # P3/8 55 | - [ 105,124, 207,107, 113,259 ] # P4/16 56 | - [ 247,238, 222,500, 563,227 ] # P5/32 57 | - [ 501,476, 376,939, 749,711 ] # P6/64 58 | - [ 1126,489, 801,1222, 1618,1227 ] # P7/128 59 | -------------------------------------------------------------------------------- /models/hub/yolov3-spp.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # darknet53 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Conv, [32, 3, 1]], # 0 16 | [-1, 1, Conv, [64, 3, 2]], # 1-P1/2 17 | [-1, 1, Bottleneck, [64]], 18 | [-1, 1, Conv, [128, 3, 2]], # 3-P2/4 19 | [-1, 2, Bottleneck, [128]], 20 | [-1, 1, Conv, [256, 3, 2]], # 5-P3/8 21 | [-1, 8, Bottleneck, [256]], 22 | [-1, 1, Conv, [512, 3, 2]], # 7-P4/16 23 | [-1, 8, Bottleneck, [512]], 24 | [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32 25 | [-1, 4, Bottleneck, [1024]], # 10 26 | ] 27 | 28 | # YOLOv3-SPP head 29 | head: 30 | [[-1, 1, Bottleneck, [1024, False]], 31 | [-1, 1, SPP, [512, [5, 9, 13]]], 32 | [-1, 1, Conv, [1024, 3, 1]], 33 | [-1, 1, Conv, [512, 1, 1]], 34 | [-1, 1, Conv, [1024, 3, 1]], # 15 (P5/32-large) 35 | 36 | [-2, 1, Conv, [256, 1, 1]], 37 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 38 | [[-1, 8], 1, Concat, [1]], # cat backbone P4 39 | [-1, 1, Bottleneck, [512, False]], 40 | [-1, 1, Bottleneck, [512, False]], 41 | [-1, 1, Conv, [256, 1, 1]], 42 | [-1, 1, Conv, [512, 3, 1]], # 22 (P4/16-medium) 43 | 44 | [-2, 1, Conv, [128, 1, 1]], 45 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 46 | [[-1, 6], 1, Concat, [1]], # cat backbone P3 47 | [-1, 1, Bottleneck, [256, False]], 48 | [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small) 49 | 50 | [[27, 22, 15], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 51 | ] 52 | -------------------------------------------------------------------------------- /models/hub/yolov3-tiny.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,14, 23,27, 37,58] # P4/16 9 | - [81,82, 135,169, 344,319] # P5/32 10 | 11 | # YOLOv3-tiny backbone 12 | backbone: 13 | # [from, number, module, args] 14 | [[-1, 1, Conv, [16, 3, 1]], # 0 15 | [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 1-P1/2 16 | [-1, 1, Conv, [32, 3, 1]], 17 | [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 3-P2/4 18 | [-1, 1, Conv, [64, 3, 1]], 19 | [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 5-P3/8 20 | [-1, 1, Conv, [128, 3, 1]], 21 | [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 7-P4/16 22 | [-1, 1, Conv, [256, 3, 1]], 23 | [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 9-P5/32 24 | [-1, 1, Conv, [512, 3, 1]], 25 | [-1, 1, nn.ZeroPad2d, [[0, 1, 0, 1]]], # 11 26 | [-1, 1, nn.MaxPool2d, [2, 1, 0]], # 12 27 | ] 28 | 29 | # YOLOv3-tiny head 30 | head: 31 | [[-1, 1, Conv, [1024, 3, 1]], 32 | [-1, 1, Conv, [256, 1, 1]], 33 | [-1, 1, Conv, [512, 3, 1]], # 15 (P5/32-large) 34 | 35 | [-2, 1, Conv, [128, 1, 1]], 36 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 37 | [[-1, 8], 1, Concat, [1]], # cat backbone P4 38 | [-1, 1, Conv, [256, 3, 1]], # 19 (P4/16-medium) 39 | 40 | [[19, 15], 1, Detect, [nc, anchors]], # Detect(P4, P5) 41 | ] 42 | -------------------------------------------------------------------------------- /models/hub/yolov3.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # darknet53 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Conv, [32, 3, 1]], # 0 16 | [-1, 1, Conv, [64, 3, 2]], # 1-P1/2 17 | [-1, 1, Bottleneck, [64]], 18 | [-1, 1, Conv, [128, 3, 2]], # 3-P2/4 19 | [-1, 2, Bottleneck, [128]], 20 | [-1, 1, Conv, [256, 3, 2]], # 5-P3/8 21 | [-1, 8, Bottleneck, [256]], 22 | [-1, 1, Conv, [512, 3, 2]], # 7-P4/16 23 | [-1, 8, Bottleneck, [512]], 24 | [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32 25 | [-1, 4, Bottleneck, [1024]], # 10 26 | ] 27 | 28 | # YOLOv3 head 29 | head: 30 | [[-1, 1, Bottleneck, [1024, False]], 31 | [-1, 1, Conv, [512, [1, 1]]], 32 | [-1, 1, Conv, [1024, 3, 1]], 33 | [-1, 1, Conv, [512, 1, 1]], 34 | [-1, 1, Conv, [1024, 3, 1]], # 15 (P5/32-large) 35 | 36 | [-2, 1, Conv, [256, 1, 1]], 37 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 38 | [[-1, 8], 1, Concat, [1]], # cat backbone P4 39 | [-1, 1, Bottleneck, [512, False]], 40 | [-1, 1, Bottleneck, [512, False]], 41 | [-1, 1, Conv, [256, 1, 1]], 42 | [-1, 1, Conv, [512, 3, 1]], # 22 (P4/16-medium) 43 | 44 | [-2, 1, Conv, [128, 1, 1]], 45 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 46 | [[-1, 6], 1, Concat, [1]], # cat backbone P3 47 | [-1, 1, Bottleneck, [256, False]], 48 | [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small) 49 | 50 | [[27, 22, 15], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 51 | ] 52 | -------------------------------------------------------------------------------- /models/hub/yolov5-fpn.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, Bottleneck, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, BottleneckCSP, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, BottleneckCSP, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 6, BottleneckCSP, [1024]], # 9 25 | ] 26 | 27 | # YOLOv5 FPN head 28 | head: 29 | [[-1, 3, BottleneckCSP, [1024, False]], # 10 (P5/32-large) 30 | 31 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 32 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 33 | [-1, 1, Conv, [512, 1, 1]], 34 | [-1, 3, BottleneckCSP, [512, False]], # 14 (P4/16-medium) 35 | 36 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 37 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 38 | [-1, 1, Conv, [256, 1, 1]], 39 | [-1, 3, BottleneckCSP, [256, False]], # 18 (P3/8-small) 40 | 41 | [[18, 14, 10], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 42 | ] 43 | -------------------------------------------------------------------------------- /models/hub/yolov5-p2.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 3 8 | 9 | # YOLOv5 backbone 10 | backbone: 11 | # [from, number, module, args] 12 | [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 13 | [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 14 | [ -1, 3, C3, [ 128 ] ], 15 | [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 16 | [ -1, 9, C3, [ 256 ] ], 17 | [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 18 | [ -1, 9, C3, [ 512 ] ], 19 | [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 7-P5/32 20 | [ -1, 1, SPP, [ 1024, [ 5, 9, 13 ] ] ], 21 | [ -1, 3, C3, [ 1024, False ] ], # 9 22 | ] 23 | 24 | # YOLOv5 head 25 | head: 26 | [ [ -1, 1, Conv, [ 512, 1, 1 ] ], 27 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 28 | [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 29 | [ -1, 3, C3, [ 512, False ] ], # 13 30 | 31 | [ -1, 1, Conv, [ 256, 1, 1 ] ], 32 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 33 | [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 34 | [ -1, 3, C3, [ 256, False ] ], # 17 (P3/8-small) 35 | 36 | [ -1, 1, Conv, [ 128, 1, 1 ] ], 37 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 38 | [ [ -1, 2 ], 1, Concat, [ 1 ] ], # cat backbone P2 39 | [ -1, 1, C3, [ 128, False ] ], # 21 (P2/4-xsmall) 40 | 41 | [ -1, 1, Conv, [ 128, 3, 2 ] ], 42 | [ [ -1, 18 ], 1, Concat, [ 1 ] ], # cat head P3 43 | [ -1, 3, C3, [ 256, False ] ], # 24 (P3/8-small) 44 | 45 | [ -1, 1, Conv, [ 256, 3, 2 ] ], 46 | [ [ -1, 14 ], 1, Concat, [ 1 ] ], # cat head P4 47 | [ -1, 3, C3, [ 512, False ] ], # 27 (P4/16-medium) 48 | 49 | [ -1, 1, Conv, [ 512, 3, 2 ] ], 50 | [ [ -1, 10 ], 1, Concat, [ 1 ] ], # cat head P5 51 | [ -1, 3, C3, [ 1024, False ] ], # 30 (P5/32-large) 52 | 53 | [ [ 24, 27, 30 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5) 54 | ] 55 | -------------------------------------------------------------------------------- /models/hub/yolov5-p6.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 3 8 | 9 | # YOLOv5 backbone 10 | backbone: 11 | # [from, number, module, args] 12 | [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 13 | [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 14 | [ -1, 3, C3, [ 128 ] ], 15 | [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 16 | [ -1, 9, C3, [ 256 ] ], 17 | [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 18 | [ -1, 9, C3, [ 512 ] ], 19 | [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 20 | [ -1, 3, C3, [ 768 ] ], 21 | [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 22 | [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], 23 | [ -1, 3, C3, [ 1024, False ] ], # 11 24 | ] 25 | 26 | # YOLOv5 head 27 | head: 28 | [ [ -1, 1, Conv, [ 768, 1, 1 ] ], 29 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 30 | [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 31 | [ -1, 3, C3, [ 768, False ] ], # 15 32 | 33 | [ -1, 1, Conv, [ 512, 1, 1 ] ], 34 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 35 | [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 36 | [ -1, 3, C3, [ 512, False ] ], # 19 37 | 38 | [ -1, 1, Conv, [ 256, 1, 1 ] ], 39 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 40 | [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 41 | [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) 42 | 43 | [ -1, 1, Conv, [ 256, 3, 2 ] ], 44 | [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 45 | [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) 46 | 47 | [ -1, 1, Conv, [ 512, 3, 2 ] ], 48 | [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 49 | [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) 50 | 51 | [ -1, 1, Conv, [ 768, 3, 2 ] ], 52 | [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 53 | [ -1, 3, C3, [ 1024, False ] ], # 32 (P5/64-xlarge) 54 | 55 | [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) 56 | ] 57 | -------------------------------------------------------------------------------- /models/hub/yolov5-p7.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 3 8 | 9 | # YOLOv5 backbone 10 | backbone: 11 | # [from, number, module, args] 12 | [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 13 | [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 14 | [ -1, 3, C3, [ 128 ] ], 15 | [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 16 | [ -1, 9, C3, [ 256 ] ], 17 | [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 18 | [ -1, 9, C3, [ 512 ] ], 19 | [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 20 | [ -1, 3, C3, [ 768 ] ], 21 | [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 22 | [ -1, 3, C3, [ 1024 ] ], 23 | [ -1, 1, Conv, [ 1280, 3, 2 ] ], # 11-P7/128 24 | [ -1, 1, SPP, [ 1280, [ 3, 5 ] ] ], 25 | [ -1, 3, C3, [ 1280, False ] ], # 13 26 | ] 27 | 28 | # YOLOv5 head 29 | head: 30 | [ [ -1, 1, Conv, [ 1024, 1, 1 ] ], 31 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 32 | [ [ -1, 10 ], 1, Concat, [ 1 ] ], # cat backbone P6 33 | [ -1, 3, C3, [ 1024, False ] ], # 17 34 | 35 | [ -1, 1, Conv, [ 768, 1, 1 ] ], 36 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 37 | [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 38 | [ -1, 3, C3, [ 768, False ] ], # 21 39 | 40 | [ -1, 1, Conv, [ 512, 1, 1 ] ], 41 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 42 | [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 43 | [ -1, 3, C3, [ 512, False ] ], # 25 44 | 45 | [ -1, 1, Conv, [ 256, 1, 1 ] ], 46 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 47 | [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 48 | [ -1, 3, C3, [ 256, False ] ], # 29 (P3/8-small) 49 | 50 | [ -1, 1, Conv, [ 256, 3, 2 ] ], 51 | [ [ -1, 26 ], 1, Concat, [ 1 ] ], # cat head P4 52 | [ -1, 3, C3, [ 512, False ] ], # 32 (P4/16-medium) 53 | 54 | [ -1, 1, Conv, [ 512, 3, 2 ] ], 55 | [ [ -1, 22 ], 1, Concat, [ 1 ] ], # cat head P5 56 | [ -1, 3, C3, [ 768, False ] ], # 35 (P5/32-large) 57 | 58 | [ -1, 1, Conv, [ 768, 3, 2 ] ], 59 | [ [ -1, 18 ], 1, Concat, [ 1 ] ], # cat head P6 60 | [ -1, 3, C3, [ 1024, False ] ], # 38 (P6/64-xlarge) 61 | 62 | [ -1, 1, Conv, [ 1024, 3, 2 ] ], 63 | [ [ -1, 14 ], 1, Concat, [ 1 ] ], # cat head P7 64 | [ -1, 3, C3, [ 1280, False ] ], # 41 (P7/128-xxlarge) 65 | 66 | [ [ 29, 32, 35, 38, 41 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6, P7) 67 | ] 68 | -------------------------------------------------------------------------------- /models/hub/yolov5-panet.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, BottleneckCSP, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, BottleneckCSP, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, BottleneckCSP, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, BottleneckCSP, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 PANet head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, BottleneckCSP, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, BottleneckCSP, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, BottleneckCSP, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, BottleneckCSP, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/hub/yolov5l6.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [ 19,27, 44,40, 38,94 ] # P3/8 9 | - [ 96,68, 86,152, 180,137 ] # P4/16 10 | - [ 140,301, 303,264, 238,542 ] # P5/32 11 | - [ 436,615, 739,380, 925,792 ] # P6/64 12 | 13 | # YOLOv5 backbone 14 | backbone: 15 | # [from, number, module, args] 16 | [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 17 | [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 18 | [ -1, 3, C3, [ 128 ] ], 19 | [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 20 | [ -1, 9, C3, [ 256 ] ], 21 | [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 22 | [ -1, 9, C3, [ 512 ] ], 23 | [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 24 | [ -1, 3, C3, [ 768 ] ], 25 | [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 26 | [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], 27 | [ -1, 3, C3, [ 1024, False ] ], # 11 28 | ] 29 | 30 | # YOLOv5 head 31 | head: 32 | [ [ -1, 1, Conv, [ 768, 1, 1 ] ], 33 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 34 | [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 35 | [ -1, 3, C3, [ 768, False ] ], # 15 36 | 37 | [ -1, 1, Conv, [ 512, 1, 1 ] ], 38 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 39 | [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 40 | [ -1, 3, C3, [ 512, False ] ], # 19 41 | 42 | [ -1, 1, Conv, [ 256, 1, 1 ] ], 43 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 44 | [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 45 | [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) 46 | 47 | [ -1, 1, Conv, [ 256, 3, 2 ] ], 48 | [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 49 | [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) 50 | 51 | [ -1, 1, Conv, [ 512, 3, 2 ] ], 52 | [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 53 | [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) 54 | 55 | [ -1, 1, Conv, [ 768, 3, 2 ] ], 56 | [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 57 | [ -1, 3, C3, [ 1024, False ] ], # 32 (P6/64-xlarge) 58 | 59 | [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) 60 | ] 61 | -------------------------------------------------------------------------------- /models/hub/yolov5m6.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.67 # model depth multiple 4 | width_multiple: 0.75 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [ 19,27, 44,40, 38,94 ] # P3/8 9 | - [ 96,68, 86,152, 180,137 ] # P4/16 10 | - [ 140,301, 303,264, 238,542 ] # P5/32 11 | - [ 436,615, 739,380, 925,792 ] # P6/64 12 | 13 | # YOLOv5 backbone 14 | backbone: 15 | # [from, number, module, args] 16 | [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 17 | [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 18 | [ -1, 3, C3, [ 128 ] ], 19 | [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 20 | [ -1, 9, C3, [ 256 ] ], 21 | [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 22 | [ -1, 9, C3, [ 512 ] ], 23 | [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 24 | [ -1, 3, C3, [ 768 ] ], 25 | [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 26 | [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], 27 | [ -1, 3, C3, [ 1024, False ] ], # 11 28 | ] 29 | 30 | # YOLOv5 head 31 | head: 32 | [ [ -1, 1, Conv, [ 768, 1, 1 ] ], 33 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 34 | [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 35 | [ -1, 3, C3, [ 768, False ] ], # 15 36 | 37 | [ -1, 1, Conv, [ 512, 1, 1 ] ], 38 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 39 | [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 40 | [ -1, 3, C3, [ 512, False ] ], # 19 41 | 42 | [ -1, 1, Conv, [ 256, 1, 1 ] ], 43 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 44 | [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 45 | [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) 46 | 47 | [ -1, 1, Conv, [ 256, 3, 2 ] ], 48 | [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 49 | [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) 50 | 51 | [ -1, 1, Conv, [ 512, 3, 2 ] ], 52 | [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 53 | [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) 54 | 55 | [ -1, 1, Conv, [ 768, 3, 2 ] ], 56 | [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 57 | [ -1, 3, C3, [ 1024, False ] ], # 32 (P6/64-xlarge) 58 | 59 | [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) 60 | ] 61 | -------------------------------------------------------------------------------- /models/hub/yolov5s-transformer.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3TR, [1024, False]], # 9 <-------- C3TR() Transformer module 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/hub/yolov5s6.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [ 19,27, 44,40, 38,94 ] # P3/8 9 | - [ 96,68, 86,152, 180,137 ] # P4/16 10 | - [ 140,301, 303,264, 238,542 ] # P5/32 11 | - [ 436,615, 739,380, 925,792 ] # P6/64 12 | 13 | # YOLOv5 backbone 14 | backbone: 15 | # [from, number, module, args] 16 | [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 17 | [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 18 | [ -1, 3, C3, [ 128 ] ], 19 | [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 20 | [ -1, 9, C3, [ 256 ] ], 21 | [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 22 | [ -1, 9, C3, [ 512 ] ], 23 | [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 24 | [ -1, 3, C3, [ 768 ] ], 25 | [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 26 | [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], 27 | [ -1, 3, C3, [ 1024, False ] ], # 11 28 | ] 29 | 30 | # YOLOv5 head 31 | head: 32 | [ [ -1, 1, Conv, [ 768, 1, 1 ] ], 33 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 34 | [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 35 | [ -1, 3, C3, [ 768, False ] ], # 15 36 | 37 | [ -1, 1, Conv, [ 512, 1, 1 ] ], 38 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 39 | [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 40 | [ -1, 3, C3, [ 512, False ] ], # 19 41 | 42 | [ -1, 1, Conv, [ 256, 1, 1 ] ], 43 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 44 | [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 45 | [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) 46 | 47 | [ -1, 1, Conv, [ 256, 3, 2 ] ], 48 | [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 49 | [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) 50 | 51 | [ -1, 1, Conv, [ 512, 3, 2 ] ], 52 | [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 53 | [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) 54 | 55 | [ -1, 1, Conv, [ 768, 3, 2 ] ], 56 | [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 57 | [ -1, 3, C3, [ 1024, False ] ], # 32 (P6/64-xlarge) 58 | 59 | [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) 60 | ] 61 | -------------------------------------------------------------------------------- /models/hub/yolov5x6.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.33 # model depth multiple 4 | width_multiple: 1.25 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [ 19,27, 44,40, 38,94 ] # P3/8 9 | - [ 96,68, 86,152, 180,137 ] # P4/16 10 | - [ 140,301, 303,264, 238,542 ] # P5/32 11 | - [ 436,615, 739,380, 925,792 ] # P6/64 12 | 13 | # YOLOv5 backbone 14 | backbone: 15 | # [from, number, module, args] 16 | [ [ -1, 1, Focus, [ 64, 3 ] ], # 0-P1/2 17 | [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 18 | [ -1, 3, C3, [ 128 ] ], 19 | [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 20 | [ -1, 9, C3, [ 256 ] ], 21 | [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 22 | [ -1, 9, C3, [ 512 ] ], 23 | [ -1, 1, Conv, [ 768, 3, 2 ] ], # 7-P5/32 24 | [ -1, 3, C3, [ 768 ] ], 25 | [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 9-P6/64 26 | [ -1, 1, SPP, [ 1024, [ 3, 5, 7 ] ] ], 27 | [ -1, 3, C3, [ 1024, False ] ], # 11 28 | ] 29 | 30 | # YOLOv5 head 31 | head: 32 | [ [ -1, 1, Conv, [ 768, 1, 1 ] ], 33 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 34 | [ [ -1, 8 ], 1, Concat, [ 1 ] ], # cat backbone P5 35 | [ -1, 3, C3, [ 768, False ] ], # 15 36 | 37 | [ -1, 1, Conv, [ 512, 1, 1 ] ], 38 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 39 | [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 40 | [ -1, 3, C3, [ 512, False ] ], # 19 41 | 42 | [ -1, 1, Conv, [ 256, 1, 1 ] ], 43 | [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], 44 | [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 45 | [ -1, 3, C3, [ 256, False ] ], # 23 (P3/8-small) 46 | 47 | [ -1, 1, Conv, [ 256, 3, 2 ] ], 48 | [ [ -1, 20 ], 1, Concat, [ 1 ] ], # cat head P4 49 | [ -1, 3, C3, [ 512, False ] ], # 26 (P4/16-medium) 50 | 51 | [ -1, 1, Conv, [ 512, 3, 2 ] ], 52 | [ [ -1, 16 ], 1, Concat, [ 1 ] ], # cat head P5 53 | [ -1, 3, C3, [ 768, False ] ], # 29 (P5/32-large) 54 | 55 | [ -1, 1, Conv, [ 768, 3, 2 ] ], 56 | [ [ -1, 12 ], 1, Concat, [ 1 ] ], # cat head P6 57 | [ -1, 3, C3, [ 1024, False ] ], # 32 (P6/64-xlarge) 58 | 59 | [ [ 23, 26, 29, 32 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4, P5, P6) 60 | ] 61 | -------------------------------------------------------------------------------- /models/pruned_common.py: -------------------------------------------------------------------------------- 1 | # YOLOv5 pruned common modules 2 | 3 | import math 4 | from copy import copy 5 | from pathlib import Path 6 | 7 | import numpy as np 8 | import pandas as pd 9 | import requests 10 | import torch 11 | import torch.nn as nn 12 | from torch.cuda import amp 13 | from models.common import Conv 14 | 15 | 16 | class BottleneckPruned(nn.Module): 17 | # Pruned bottleneck 18 | def __init__(self, cv1in, cv1out, cv2out, shortcut=True, g=1): # ch_in, ch_out, shortcut, groups, expansion 19 | super(BottleneckPruned, self).__init__() 20 | self.cv1 = Conv(cv1in, cv1out, 1, 1) 21 | self.cv2 = Conv(cv1out, cv2out, 3, 1, g=g) 22 | self.add = shortcut and cv1in == cv2out 23 | 24 | def forward(self, x): 25 | return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 26 | 27 | 28 | class C3Pruned(nn.Module): 29 | # CSP Bottleneck with 3 convolutions 30 | def __init__(self, cv1in, cv1out, cv2out, cv3out, bottle_args, n=1, shortcut=True, g=1): # ch_in, ch_out, number, shortcut, groups, expansion 31 | super(C3Pruned, self).__init__() 32 | cv3in = bottle_args[-1][-1] 33 | self.cv1 = Conv(cv1in, cv1out, 1, 1) 34 | self.cv2 = Conv(cv1in, cv2out, 1, 1) 35 | self.cv3 = Conv(cv3in+cv2out, cv3out, 1) 36 | self.m = nn.Sequential(*[BottleneckPruned(*bottle_args[k], shortcut, g) for k in range(n)]) 37 | 38 | def forward(self, x): 39 | return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1)) 40 | 41 | 42 | class C3PrunedNoRes(nn.Module): 43 | # CSP Bottleneck with 3 convolutions 44 | def __init__(self, cv1in, cv1out, cv2out, cv3out): # ch_in, ch_out 45 | super(C3PrunedNoRes, self).__init__() 46 | self.cv1 = Conv(cv1in, cv1out, 1, 1) 47 | self.cv2 = Conv(cv1in, cv2out, 1, 1) 48 | self.cv3 = Conv(cv1out+cv2out, cv3out, 1) 49 | 50 | def forward(self, x): 51 | return self.cv3(torch.cat((self.cv1(x), self.cv2(x)), dim=1)) 52 | 53 | class SPPPruned(nn.Module): 54 | # Spatial pyramid pooling layer used in YOLOv3-SPP 55 | def __init__(self, cv1in, cv1out, cv2out, k=(5, 9, 13)): 56 | super(SPPPruned, self).__init__() 57 | self.cv1 = Conv(cv1in, cv1out, 1, 1) 58 | self.cv2 = Conv(cv1out * (len(k) + 1), cv2out, 1, 1) 59 | self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k]) 60 | 61 | def forward(self, x): 62 | x = self.cv1(x) 63 | return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1)) 64 | 65 | 66 | -------------------------------------------------------------------------------- /models/yolov5l.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 22 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/yolov5m.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.67 # model depth multiple 4 | width_multiple: 0.75 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/yolov5s.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/yolov5slite.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 22 # number of classes 3 | depth_multiple: 0.25 # model depth multiple 4 | width_multiple: 0.25 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 7, 9]]], 24 | [-1, 3, C3, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/yolov5sprune.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [0:[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | 1:[-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | 2:[-1, 3, C3, [128]], 18 | 3:[-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | 4:[-1, 9, C3, [256]], 20 | 5:[-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | 6:[-1, 9, C3, [512]], 22 | 7:[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | 8:[-1, 1, SPP, [1024, [5, 9, 13]]], 24 | 9:[-1, 3, C3, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [10:[-1, 1, Conv, [512, 1, 1]], 30 | 11:[-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | 12:[[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | 13:[-1, 3, C3, [512, False]], # 13 33 | 34 | 14:[-1, 1, Conv, [256, 1, 1]], 35 | 15:[-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | 16:[[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | 17:[-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | 18:[-1, 1, Conv, [256, 3, 2]], 40 | 19:[[-1, 14], 1, Concat, [1]], # cat head P4 41 | 20:[-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | 21:[-1, 1, Conv, [512, 3, 2]], 44 | 22:[[-1, 10], 1, Concat, [1]], # cat head P5 45 | 23:[-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | 24:[[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/yolov5x.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.33 # model depth multiple 4 | width_multiple: 1.25 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /prune_utils.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | # @Time : 2021/5/24 下午4:36 3 | # @Author : midaskong 4 | # @File : prune_utils.py 5 | # @Description: 6 | 7 | import torch 8 | from copy import deepcopy 9 | import numpy as np 10 | import torch.nn.functional as F 11 | 12 | 13 | def gather_bn_weights(module_list): 14 | prune_idx = list(range(len(module_list))) 15 | size_list = [idx.weight.data.shape[0] for idx in module_list.values()] 16 | bn_weights = torch.zeros(sum(size_list)) 17 | index = 0 18 | for i, idx in enumerate(module_list.values()): 19 | size = size_list[i] 20 | bn_weights[index:(index + size)] = idx.weight.data.abs().clone() 21 | index += size 22 | return bn_weights 23 | 24 | def gather_conv_weights(module_list): 25 | prune_idx = list(range(len(module_list))) 26 | size_list = [idx.weight.data.shape[0] for idx in module_list.values()] 27 | 28 | conv_weights = torch.zeros(sum(size_list)) 29 | index = 0 30 | for i, idx in enumerate(module_list.values()): 31 | size = size_list[i] 32 | conv_weights[index:(index + size)] = idx.weight.data.abs().sum(dim=1).sum(dim=1).sum(dim=1).clone() 33 | index += size 34 | return conv_weights 35 | 36 | 37 | def obtain_bn_mask(bn_module, thre): 38 | 39 | thre = thre.cuda() 40 | mask = bn_module.weight.data.abs().ge(thre).float() 41 | 42 | return mask 43 | 44 | 45 | def obtain_conv_mask(conv_module, thre): 46 | thre = thre.cuda() 47 | mask = conv_module.weight.data.abs().sum(dim=1).sum(dim=1).sum(dim=1).ge(thre).float() 48 | return mask 49 | 50 | def uodate_pruned_yolov5_cfg(model, maskbndict): 51 | # save pruned yolov5 model in yaml format: 52 | # model: 53 | # model to be pruned 54 | # maskbndict: 55 | # key : module name 56 | # value : bn layer mask index 57 | return -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | # pip install -r requirements.txt 2 | 3 | # base ---------------------------------------- 4 | matplotlib>=3.2.2 5 | numpy>=1.18.5 6 | opencv-python>=4.1.2 7 | Pillow 8 | PyYAML>=5.3.1 9 | scipy>=1.4.1 10 | torch>=1.7.0 11 | torchvision>=0.8.1 12 | tqdm>=4.41.0 13 | 14 | # logging ------------------------------------- 15 | tensorboard>=2.4.1 16 | # wandb 17 | 18 | # plotting ------------------------------------ 19 | seaborn>=0.11.0 20 | pandas 21 | 22 | # export -------------------------------------- 23 | # coremltools>=4.1 24 | # onnx>=1.8.1 25 | # scikit-learn==0.19.2 # for coreml quantization 26 | 27 | # extras -------------------------------------- 28 | thop # FLOPS computation 29 | pycocotools>=2.0 # COCO mAP 30 | -------------------------------------------------------------------------------- /showbn.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import time 3 | from pathlib import Path 4 | 5 | import cv2 6 | import torch 7 | import torch.backends.cudnn as cudnn 8 | from numpy import random 9 | 10 | from models.experimental import attempt_load 11 | from utils.datasets import LoadStreams, LoadImages 12 | from utils.general import check_img_size, check_requirements, check_imshow, non_max_suppression, apply_classifier, \ 13 | scale_coords, xyxy2xywh, strip_optimizer, set_logging, increment_path, save_one_box 14 | from utils.plots import colors, plot_one_box 15 | from utils.torch_utils import select_device, load_classifier, time_synchronized 16 | 17 | 18 | def detect(opt): 19 | source, weights, view_img, save_txt, imgsz = opt.source, opt.weights, opt.view_img, opt.save_txt, opt.img_size 20 | save_img = not opt.nosave and not source.endswith('.txt') # save inference images 21 | webcam = source.isnumeric() or source.endswith('.txt') or source.lower().startswith( 22 | ('rtsp://', 'rtmp://', 'http://', 'https://')) 23 | 24 | # Directories 25 | save_dir = increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok) # increment run 26 | (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir 27 | 28 | # Initialize 29 | set_logging() 30 | device = select_device(opt.device) 31 | half = device.type != 'cpu' # half precision only supported on CUDA 32 | 33 | # Load model 34 | model = attempt_load(weights, map_location=device) # load FP32 model 35 | stride = int(model.stride.max()) # model stride 36 | imgsz = check_img_size(imgsz, s=stride) # check img_size 37 | names = model.module.names if hasattr(model, 'module') else model.names # get class names 38 | 39 | # Run inference 40 | if device.type != 'cpu': 41 | model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters()))) # run once 42 | t0 = time.time() 43 | print(model) 44 | for i,layer in enumerate(model.state_dict().keys()): 45 | print(f"{i} : {layer} : {model.state_dict()[layer].size()}") 46 | 47 | 48 | if __name__ == '__main__': 49 | parser = argparse.ArgumentParser() 50 | parser.add_argument('--weights', nargs='+', type=str, default='/home/kong/yolov5/yolov5s.pt', help='model.pt path(s)') 51 | parser.add_argument('--source', type=str, default='/home/kong/yolov5/data/test', help='source') # file/folder, 0 for webcam 52 | parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)') 53 | parser.add_argument('--conf-thres', type=float, default=0.25, help='object confidence threshold') 54 | parser.add_argument('--iou-thres', type=float, default=0.45, help='IOU threshold for NMS') 55 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 56 | parser.add_argument('--view-img', action='store_true', help='display results') 57 | parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') 58 | parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') 59 | parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes') 60 | parser.add_argument('--nosave', action='store_true', help='do not save images/videos') 61 | parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --class 0, or --class 0 2 3') 62 | parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') 63 | parser.add_argument('--augment', action='store_true', help='augmented inference') 64 | parser.add_argument('--update', action='store_true', help='update all models') 65 | parser.add_argument('--project', default='runs/detect', help='save results to project/name') 66 | parser.add_argument('--name', default='exp', help='save results to project/name') 67 | parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') 68 | parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)') 69 | parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels') 70 | parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences') 71 | opt = parser.parse_args() 72 | print(opt) 73 | check_requirements(exclude=('tensorboard', 'pycocotools', 'thop')) 74 | 75 | with torch.no_grad(): 76 | if opt.update: # update all models (to fix SourceChangeWarning) 77 | for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']: 78 | detect(opt=opt) 79 | strip_optimizer(opt.weights) 80 | else: 81 | detect(opt=opt) 82 | -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/utils/__init__.py -------------------------------------------------------------------------------- /utils/activations.py: -------------------------------------------------------------------------------- 1 | # Activation functions 2 | 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | 7 | 8 | # SiLU https://arxiv.org/pdf/1606.08415.pdf ---------------------------------------------------------------------------- 9 | class SiLU(nn.Module): # export-friendly version of nn.SiLU() 10 | @staticmethod 11 | def forward(x): 12 | return x * torch.sigmoid(x) 13 | 14 | 15 | class Hardswish(nn.Module): # export-friendly version of nn.Hardswish() 16 | @staticmethod 17 | def forward(x): 18 | # return x * F.hardsigmoid(x) # for torchscript and CoreML 19 | return x * F.hardtanh(x + 3, 0., 6.) / 6. # for torchscript, CoreML and ONNX 20 | 21 | 22 | # Mish https://github.com/digantamisra98/Mish -------------------------------------------------------------------------- 23 | class Mish(nn.Module): 24 | @staticmethod 25 | def forward(x): 26 | return x * F.softplus(x).tanh() 27 | 28 | 29 | class MemoryEfficientMish(nn.Module): 30 | class F(torch.autograd.Function): 31 | @staticmethod 32 | def forward(ctx, x): 33 | ctx.save_for_backward(x) 34 | return x.mul(torch.tanh(F.softplus(x))) # x * tanh(ln(1 + exp(x))) 35 | 36 | @staticmethod 37 | def backward(ctx, grad_output): 38 | x = ctx.saved_tensors[0] 39 | sx = torch.sigmoid(x) 40 | fx = F.softplus(x).tanh() 41 | return grad_output * (fx + x * sx * (1 - fx * fx)) 42 | 43 | def forward(self, x): 44 | return self.F.apply(x) 45 | 46 | 47 | # FReLU https://arxiv.org/abs/2007.11824 ------------------------------------------------------------------------------- 48 | class FReLU(nn.Module): 49 | def __init__(self, c1, k=3): # ch_in, kernel 50 | super().__init__() 51 | self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False) 52 | self.bn = nn.BatchNorm2d(c1) 53 | 54 | def forward(self, x): 55 | return torch.max(x, self.bn(self.conv(x))) 56 | 57 | 58 | # ACON https://arxiv.org/pdf/2009.04759.pdf ---------------------------------------------------------------------------- 59 | class AconC(nn.Module): 60 | r""" ACON activation (activate or not). 61 | AconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is a learnable parameter 62 | according to "Activate or Not: Learning Customized Activation" . 63 | """ 64 | 65 | def __init__(self, c1): 66 | super().__init__() 67 | self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1)) 68 | self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1)) 69 | self.beta = nn.Parameter(torch.ones(1, c1, 1, 1)) 70 | 71 | def forward(self, x): 72 | dpx = (self.p1 - self.p2) * x 73 | return dpx * torch.sigmoid(self.beta * dpx) + self.p2 * x 74 | 75 | 76 | class MetaAconC(nn.Module): 77 | r""" ACON activation (activate or not). 78 | MetaAconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is generated by a small network 79 | according to "Activate or Not: Learning Customized Activation" . 80 | """ 81 | 82 | def __init__(self, c1, k=1, s=1, r=16): # ch_in, kernel, stride, r 83 | super().__init__() 84 | c2 = max(r, c1 // r) 85 | self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1)) 86 | self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1)) 87 | self.fc1 = nn.Conv2d(c1, c2, k, s, bias=True) 88 | self.fc2 = nn.Conv2d(c2, c1, k, s, bias=True) 89 | # self.bn1 = nn.BatchNorm2d(c2) 90 | # self.bn2 = nn.BatchNorm2d(c1) 91 | 92 | def forward(self, x): 93 | y = x.mean(dim=2, keepdims=True).mean(dim=3, keepdims=True) 94 | # batch-size 1 bug/instabilities https://github.com/ultralytics/yolov5/issues/2891 95 | # beta = torch.sigmoid(self.bn2(self.fc2(self.bn1(self.fc1(y))))) # bug/unstable 96 | beta = torch.sigmoid(self.fc2(self.fc1(y))) # bug patch BN layers removed 97 | dpx = (self.p1 - self.p2) * x 98 | return dpx * torch.sigmoid(beta * dpx) + self.p2 * x 99 | -------------------------------------------------------------------------------- /utils/autoanchor.py: -------------------------------------------------------------------------------- 1 | # Auto-anchor utils 2 | 3 | import numpy as np 4 | import torch 5 | import yaml 6 | from tqdm import tqdm 7 | 8 | from utils.general import colorstr 9 | 10 | 11 | def check_anchor_order(m): 12 | # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary 13 | a = m.anchor_grid.prod(-1).view(-1) # anchor area 14 | da = a[-1] - a[0] # delta a 15 | ds = m.stride[-1] - m.stride[0] # delta s 16 | if da.sign() != ds.sign(): # same order 17 | print('Reversing anchor order') 18 | m.anchors[:] = m.anchors.flip(0) 19 | m.anchor_grid[:] = m.anchor_grid.flip(0) 20 | 21 | 22 | def check_anchors(dataset, model, thr=4.0, imgsz=640): 23 | # Check anchor fit to data, recompute if necessary 24 | prefix = colorstr('autoanchor: ') 25 | print(f'\n{prefix}Analyzing anchors... ', end='') 26 | m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1] # Detect() 27 | shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True) 28 | scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1)) # augment scale 29 | wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float() # wh 30 | 31 | def metric(k): # compute metric 32 | r = wh[:, None] / k[None] 33 | x = torch.min(r, 1. / r).min(2)[0] # ratio metric 34 | best = x.max(1)[0] # best_x 35 | aat = (x > 1. / thr).float().sum(1).mean() # anchors above threshold 36 | bpr = (best > 1. / thr).float().mean() # best possible recall 37 | return bpr, aat 38 | 39 | anchors = m.anchor_grid.clone().cpu().view(-1, 2) # current anchors 40 | bpr, aat = metric(anchors) 41 | print(f'anchors/target = {aat:.2f}, Best Possible Recall (BPR) = {bpr:.4f}', end='') 42 | if bpr < 0.98: # threshold to recompute 43 | print('. Attempting to improve anchors, please wait...') 44 | na = m.anchor_grid.numel() // 2 # number of anchors 45 | try: 46 | anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False) 47 | except Exception as e: 48 | print(f'{prefix}ERROR: {e}') 49 | new_bpr = metric(anchors)[0] 50 | if new_bpr > bpr: # replace anchors 51 | anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors) 52 | m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid) # for inference 53 | m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1) # loss 54 | check_anchor_order(m) 55 | print(f'{prefix}New anchors saved to model. Update model *.yaml to use these anchors in the future.') 56 | else: 57 | print(f'{prefix}Original anchors better than new anchors. Proceeding with original anchors.') 58 | print('') # newline 59 | 60 | 61 | def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True): 62 | """ Creates kmeans-evolved anchors from training dataset 63 | 64 | Arguments: 65 | path: path to dataset *.yaml, or a loaded dataset 66 | n: number of anchors 67 | img_size: image size used for training 68 | thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0 69 | gen: generations to evolve anchors using genetic algorithm 70 | verbose: print all results 71 | 72 | Return: 73 | k: kmeans evolved anchors 74 | 75 | Usage: 76 | from utils.autoanchor import *; _ = kmean_anchors() 77 | """ 78 | from scipy.cluster.vq import kmeans 79 | 80 | thr = 1. / thr 81 | prefix = colorstr('autoanchor: ') 82 | 83 | def metric(k, wh): # compute metrics 84 | r = wh[:, None] / k[None] 85 | x = torch.min(r, 1. / r).min(2)[0] # ratio metric 86 | # x = wh_iou(wh, torch.tensor(k)) # iou metric 87 | return x, x.max(1)[0] # x, best_x 88 | 89 | def anchor_fitness(k): # mutation fitness 90 | _, best = metric(torch.tensor(k, dtype=torch.float32), wh) 91 | return (best * (best > thr).float()).mean() # fitness 92 | 93 | def print_results(k): 94 | k = k[np.argsort(k.prod(1))] # sort small to large 95 | x, best = metric(k, wh0) 96 | bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n # best possible recall, anch > thr 97 | print(f'{prefix}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr') 98 | print(f'{prefix}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, ' 99 | f'past_thr={x[x > thr].mean():.3f}-mean: ', end='') 100 | for i, x in enumerate(k): 101 | print('%i,%i' % (round(x[0]), round(x[1])), end=', ' if i < len(k) - 1 else '\n') # use in *.cfg 102 | return k 103 | 104 | if isinstance(path, str): # *.yaml file 105 | with open(path) as f: 106 | data_dict = yaml.safe_load(f) # model dict 107 | from utils.datasets import LoadImagesAndLabels 108 | dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True) 109 | else: 110 | dataset = path # dataset 111 | 112 | # Get label wh 113 | shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True) 114 | wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh 115 | 116 | # Filter 117 | i = (wh0 < 3.0).any(1).sum() 118 | if i: 119 | print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.') 120 | wh = wh0[(wh0 >= 2.0).any(1)] # filter > 2 pixels 121 | # wh = wh * (np.random.rand(wh.shape[0], 1) * 0.9 + 0.1) # multiply by random scale 0-1 122 | 123 | # Kmeans calculation 124 | print(f'{prefix}Running kmeans for {n} anchors on {len(wh)} points...') 125 | s = wh.std(0) # sigmas for whitening 126 | k, dist = kmeans(wh / s, n, iter=30) # points, mean distance 127 | assert len(k) == n, print(f'{prefix}ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len(k)}') 128 | k *= s 129 | wh = torch.tensor(wh, dtype=torch.float32) # filtered 130 | wh0 = torch.tensor(wh0, dtype=torch.float32) # unfiltered 131 | k = print_results(k) 132 | 133 | # Plot 134 | # k, d = [None] * 20, [None] * 20 135 | # for i in tqdm(range(1, 21)): 136 | # k[i-1], d[i-1] = kmeans(wh / s, i) # points, mean distance 137 | # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True) 138 | # ax = ax.ravel() 139 | # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.') 140 | # fig, ax = plt.subplots(1, 2, figsize=(14, 7)) # plot wh 141 | # ax[0].hist(wh[wh[:, 0]<100, 0],400) 142 | # ax[1].hist(wh[wh[:, 1]<100, 1],400) 143 | # fig.savefig('wh.png', dpi=200) 144 | 145 | # Evolve 146 | npr = np.random 147 | f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1 # fitness, generations, mutation prob, sigma 148 | pbar = tqdm(range(gen), desc=f'{prefix}Evolving anchors with Genetic Algorithm:') # progress bar 149 | for _ in pbar: 150 | v = np.ones(sh) 151 | while (v == 1).all(): # mutate until a change occurs (prevent duplicates) 152 | v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0) 153 | kg = (k.copy() * v).clip(min=2.0) 154 | fg = anchor_fitness(kg) 155 | if fg > f: 156 | f, k = fg, kg.copy() 157 | pbar.desc = f'{prefix}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}' 158 | if verbose: 159 | print_results(k) 160 | 161 | return print_results(k) 162 | -------------------------------------------------------------------------------- /utils/aws/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/utils/aws/__init__.py -------------------------------------------------------------------------------- /utils/aws/mime.sh: -------------------------------------------------------------------------------- 1 | # AWS EC2 instance startup 'MIME' script https://aws.amazon.com/premiumsupport/knowledge-center/execute-user-data-ec2/ 2 | # This script will run on every instance restart, not only on first start 3 | # --- DO NOT COPY ABOVE COMMENTS WHEN PASTING INTO USERDATA --- 4 | 5 | Content-Type: multipart/mixed; boundary="//" 6 | MIME-Version: 1.0 7 | 8 | --// 9 | Content-Type: text/cloud-config; charset="us-ascii" 10 | MIME-Version: 1.0 11 | Content-Transfer-Encoding: 7bit 12 | Content-Disposition: attachment; filename="cloud-config.txt" 13 | 14 | #cloud-config 15 | cloud_final_modules: 16 | - [scripts-user, always] 17 | 18 | --// 19 | Content-Type: text/x-shellscript; charset="us-ascii" 20 | MIME-Version: 1.0 21 | Content-Transfer-Encoding: 7bit 22 | Content-Disposition: attachment; filename="userdata.txt" 23 | 24 | #!/bin/bash 25 | # --- paste contents of userdata.sh here --- 26 | --// 27 | -------------------------------------------------------------------------------- /utils/aws/resume.py: -------------------------------------------------------------------------------- 1 | # Resume all interrupted trainings in yolov5/ dir including DDP trainings 2 | # Usage: $ python utils/aws/resume.py 3 | 4 | import os 5 | import sys 6 | from pathlib import Path 7 | 8 | import torch 9 | import yaml 10 | 11 | sys.path.append('./') # to run '$ python *.py' files in subdirectories 12 | 13 | port = 0 # --master_port 14 | path = Path('').resolve() 15 | for last in path.rglob('*/**/last.pt'): 16 | ckpt = torch.load(last) 17 | if ckpt['optimizer'] is None: 18 | continue 19 | 20 | # Load opt.yaml 21 | with open(last.parent.parent / 'opt.yaml') as f: 22 | opt = yaml.safe_load(f) 23 | 24 | # Get device count 25 | d = opt['device'].split(',') # devices 26 | nd = len(d) # number of devices 27 | ddp = nd > 1 or (nd == 0 and torch.cuda.device_count() > 1) # distributed data parallel 28 | 29 | if ddp: # multi-GPU 30 | port += 1 31 | cmd = f'python -m torch.distributed.launch --nproc_per_node {nd} --master_port {port} train.py --resume {last}' 32 | else: # single-GPU 33 | cmd = f'python train.py --resume {last}' 34 | 35 | cmd += ' > /dev/null 2>&1 &' # redirect output to dev/null and run in daemon thread 36 | print(cmd) 37 | os.system(cmd) 38 | -------------------------------------------------------------------------------- /utils/aws/userdata.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # AWS EC2 instance startup script https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/user-data.html 3 | # This script will run only once on first instance start (for a re-start script see mime.sh) 4 | # /home/ubuntu (ubuntu) or /home/ec2-user (amazon-linux) is working dir 5 | # Use >300 GB SSD 6 | 7 | cd home/ubuntu 8 | if [ ! -d yolov5 ]; then 9 | echo "Running first-time script." # install dependencies, download COCO, pull Docker 10 | git clone https://github.com/ultralytics/yolov5 && sudo chmod -R 777 yolov5 11 | cd yolov5 12 | bash data/scripts/get_coco.sh && echo "Data done." & 13 | sudo docker pull ultralytics/yolov5:latest && echo "Docker done." & 14 | python -m pip install --upgrade pip && pip install -r requirements.txt && python detect.py && echo "Requirements done." & 15 | wait && echo "All tasks done." # finish background tasks 16 | else 17 | echo "Running re-start script." # resume interrupted runs 18 | i=0 19 | list=$(sudo docker ps -qa) # container list i.e. $'one\ntwo\nthree\nfour' 20 | while IFS= read -r id; do 21 | ((i++)) 22 | echo "restarting container $i: $id" 23 | sudo docker start $id 24 | # sudo docker exec -it $id python train.py --resume # single-GPU 25 | sudo docker exec -d $id python utils/aws/resume.py # multi-scenario 26 | done <<<"$list" 27 | fi 28 | -------------------------------------------------------------------------------- /utils/flask_rest_api/README.md: -------------------------------------------------------------------------------- 1 | # Flask REST API 2 | [REST](https://en.wikipedia.org/wiki/Representational_state_transfer) [API](https://en.wikipedia.org/wiki/API)s are commonly used to expose Machine Learning (ML) models to other services. This folder contains an example REST API created using Flask to expose the `yolov5s` model from [PyTorch Hub](https://pytorch.org/hub/ultralytics_yolov5/). 3 | 4 | ## Requirements 5 | 6 | [Flask](https://palletsprojects.com/p/flask/) is required. Install with: 7 | ```shell 8 | $ pip install Flask 9 | ``` 10 | 11 | ## Run 12 | 13 | After Flask installation run: 14 | 15 | ```shell 16 | $ python3 restapi.py --port 5000 17 | ``` 18 | 19 | Then use [curl](https://curl.se/) to perform a request: 20 | 21 | ```shell 22 | $ curl -X POST -F image=@zidane.jpg 'http://localhost:5000/v1/object-detection/yolov5s'` 23 | ``` 24 | 25 | The model inference results are returned: 26 | 27 | ```shell 28 | [{'class': 0, 29 | 'confidence': 0.8197850585, 30 | 'name': 'person', 31 | 'xmax': 1159.1403808594, 32 | 'xmin': 750.912902832, 33 | 'ymax': 711.2583007812, 34 | 'ymin': 44.0350036621}, 35 | {'class': 0, 36 | 'confidence': 0.5667674541, 37 | 'name': 'person', 38 | 'xmax': 1065.5523681641, 39 | 'xmin': 116.0448303223, 40 | 'ymax': 713.8904418945, 41 | 'ymin': 198.4603881836}, 42 | {'class': 27, 43 | 'confidence': 0.5661227107, 44 | 'name': 'tie', 45 | 'xmax': 516.7975463867, 46 | 'xmin': 416.6880187988, 47 | 'ymax': 717.0524902344, 48 | 'ymin': 429.2020568848}] 49 | ``` 50 | 51 | An example python script to perform inference using [requests](https://docs.python-requests.org/en/master/) is given in `example_request.py` 52 | -------------------------------------------------------------------------------- /utils/flask_rest_api/example_request.py: -------------------------------------------------------------------------------- 1 | """Perform test request""" 2 | import pprint 3 | 4 | import requests 5 | 6 | DETECTION_URL = "http://localhost:5000/v1/object-detection/yolov5s" 7 | TEST_IMAGE = "zidane.jpg" 8 | 9 | image_data = open(TEST_IMAGE, "rb").read() 10 | 11 | response = requests.post(DETECTION_URL, files={"image": image_data}).json() 12 | 13 | pprint.pprint(response) 14 | -------------------------------------------------------------------------------- /utils/flask_rest_api/restapi.py: -------------------------------------------------------------------------------- 1 | """ 2 | Run a rest API exposing the yolov5s object detection model 3 | """ 4 | import argparse 5 | import io 6 | 7 | import torch 8 | from PIL import Image 9 | from flask import Flask, request 10 | 11 | app = Flask(__name__) 12 | 13 | DETECTION_URL = "/v1/object-detection/yolov5s" 14 | 15 | 16 | @app.route(DETECTION_URL, methods=["POST"]) 17 | def predict(): 18 | if not request.method == "POST": 19 | return 20 | 21 | if request.files.get("image"): 22 | image_file = request.files["image"] 23 | image_bytes = image_file.read() 24 | 25 | img = Image.open(io.BytesIO(image_bytes)) 26 | 27 | results = model(img, size=640) # reduce size=320 for faster inference 28 | return results.pandas().xyxy[0].to_json(orient="records") 29 | 30 | 31 | if __name__ == "__main__": 32 | parser = argparse.ArgumentParser(description="Flask API exposing YOLOv5 model") 33 | parser.add_argument("--port", default=5000, type=int, help="port number") 34 | args = parser.parse_args() 35 | 36 | model = torch.hub.load("ultralytics/yolov5", "yolov5s", force_reload=True) # force_reload to recache 37 | app.run(host="0.0.0.0", port=args.port) # debug=True causes Restarting with stat 38 | -------------------------------------------------------------------------------- /utils/google_app_engine/Dockerfile: -------------------------------------------------------------------------------- 1 | FROM gcr.io/google-appengine/python 2 | 3 | # Create a virtualenv for dependencies. This isolates these packages from 4 | # system-level packages. 5 | # Use -p python3 or -p python3.7 to select python version. Default is version 2. 6 | RUN virtualenv /env -p python3 7 | 8 | # Setting these environment variables are the same as running 9 | # source /env/bin/activate. 10 | ENV VIRTUAL_ENV /env 11 | ENV PATH /env/bin:$PATH 12 | 13 | RUN apt-get update && apt-get install -y python-opencv 14 | 15 | # Copy the application's requirements.txt and run pip to install all 16 | # dependencies into the virtualenv. 17 | ADD requirements.txt /app/requirements.txt 18 | RUN pip install -r /app/requirements.txt 19 | 20 | # Add the application source code. 21 | ADD . /app 22 | 23 | # Run a WSGI server to serve the application. gunicorn must be declared as 24 | # a dependency in requirements.txt. 25 | CMD gunicorn -b :$PORT main:app 26 | -------------------------------------------------------------------------------- /utils/google_app_engine/additional_requirements.txt: -------------------------------------------------------------------------------- 1 | # add these requirements in your app on top of the existing ones 2 | pip==18.1 3 | Flask==1.0.2 4 | gunicorn==19.9.0 5 | -------------------------------------------------------------------------------- /utils/google_app_engine/app.yaml: -------------------------------------------------------------------------------- 1 | runtime: custom 2 | env: flex 3 | 4 | service: yolov5app 5 | 6 | liveness_check: 7 | initial_delay_sec: 600 8 | 9 | manual_scaling: 10 | instances: 1 11 | resources: 12 | cpu: 1 13 | memory_gb: 4 14 | disk_size_gb: 20 -------------------------------------------------------------------------------- /utils/google_utils.py: -------------------------------------------------------------------------------- 1 | # Google utils: https://cloud.google.com/storage/docs/reference/libraries 2 | 3 | import os 4 | import platform 5 | import subprocess 6 | import time 7 | from pathlib import Path 8 | 9 | import requests 10 | import torch 11 | 12 | 13 | def gsutil_getsize(url=''): 14 | # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du 15 | s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8') 16 | return eval(s.split(' ')[0]) if len(s) else 0 # bytes 17 | 18 | 19 | def attempt_download(file, repo='ultralytics/yolov5'): 20 | # Attempt file download if does not exist 21 | file = Path(str(file).strip().replace("'", '')) 22 | 23 | if not file.exists(): 24 | file.parent.mkdir(parents=True, exist_ok=True) # make parent dir (if required) 25 | try: 26 | response = requests.get(f'https://api.github.com/repos/{repo}/releases/latest').json() # github api 27 | assets = [x['name'] for x in response['assets']] # release assets, i.e. ['yolov5s.pt', 'yolov5m.pt', ...] 28 | tag = response['tag_name'] # i.e. 'v1.0' 29 | except: # fallback plan 30 | assets = ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt', 31 | 'yolov5s6.pt', 'yolov5m6.pt', 'yolov5l6.pt', 'yolov5x6.pt'] 32 | try: 33 | tag = subprocess.check_output('git tag', shell=True, stderr=subprocess.STDOUT).decode().split()[-1] 34 | except: 35 | tag = 'v5.0' # current release 36 | 37 | name = file.name 38 | if name in assets: 39 | msg = f'{file} missing, try downloading from https://github.com/{repo}/releases/' 40 | redundant = False # second download option 41 | try: # GitHub 42 | url = f'https://github.com/{repo}/releases/download/{tag}/{name}' 43 | print(f'Downloading {url} to {file}...') 44 | torch.hub.download_url_to_file(url, file) 45 | assert file.exists() and file.stat().st_size > 1E6 # check 46 | except Exception as e: # GCP 47 | print(f'Download error: {e}') 48 | assert redundant, 'No secondary mirror' 49 | url = f'https://storage.googleapis.com/{repo}/ckpt/{name}' 50 | print(f'Downloading {url} to {file}...') 51 | os.system(f"curl -L '{url}' -o '{file}' --retry 3 -C -") # curl download, retry and resume on fail 52 | finally: 53 | if not file.exists() or file.stat().st_size < 1E6: # check 54 | file.unlink(missing_ok=True) # remove partial downloads 55 | print(f'ERROR: Download failure: {msg}') 56 | print('') 57 | return 58 | 59 | 60 | def gdrive_download(id='16TiPfZj7htmTyhntwcZyEEAejOUxuT6m', file='tmp.zip'): 61 | # Downloads a file from Google Drive. from yolov5.utils.google_utils import *; gdrive_download() 62 | t = time.time() 63 | file = Path(file) 64 | cookie = Path('cookie') # gdrive cookie 65 | print(f'Downloading https://drive.google.com/uc?export=download&id={id} as {file}... ', end='') 66 | file.unlink(missing_ok=True) # remove existing file 67 | cookie.unlink(missing_ok=True) # remove existing cookie 68 | 69 | # Attempt file download 70 | out = "NUL" if platform.system() == "Windows" else "/dev/null" 71 | os.system(f'curl -c ./cookie -s -L "drive.google.com/uc?export=download&id={id}" > {out}') 72 | if os.path.exists('cookie'): # large file 73 | s = f'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm={get_token()}&id={id}" -o {file}' 74 | else: # small file 75 | s = f'curl -s -L -o {file} "drive.google.com/uc?export=download&id={id}"' 76 | r = os.system(s) # execute, capture return 77 | cookie.unlink(missing_ok=True) # remove existing cookie 78 | 79 | # Error check 80 | if r != 0: 81 | file.unlink(missing_ok=True) # remove partial 82 | print('Download error ') # raise Exception('Download error') 83 | return r 84 | 85 | # Unzip if archive 86 | if file.suffix == '.zip': 87 | print('unzipping... ', end='') 88 | os.system(f'unzip -q {file}') # unzip 89 | file.unlink() # remove zip to free space 90 | 91 | print(f'Done ({time.time() - t:.1f}s)') 92 | return r 93 | 94 | 95 | def get_token(cookie="./cookie"): 96 | with open(cookie) as f: 97 | for line in f: 98 | if "download" in line: 99 | return line.split()[-1] 100 | return "" 101 | 102 | # def upload_blob(bucket_name, source_file_name, destination_blob_name): 103 | # # Uploads a file to a bucket 104 | # # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python 105 | # 106 | # storage_client = storage.Client() 107 | # bucket = storage_client.get_bucket(bucket_name) 108 | # blob = bucket.blob(destination_blob_name) 109 | # 110 | # blob.upload_from_filename(source_file_name) 111 | # 112 | # print('File {} uploaded to {}.'.format( 113 | # source_file_name, 114 | # destination_blob_name)) 115 | # 116 | # 117 | # def download_blob(bucket_name, source_blob_name, destination_file_name): 118 | # # Uploads a blob from a bucket 119 | # storage_client = storage.Client() 120 | # bucket = storage_client.get_bucket(bucket_name) 121 | # blob = bucket.blob(source_blob_name) 122 | # 123 | # blob.download_to_filename(destination_file_name) 124 | # 125 | # print('Blob {} downloaded to {}.'.format( 126 | # source_blob_name, 127 | # destination_file_name)) 128 | -------------------------------------------------------------------------------- /utils/loss.py: -------------------------------------------------------------------------------- 1 | # Loss functions 2 | 3 | import torch 4 | import torch.nn as nn 5 | 6 | from utils.general import bbox_iou 7 | from utils.torch_utils import is_parallel 8 | 9 | 10 | def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441 11 | # return positive, negative label smoothing BCE targets 12 | return 1.0 - 0.5 * eps, 0.5 * eps 13 | 14 | 15 | class BCEBlurWithLogitsLoss(nn.Module): 16 | # BCEwithLogitLoss() with reduced missing label effects. 17 | def __init__(self, alpha=0.05): 18 | super(BCEBlurWithLogitsLoss, self).__init__() 19 | self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss() 20 | self.alpha = alpha 21 | 22 | def forward(self, pred, true): 23 | loss = self.loss_fcn(pred, true) 24 | pred = torch.sigmoid(pred) # prob from logits 25 | dx = pred - true # reduce only missing label effects 26 | # dx = (pred - true).abs() # reduce missing label and false label effects 27 | alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4)) 28 | loss *= alpha_factor 29 | return loss.mean() 30 | 31 | 32 | class FocalLoss(nn.Module): 33 | # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) 34 | def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): 35 | super(FocalLoss, self).__init__() 36 | self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() 37 | self.gamma = gamma 38 | self.alpha = alpha 39 | self.reduction = loss_fcn.reduction 40 | self.loss_fcn.reduction = 'none' # required to apply FL to each element 41 | 42 | def forward(self, pred, true): 43 | loss = self.loss_fcn(pred, true) 44 | # p_t = torch.exp(-loss) 45 | # loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability 46 | 47 | # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py 48 | pred_prob = torch.sigmoid(pred) # prob from logits 49 | p_t = true * pred_prob + (1 - true) * (1 - pred_prob) 50 | alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) 51 | modulating_factor = (1.0 - p_t) ** self.gamma 52 | loss *= alpha_factor * modulating_factor 53 | 54 | if self.reduction == 'mean': 55 | return loss.mean() 56 | elif self.reduction == 'sum': 57 | return loss.sum() 58 | else: # 'none' 59 | return loss 60 | 61 | 62 | class QFocalLoss(nn.Module): 63 | # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) 64 | def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): 65 | super(QFocalLoss, self).__init__() 66 | self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() 67 | self.gamma = gamma 68 | self.alpha = alpha 69 | self.reduction = loss_fcn.reduction 70 | self.loss_fcn.reduction = 'none' # required to apply FL to each element 71 | 72 | def forward(self, pred, true): 73 | loss = self.loss_fcn(pred, true) 74 | 75 | pred_prob = torch.sigmoid(pred) # prob from logits 76 | alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) 77 | modulating_factor = torch.abs(true - pred_prob) ** self.gamma 78 | loss *= alpha_factor * modulating_factor 79 | 80 | if self.reduction == 'mean': 81 | return loss.mean() 82 | elif self.reduction == 'sum': 83 | return loss.sum() 84 | else: # 'none' 85 | return loss 86 | 87 | 88 | class ComputeLoss: 89 | # Compute losses 90 | def __init__(self, model, autobalance=False): 91 | super(ComputeLoss, self).__init__() 92 | device = next(model.parameters()).device # get model device 93 | h = model.hyp # hyperparameters 94 | 95 | # Define criteria 96 | BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device)) 97 | BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device)) 98 | 99 | # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3 100 | self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets 101 | 102 | # Focal loss 103 | g = h['fl_gamma'] # focal loss gamma 104 | if g > 0: 105 | BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g) 106 | 107 | det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module 108 | self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02]) # P3-P7 109 | self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index 110 | self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance 111 | for k in 'na', 'nc', 'nl', 'anchors': 112 | setattr(self, k, getattr(det, k)) 113 | 114 | def __call__(self, p, targets): # predictions, targets, model 115 | device = targets.device 116 | lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device) 117 | tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets 118 | 119 | # Losses 120 | for i, pi in enumerate(p): # layer index, layer predictions 121 | b, a, gj, gi = indices[i] # image, anchor, gridy, gridx 122 | tobj = torch.zeros_like(pi[..., 0], device=device) # target obj 123 | 124 | n = b.shape[0] # number of targets 125 | if n: 126 | ps = pi[b, a, gj, gi] # prediction subset corresponding to targets 127 | 128 | # Regression 129 | pxy = ps[:, :2].sigmoid() * 2. - 0.5 130 | pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i] 131 | pbox = torch.cat((pxy, pwh), 1) # predicted box 132 | iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True) # iou(prediction, target) 133 | lbox += (1.0 - iou).mean() # iou loss 134 | 135 | # Objectness 136 | tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio 137 | 138 | # Classification 139 | if self.nc > 1: # cls loss (only if multiple classes) 140 | t = torch.full_like(ps[:, 5:], self.cn, device=device) # targets 141 | t[range(n), tcls[i]] = self.cp 142 | lcls += self.BCEcls(ps[:, 5:], t) # BCE 143 | 144 | # Append targets to text file 145 | # with open('targets.txt', 'a') as file: 146 | # [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)] 147 | 148 | obji = self.BCEobj(pi[..., 4], tobj) 149 | lobj += obji * self.balance[i] # obj loss 150 | if self.autobalance: 151 | self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item() 152 | 153 | if self.autobalance: 154 | self.balance = [x / self.balance[self.ssi] for x in self.balance] 155 | lbox *= self.hyp['box'] 156 | lobj *= self.hyp['obj'] 157 | lcls *= self.hyp['cls'] 158 | bs = tobj.shape[0] # batch size 159 | 160 | loss = lbox + lobj + lcls 161 | return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach() 162 | 163 | def build_targets(self, p, targets): 164 | # Build targets for compute_loss(), input targets(image,class,x,y,w,h) 165 | na, nt = self.na, targets.shape[0] # number of anchors, targets 166 | tcls, tbox, indices, anch = [], [], [], [] 167 | gain = torch.ones(7, device=targets.device) # normalized to gridspace gain 168 | ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt) 169 | targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices 170 | 171 | g = 0.5 # bias 172 | off = torch.tensor([[0, 0], 173 | [1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m 174 | # [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm 175 | ], device=targets.device).float() * g # offsets 176 | 177 | for i in range(self.nl): 178 | anchors = self.anchors[i] 179 | gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain 180 | 181 | # Match targets to anchors 182 | t = targets * gain 183 | if nt: 184 | # Matches 185 | r = t[:, :, 4:6] / anchors[:, None] # wh ratio 186 | j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t'] # compare 187 | # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2)) 188 | t = t[j] # filter 189 | 190 | # Offsets 191 | gxy = t[:, 2:4] # grid xy 192 | gxi = gain[[2, 3]] - gxy # inverse 193 | j, k = ((gxy % 1. < g) & (gxy > 1.)).T 194 | l, m = ((gxi % 1. < g) & (gxi > 1.)).T 195 | j = torch.stack((torch.ones_like(j), j, k, l, m)) 196 | t = t.repeat((5, 1, 1))[j] 197 | offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j] 198 | else: 199 | t = targets[0] 200 | offsets = 0 201 | 202 | # Define 203 | b, c = t[:, :2].long().T # image, class 204 | gxy = t[:, 2:4] # grid xy 205 | gwh = t[:, 4:6] # grid wh 206 | gij = (gxy - offsets).long() 207 | gi, gj = gij.T # grid xy indices 208 | 209 | # Append 210 | a = t[:, 6].long() # anchor indices 211 | indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices 212 | tbox.append(torch.cat((gxy - gij, gwh), 1)) # box 213 | anch.append(anchors[a]) # anchors 214 | tcls.append(c) # class 215 | 216 | return tcls, tbox, indices, anch 217 | -------------------------------------------------------------------------------- /utils/metrics.py: -------------------------------------------------------------------------------- 1 | # Model validation metrics 2 | 3 | from pathlib import Path 4 | 5 | import matplotlib.pyplot as plt 6 | import numpy as np 7 | import torch 8 | 9 | from . import general 10 | 11 | 12 | def fitness(x): 13 | # Model fitness as a weighted combination of metrics 14 | w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95] 15 | return (x[:, :4] * w).sum(1) 16 | 17 | 18 | def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=()): 19 | """ Compute the average precision, given the recall and precision curves. 20 | Source: https://github.com/rafaelpadilla/Object-Detection-Metrics. 21 | # Arguments 22 | tp: True positives (nparray, nx1 or nx10). 23 | conf: Objectness value from 0-1 (nparray). 24 | pred_cls: Predicted object classes (nparray). 25 | target_cls: True object classes (nparray). 26 | plot: Plot precision-recall curve at mAP@0.5 27 | save_dir: Plot save directory 28 | # Returns 29 | The average precision as computed in py-faster-rcnn. 30 | """ 31 | 32 | # Sort by objectness 33 | i = np.argsort(-conf) 34 | tp, conf, pred_cls = tp[i], conf[i], pred_cls[i] 35 | 36 | # Find unique classes 37 | unique_classes = np.unique(target_cls) 38 | nc = unique_classes.shape[0] # number of classes, number of detections 39 | 40 | # Create Precision-Recall curve and compute AP for each class 41 | px, py = np.linspace(0, 1, 1000), [] # for plotting 42 | ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000)) 43 | for ci, c in enumerate(unique_classes): 44 | i = pred_cls == c 45 | n_l = (target_cls == c).sum() # number of labels 46 | n_p = i.sum() # number of predictions 47 | 48 | if n_p == 0 or n_l == 0: 49 | continue 50 | else: 51 | # Accumulate FPs and TPs 52 | fpc = (1 - tp[i]).cumsum(0) 53 | tpc = tp[i].cumsum(0) 54 | 55 | # Recall 56 | recall = tpc / (n_l + 1e-16) # recall curve 57 | r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0) # negative x, xp because xp decreases 58 | 59 | # Precision 60 | precision = tpc / (tpc + fpc) # precision curve 61 | p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1) # p at pr_score 62 | 63 | # AP from recall-precision curve 64 | for j in range(tp.shape[1]): 65 | ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j]) 66 | if plot and j == 0: 67 | py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5 68 | 69 | # Compute F1 (harmonic mean of precision and recall) 70 | f1 = 2 * p * r / (p + r + 1e-16) 71 | if plot: 72 | plot_pr_curve(px, py, ap, Path(save_dir) / 'PR_curve.png', names) 73 | plot_mc_curve(px, f1, Path(save_dir) / 'F1_curve.png', names, ylabel='F1') 74 | plot_mc_curve(px, p, Path(save_dir) / 'P_curve.png', names, ylabel='Precision') 75 | plot_mc_curve(px, r, Path(save_dir) / 'R_curve.png', names, ylabel='Recall') 76 | 77 | i = f1.mean(0).argmax() # max F1 index 78 | return p[:, i], r[:, i], ap, f1[:, i], unique_classes.astype('int32') 79 | 80 | 81 | def compute_ap(recall, precision): 82 | """ Compute the average precision, given the recall and precision curves 83 | # Arguments 84 | recall: The recall curve (list) 85 | precision: The precision curve (list) 86 | # Returns 87 | Average precision, precision curve, recall curve 88 | """ 89 | 90 | # Append sentinel values to beginning and end 91 | mrec = np.concatenate(([0.], recall, [recall[-1] + 0.01])) 92 | mpre = np.concatenate(([1.], precision, [0.])) 93 | 94 | # Compute the precision envelope 95 | mpre = np.flip(np.maximum.accumulate(np.flip(mpre))) 96 | 97 | # Integrate area under curve 98 | method = 'interp' # methods: 'continuous', 'interp' 99 | if method == 'interp': 100 | x = np.linspace(0, 1, 101) # 101-point interp (COCO) 101 | ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate 102 | else: # 'continuous' 103 | i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes 104 | ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve 105 | 106 | return ap, mpre, mrec 107 | 108 | 109 | class ConfusionMatrix: 110 | # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix 111 | def __init__(self, nc, conf=0.25, iou_thres=0.45): 112 | self.matrix = np.zeros((nc + 1, nc + 1)) 113 | self.nc = nc # number of classes 114 | self.conf = conf 115 | self.iou_thres = iou_thres 116 | 117 | def process_batch(self, detections, labels): 118 | """ 119 | Return intersection-over-union (Jaccard index) of boxes. 120 | Both sets of boxes are expected to be in (x1, y1, x2, y2) format. 121 | Arguments: 122 | detections (Array[N, 6]), x1, y1, x2, y2, conf, class 123 | labels (Array[M, 5]), class, x1, y1, x2, y2 124 | Returns: 125 | None, updates confusion matrix accordingly 126 | """ 127 | detections = detections[detections[:, 4] > self.conf] 128 | gt_classes = labels[:, 0].int() 129 | detection_classes = detections[:, 5].int() 130 | iou = general.box_iou(labels[:, 1:], detections[:, :4]) 131 | 132 | x = torch.where(iou > self.iou_thres) 133 | if x[0].shape[0]: 134 | matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() 135 | if x[0].shape[0] > 1: 136 | matches = matches[matches[:, 2].argsort()[::-1]] 137 | matches = matches[np.unique(matches[:, 1], return_index=True)[1]] 138 | matches = matches[matches[:, 2].argsort()[::-1]] 139 | matches = matches[np.unique(matches[:, 0], return_index=True)[1]] 140 | else: 141 | matches = np.zeros((0, 3)) 142 | 143 | n = matches.shape[0] > 0 144 | m0, m1, _ = matches.transpose().astype(np.int16) 145 | for i, gc in enumerate(gt_classes): 146 | j = m0 == i 147 | if n and sum(j) == 1: 148 | self.matrix[detection_classes[m1[j]], gc] += 1 # correct 149 | else: 150 | self.matrix[self.nc, gc] += 1 # background FP 151 | 152 | if n: 153 | for i, dc in enumerate(detection_classes): 154 | if not any(m1 == i): 155 | self.matrix[dc, self.nc] += 1 # background FN 156 | 157 | def matrix(self): 158 | return self.matrix 159 | 160 | def plot(self, save_dir='', names=()): 161 | try: 162 | import seaborn as sn 163 | 164 | array = self.matrix / (self.matrix.sum(0).reshape(1, self.nc + 1) + 1E-6) # normalize 165 | array[array < 0.005] = np.nan # don't annotate (would appear as 0.00) 166 | 167 | fig = plt.figure(figsize=(12, 9), tight_layout=True) 168 | sn.set(font_scale=1.0 if self.nc < 50 else 0.8) # for label size 169 | labels = (0 < len(names) < 99) and len(names) == self.nc # apply names to ticklabels 170 | sn.heatmap(array, annot=self.nc < 30, annot_kws={"size": 8}, cmap='Blues', fmt='.2f', square=True, 171 | xticklabels=names + ['background FP'] if labels else "auto", 172 | yticklabels=names + ['background FN'] if labels else "auto").set_facecolor((1, 1, 1)) 173 | fig.axes[0].set_xlabel('True') 174 | fig.axes[0].set_ylabel('Predicted') 175 | fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250) 176 | except Exception as e: 177 | pass 178 | 179 | def print(self): 180 | for i in range(self.nc + 1): 181 | print(' '.join(map(str, self.matrix[i]))) 182 | 183 | 184 | # Plots ---------------------------------------------------------------------------------------------------------------- 185 | 186 | def plot_pr_curve(px, py, ap, save_dir='pr_curve.png', names=()): 187 | # Precision-recall curve 188 | fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) 189 | py = np.stack(py, axis=1) 190 | 191 | if 0 < len(names) < 21: # display per-class legend if < 21 classes 192 | for i, y in enumerate(py.T): 193 | ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}') # plot(recall, precision) 194 | else: 195 | ax.plot(px, py, linewidth=1, color='grey') # plot(recall, precision) 196 | 197 | ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean()) 198 | ax.set_xlabel('Recall') 199 | ax.set_ylabel('Precision') 200 | ax.set_xlim(0, 1) 201 | ax.set_ylim(0, 1) 202 | plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") 203 | fig.savefig(Path(save_dir), dpi=250) 204 | 205 | 206 | def plot_mc_curve(px, py, save_dir='mc_curve.png', names=(), xlabel='Confidence', ylabel='Metric'): 207 | # Metric-confidence curve 208 | fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) 209 | 210 | if 0 < len(names) < 21: # display per-class legend if < 21 classes 211 | for i, y in enumerate(py): 212 | ax.plot(px, y, linewidth=1, label=f'{names[i]}') # plot(confidence, metric) 213 | else: 214 | ax.plot(px, py.T, linewidth=1, color='grey') # plot(confidence, metric) 215 | 216 | y = py.mean(0) 217 | ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}') 218 | ax.set_xlabel(xlabel) 219 | ax.set_ylabel(ylabel) 220 | ax.set_xlim(0, 1) 221 | ax.set_ylim(0, 1) 222 | plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") 223 | fig.savefig(Path(save_dir), dpi=250) 224 | -------------------------------------------------------------------------------- /utils/wandb_logging/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/midasklr/yolov5prune/8a0eff3edd2225ef9e894c72f1a9d978de37b042/utils/wandb_logging/__init__.py -------------------------------------------------------------------------------- /utils/wandb_logging/log_dataset.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | 3 | import yaml 4 | 5 | from wandb_utils import WandbLogger 6 | 7 | WANDB_ARTIFACT_PREFIX = 'wandb-artifact://' 8 | 9 | 10 | def create_dataset_artifact(opt): 11 | with open(opt.data) as f: 12 | data = yaml.safe_load(f) # data dict 13 | logger = WandbLogger(opt, '', None, data, job_type='Dataset Creation') 14 | 15 | 16 | if __name__ == '__main__': 17 | parser = argparse.ArgumentParser() 18 | parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path') 19 | parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset') 20 | parser.add_argument('--project', type=str, default='YOLOv5', help='name of W&B Project') 21 | opt = parser.parse_args() 22 | opt.resume = False # Explicitly disallow resume check for dataset upload job 23 | 24 | create_dataset_artifact(opt) 25 | -------------------------------------------------------------------------------- /weights/download_weights.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Download latest models from https://github.com/ultralytics/yolov5/releases 3 | # Usage: 4 | # $ bash weights/download_weights.sh 5 | 6 | python - <