yolov5模型修剪/稀疏性
在你开始前
克隆 YOLOv5 存储库并安装requirements.txt依赖项，包括Python>=3.8和PyTorch>=1.7。
$ git clone https://github.com/ultralytics/yolov5 # clone repo
$ cd yolov5
$ pip install -r requirements.txt # install
在 COCO 上测试 YOLOv5x（默认）
此命令在 COCO val2017 上以 640 像素的图像大小测试 YOLOv5x 以建立标称基线。yolov5x.pt是可用的最大和最准确的模型。其他选项是yolov5s.pt, yolov5m.ptand yolov5l.pt, 或者您拥有训练自定义数据集的检查点./weights/best.pt。有关所有可用模型的详细信息，请参阅我们的自述文件表。
$ python test.py --weights yolov5x.pt --data coco.yaml --img 640 --iou 0.65
默认输出：
YOLOv5 🚀 v4.0-174-g9c803f2 torch 1.8.0+cu101 CUDA:0 (Tesla V100-SXM2-16GB, 16160.5MB)
Fusing layers...
Model Summary: 476 layers, 87730285 parameters, 0 gradients, 218.8 GFLOPS
val: Scanning '../coco/val2017.cache' images and labels... 4952 found, 48 missing, 0 empty, 0 corrupted: 100% 5000/5000 [00:00<00:00, 50901747.57it/s]
Class Images Labels P R mAP@.5 mAP@.5:.95: 100% 157/157 [01:20<00:00, 1.95it/s]
all 5000 36335 0.749 0.619 0.68 0.486
Speed: 5.2/1.6/6.8 ms inference/NMS/total per 640x640 image at batch-size 32 < -------- speed
Evaluating pycocotools mAP... saving runs/test/exp/yolov5x_predictions.json...
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.501 < -------- mAP
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.687
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.544
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.338
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.548
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.637
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.378 < -------- mAR
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.628
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.680
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.520
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.729
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.826
Results saved to runs/test/exp
在 COCO 上测试 YOLOv5x（0.30 稀疏度）
我们使用torch_utils.prune()命令对修剪后的模型重复上述测试。我们更新test.py以将 YOLOv5x 修剪为 0.3 稀疏度：
30% 修剪输出：
YOLOv5 🚀 v4.0-174-g9c803f2 torch 1.8.0+cu101 CUDA:0 (Tesla V100-SXM2-16GB, 16160.5MB)
Fusing layers...
Model Summary: 476 layers, 87730285 parameters, 0 gradients, 218.8 GFLOPS
Pruning model... 0.3 global sparsity
val: Scanning '../coco/val2017.cache' images and labels... 4952 found, 48 missing, 0 empty, 0 corrupted: 100% 5000/5000 [00:00<00:00, 50901747.57it/s]
Class Images Labels P R mAP@.5 mAP@.5:.95: 100% 157/157 [01:17<00:00, 2.01it/s]
all 5000 36335 0.742 0.595 0.66 0.46
Speed: 5.2/1.5/6.7 ms inference/NMS/total per 640x640 image at batch-size 32 < -------- speed
Evaluating pycocotools mAP... saving runs/test/exp/yolov5x_predictions.json...
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.477 < -------- mAP
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.667
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.525
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.313
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.530
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.619
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.366 < -------- mAR
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.603
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.654
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.475
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.709
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.794
Results saved to runs/test/exp
在结果中我们可以观察到，我们的模型在剪枝后达到了30%的稀疏性，这意味着模型nn.Conv2d层中30% 的权重参数等于 0。推理时间基本不变，而模型的AP 和 AR分数略有降低。
评论