{"id":1825,"date":"2023-08-23T11:37:35","date_gmt":"2023-08-23T03:37:35","guid":{"rendered":"https:\/\/usei.cn\/?p=1825"},"modified":"2023-09-08T17:14:12","modified_gmt":"2023-09-08T09:14:12","slug":"paddledetection-%e6%ba%90%e7%a0%81%e5%ad%a6%e4%b9%a0","status":"publish","type":"post","link":"https:\/\/usei.cn\/index.php\/2023\/08\/23\/paddledetection-%e6%ba%90%e7%a0%81%e5%ad%a6%e4%b9%a0\/","title":{"rendered":"PaddleDetection \u6e90\u7801\u5b66\u4e60"},"content":{"rendered":"\n

\u6587\u4ef6\u7ed3\u6784<\/p>\n\n\n\n

tree -f#\u663e\u793a\u663e\u793a\u5b8c\u6574\u8def\u5f84\ntree -d#\u663e\u793a\u6240\u6709\u6587\u4ef6\u5939\n\u2514\u2500\u2500 PaddleDetection #\u4e3b\u76ee\u5f55\n    \u251c\u2500\u2500 configs  #\u5b58\u653e\u914d\u7f6e\u6587\u4ef6\n    \u2502    \u251c\u2500\u2500_base_ \u5404\u4e2a\u6a21\u5757\u914d\u7f6e\n    \u251c\u2500\u2500 dataset  #\u5b58\u653e\u6570\u636e\u96c6\uff0c\u6570\u636e\u96c6\u4e0b\u8f7d\u811a\u672c\uff0c\u5bf9\u5e94\u5404\u6570\u636e\u96c6\u6587\u4ef6\u5939\n    \u2502   \u251c\u2500\u2500 coco #80\u7c7b #\u7269\u4f53\n    \u2502   \u251c\u2500\u2500 fddb #1\u7c7b  #\u4eba\u8138 \u901a\u5e38\u7528\u6765\u8bc4\u4f30\u4eba\u8138\u68c0\u6d4b\u7b97\u6cd5\n    \u2502   \u251c\u2500\u2500 mot # \u591a\u76ee\u6807\u8ddf\u8e2a\n    \u2502   \u251c\u2500\u2500 roadsign_voc #\u8f66\u9053\u7ebf\u6570\u636e 4\u7c7b\n    \u2502   \u251c\u2500\u2500 voc #20\u7c7b #\u7269\u4f53\n    \u2502   \u2514\u2500\u2500 wider_face #1\u7c7b \u4eba\u8138\n    \u251c\u2500\u2500 deploy  #\u90e8\u7f72\u76f8\u5173\n    \u2502   \u251c\u2500\u2500 cpp #C++\u90e8\u7f72\n    \u2502   \u2502   \u251c\u2500\u2500 cmake #cmake\u6587\u4ef6\n    \u2502   \u2502   \u251c\u2500\u2500 docs #\u90e8\u7f72\u6587\u6863\n    \u2502   \u2502   \u251c\u2500\u2500 include #\u5e93\u5934\u6587\u4ef6\n    \u2502   \u2502   \u251c\u2500\u2500 scripts #\u4f9d\u8d56\u5e93\u914d\u7f6e\u811a\u672c\uff0cbuild\u811a\u672c\n    \u2502   \u2502   \u2514\u2500\u2500 src #\u6e90\u7801\n    \u2502   \u2514\u2500\u2500 python #python\u90e8\u7f72\n    \u251c\u2500\u2500 ppdet   #\u98de\u6868\u7269\u4f53\u68c0\u6d4b\u5957\u4ef6\n    \u2502   \u251c\u2500\u2500 core #\u6838\u5fc3\u90e8\u5206\n    \u2502   \u2502   \u2514\u2500\u2500 config #\u5b9e\u4f8b\u3001\u6ce8\u518c\u7c7b\u7684\u914d\u7f6e\n    \u2502   \u251c\u2500\u2500 data #\u6570\u636e\u5904\u7406\n    \u2502   \u2502   \u251c\u2500\u2500 shared_queue #\u5171\u4eab\u961f\u5217\uff08\u6570\u636e\u591a\u7ebf\u7a0b\uff09\n    \u2502   \u2502   \u251c\u2500\u2500 source #\u5404\u79cd\u6570\u636e\u96c6\u7c7b\n    \u2502   \u2502   \u251c\u2500\u2500 tests #\u6d4b\u8bd5\n    \u2502   \u2502   \u251c\u2500\u2500 tools #\u5de5\u5177\uff08\u8f6ccoco\u6570\u636e\u683c\u5f0f\uff09\n    \u2502   \u2502   \u2514\u2500\u2500 transform #\u6570\u636e\u589e\u5f3a\u6a21\u5757\n    \u2502   \u251c\u2500\u2500 ext_op  #\u589e\u52a0op\n    \u2502   \u2502   \u251c\u2500\u2500 src #op\u5b9e\u73b0\u6e90\u7801\n    \u2502   \u2502   \u2514\u2500\u2500 test #op\u6d4b\u8bd5\n    \u2502   \u251c\u2500\u2500 modeling #\u6a21\u578b\u7ed3\u6784\n    \u2502   \u2502   \u251c\u2500\u2500 architectures #\u7f51\u7edc\u7ed3\u6784\n    \u2502   \u2502   \u251c\u2500\u2500 backbones #\u4e3b\u5e72\u7f51\u7edc\n    \u2502   \u2502   \u251c\u2500\u2500 heads #\u5934\uff08RPN\u3001loss\uff09\n    \u2502   \u2502   \u251c\u2500\u2500 losses #\u5934\uff08loss\uff09\n    \u2502   \u2502   \u251c\u2500\u2500 mot #\u591a\u76ee\u6807\u8ddf\u8e2a\uff08tracking\uff09\n    \u2502   \u2502   \u251c\u2500\u2500 necks #\u9888\uff08FPN\uff09\n    \u2502   \u2502   \u251c\u2500\u2500 proposal_generator #\u5efa\u8bae\u6846\u751f\u6210\uff08anchor\u3001rpnhead\uff09\n    \u2502   \u2502   \u251c\u2500\u2500 reid #\u91cd\u8bc6\u522b\n    \u2502   \u2502   \u2514\u2500\u2500 tests #\u6d4b\u8bd5\n    \u2502   \u251c\u2500\u2500 py_op #\u4e00\u4e9b\u524d\u540e\u5904\u7406\n    \u2502   \u2514\u2500\u2500 utils #\u5b9e\u7528\u5de5\u5177\n    \u2514\u2500\u2500 tools #\u8bad\u7ec3\uff0c\u6d4b\u8bd5\uff0c\u9a8c\u8bc1\uff0c\u5c06\u6765\u8fd8\u6709\u6a21\u578b\u5bfc\u51fa\n\n\n<\/code><\/pre>\n\n\n\n
\u811a\u672c\u516c\u5171\u53c2\u6570<\/p>\n\n\n\n
FLAG<\/th>\u652f\u6301\u811a\u672c<\/th>\u7528\u9014<\/th>\u9ed8\u8ba4\u503c<\/th>\u5907\u6ce8<\/th><\/tr><\/thead>
-c<\/td>ALL<\/td>\u6307\u5b9a\u914d\u7f6e\u6587\u4ef6<\/td>None<\/td>\u5fc5\u9009<\/strong>\uff0c\u4f8b\u5982-c configs\/faster_rcnn\/faster_rcnn_r50_fpn_1x_coco.yml<\/td><\/tr>
-o<\/td>ALL<\/td>\u8bbe\u7f6e\u6216\u66f4\u6539\u914d\u7f6e\u6587\u4ef6\u91cc\u7684\u53c2\u6570\u5185\u5bb9<\/td>None<\/td>\u76f8\u8f83\u4e8e-c<\/code>\u8bbe\u7f6e\u7684\u914d\u7f6e\u6587\u4ef6\u6709\u66f4\u9ad8\u4f18\u5148\u7ea7\uff0c\u4f8b\u5982\uff1a-o use_gpu=False<\/code><\/td><\/tr>
–slim_config<\/td>ALL<\/td>\u6a21\u578b\u538b\u7f29\u7b56\u7565\u914d\u7f6e\u6587\u4ef6<\/td>None<\/td>\u4f8b\u5982--slim_config configs\/slim\/prune\/yolov3_prune_l1_norm.yml<\/code><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
2.\u8bad\u7ec3<\/h1>\n\n\n\n
\u811a\u672c\uff1atools\/train.py<\/p>\n\n\n\n
\u53c2\u6570<\/p>\n\n\n\n
FLAG<\/th>\u652f\u6301\u811a\u672c<\/th>\u7528\u9014<\/th>\u9ed8\u8ba4\u503c<\/th>\u5907\u6ce8<\/th><\/tr><\/thead>
–eval<\/td>train<\/td>\u662f\u5426\u8fb9\u8bad\u7ec3\u8fb9\u6d4b\u8bd5<\/td>False<\/td>\u5982\u9700\u6307\u5b9a\uff0c\u76f4\u63a5--eval<\/code>\u5373\u53ef<\/td><\/tr>
-r\/–resume_checkpoint<\/td>train<\/td>\u6062\u590d\u8bad\u7ec3\u52a0\u8f7d\u7684\u6743\u91cd\u8def\u5f84<\/td>None<\/td>\u4f8b\u5982\uff1a-r output\/faster_rcnn_r50_1x_coco\/10000<\/code><\/td><\/tr>
–use_vdl<\/td>train<\/td>\u662f\u5426\u4f7f\u7528VisualDL<\/a>\u8bb0\u5f55\u6570\u636e\uff0c\u8fdb\u800c\u5728VisualDL\u9762\u677f\u4e2d\u663e\u793a<\/td>False<\/td>VisualDL\u9700Python>=3.5<\/td><\/tr>
–vdl_log_dir<\/td>train<\/td>\u6307\u5b9a VisualDL \u8bb0\u5f55\u6570\u636e\u7684\u5b58\u50a8\u8def\u5f84<\/td>train:vdl_log_dir\/scalar<\/code>
<\/td>		VisualDL\u9700Python>=3.5<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\u8bc4\u4f30\u6307\u6807<\/h3>\n\n\n\n
loss<\/td>\u603b\u4f53\u635f\u5931<\/td>\u603b\u4f53\u635f\u5931\uff0c\u5b83\u662f\u5404\u4e2a\u635f\u5931\u51fd\u6570\u7684\u7ec4\u5408\uff0c\u7528\u4e8e\u8861\u91cf\u6a21\u578b\u9884\u6d4b\u7ed3\u679c\u548c\u771f\u5b9e\u6807\u7b7e\u4e4b\u95f4\u7684\u5dee\u5f02\u3002\u5b83\u5305\u542b\u4e86\u7c7b\u522b\u635f\u5931\u3001IOU \u635f\u5931\u548c L1 \u635f\u5931\u7684\u52a0\u6743\u548c\u3002<\/td><\/tr>
loss_cls<\/td>\u7c7b\u522b\u635f\u5931<\/td>\u7c7b\u522b\u635f\u5931\u8861\u91cf\u4e86\u6a21\u578b\u9884\u6d4b\u7684\u7269\u4f53\u7c7b\u522b\u662f\u5426\u6b63\u786e\u3002\u5b83\u901a\u5e38\u4f7f\u7528\u4ea4\u53c9\u71b5\u635f\u5931\u6765\u5ea6\u91cf\u9884\u6d4b\u7c7b\u522b\u7684\u6982\u7387\u5206\u5e03\u4e0e\u771f\u5b9e\u6807\u7b7e\u4e4b\u95f4\u7684\u5dee\u5f02\u3002<\/td><\/tr>
loss_iou<\/td>IOU \u635f\u5931<\/td>IOU \u635f\u5931\u662f\u6307\u6a21\u578b\u9884\u6d4b\u7684\u8fb9\u754c\u6846\u4e0e\u771f\u5b9e\u8fb9\u754c\u6846\u4e4b\u95f4\u7684\u91cd\u53e0\u7a0b\u5ea6\u7684\u635f\u5931\u3002\u5b83\u53ef\u4ee5\u5e2e\u52a9\u6a21\u578b\u66f4\u597d\u5730\u5b9a\u4f4d\u76ee\u6807\u3002IOU \u635f\u5931\u901a\u5e38\u4f7f\u7528\u8fb9\u754c\u6846\u7684 IOU \u503c\uff08Intersection over Union\uff09\u7684\u5dee\u8ddd\u6765\u5ea6\u91cf\u3002<\/td><\/tr>
loss_l1<\/td>L1 \u635f\u5931<\/td>L1 \u635f\u5931\u662f\u6307\u6a21\u578b\u9884\u6d4b\u7684\u8fb9\u754c\u6846\u4f4d\u7f6e\u4e0e\u771f\u5b9e\u8fb9\u754c\u6846\u4f4d\u7f6e\u4e4b\u95f4\u7684 L1 \u8303\u6570\u5dee\u5f02\u3002\u5b83\u6709\u52a9\u4e8e\u786e\u4fdd\u9884\u6d4b\u7684\u8fb9\u754c\u6846\u4f4d\u7f6e\u4e0e\u771f\u5b9e\u4f4d\u7f6e\u66f4\u63a5\u8fd1\u3002<\/td><\/tr>
<\/td><\/td><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
3.\u6a21\u578b\u8bc4\u4f30<\/h1>\n\n\n\n
\u811a\u672c\uff1atools\/eval.py<\/p>\n\n\n\n
!python tools\/eval.py -c configs\/yolov3\/yolov3_mobilenet_v1_roadsign.yml -o weights=output\/yolov3_mobilenet_v1_roadsign\/best_model.pdparams\n<\/code><\/pre>\n\n\n\n
\u53c2\u6570\u8bf4\u660e<\/p>\n\n\n\n
FLAG<\/th>\u652f\u6301\u811a\u672c<\/th>\u7528\u9014<\/th>\u9ed8\u8ba4\u503c<\/th>\u5907\u6ce8<\/th><\/tr><\/thead>
–output_eval<\/td>eval<\/td>\u8bc4\u4f30\u9636\u6bb5\u4fdd\u5b58json\u8def\u5f84<\/td>None<\/td>\u4f8b\u5982 --output_eval=eval_output<\/code>, \u9ed8\u8ba4\u4e3a\u5f53\u524d\u8def\u5f84<\/td><\/tr>
–json_eval<\/td>eval<\/td>\u662f\u5426\u901a\u8fc7\u5df2\u5b58\u5728\u7684bbox.json\u6216\u8005mask.json\u8fdb\u884c\u8bc4\u4f30<\/td>False<\/td>\u5982\u9700\u6307\u5b9a\uff0c\u76f4\u63a5--json_eval<\/code>\u5373\u53ef\uff0c json\u6587\u4ef6\u8def\u5f84\u5728--output_eval<\/code>\u4e2d\u8bbe\u7f6e<\/td><\/tr>
–classwise<\/td>eval<\/td>\u662f\u5426\u8bc4\u4f30\u5355\u7c7bAP\u548c\u7ed8\u5236\u5355\u7c7bPR\u66f2\u7ebf<\/td>False<\/td>\u5982\u9700\u6307\u5b9a\uff0c\u76f4\u63a5--classwise<\/code>\u5373\u53ef<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
4.\u9884\u6d4b<\/h1>\n\n\n\n
\u811a\u672c\uff1atools\/infer.py<\/p>\n\n\n\n
!python tools\/infer.py -c configs\/yolov3\/yolov3_mobilenet_v1_roadsign.yml -o weights=output\/yolov3_mobilenet_v1_roadsign\/best_model.pdparams --infer_img=dataset\/roadsign_voc\/images\/road114.png<\/code><\/pre>\n\n\n\n
FLAG<\/th>\u652f\u6301\u811a\u672c<\/th>\u7528\u9014<\/th>\u9ed8\u8ba4\u503c<\/th>\u5907\u6ce8<\/th><\/tr><\/thead>
–use_vdl<\/td>infer<\/td>\u662f\u5426\u4f7f\u7528VisualDL<\/a>\u8bb0\u5f55\u6570\u636e\uff0c\u8fdb\u800c\u5728VisualDL\u9762\u677f\u4e2d\u663e\u793a<\/td>False<\/td>VisualDL\u9700Python>=3.5<\/td><\/tr>
–vdl_log_dir<\/td>infer<\/td>\u6307\u5b9a VisualDL \u8bb0\u5f55\u6570\u636e\u7684\u5b58\u50a8\u8def\u5f84<\/td>train:vdl_log_dir\/scalar<\/code> infer: vdl_log_dir\/image<\/code><\/td>VisualDL\u9700Python>=3.5<\/td><\/tr>
–output_dir<\/td>infer<\/td>\u9884\u6d4b\u540e\u7ed3\u679c\u6216\u5bfc\u51fa\u6a21\u578b\u4fdd\u5b58\u8def\u5f84<\/td>.\/output<\/code><\/td>\u4f8b\u5982--output_dir=output<\/code><\/td><\/tr>
–infer_dir<\/td>infer<\/td>\u7528\u4e8e\u9884\u6d4b\u7684\u56fe\u7247\u6587\u4ef6\u5939\u8def\u5f84<\/td>None<\/td>--infer_img<\/code>\u548c--infer_dir<\/code>\u5fc5\u987b\u81f3\u5c11\u8bbe\u7f6e\u4e00\u4e2a<\/td><\/tr>
–infer_img<\/td>infer<\/td>\u7528\u4e8e\u9884\u6d4b\u7684\u56fe\u7247\u8def\u5f84<\/td>None<\/td>--infer_img<\/code>\u548c--infer_dir<\/code>\u5fc5\u987b\u81f3\u5c11\u8bbe\u7f6e\u4e00\u4e2a\uff0cinfer_img<\/code>\u5177\u6709\u66f4\u9ad8\u4f18\u5148\u7ea7<\/td><\/tr>
–save_txt<\/td>infer<\/td>\u662f\u5426\u5728\u6587\u4ef6\u5939\u4e0b\u5c06\u56fe\u7247\u7684\u9884\u6d4b\u7ed3\u679c\u4fdd\u5b58\u5230\u6587\u672c\u6587\u4ef6\u4e2d<\/td>False<\/td>\u53ef\u9009<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
5.\u5bfc\u51fa\u6a21\u578b<\/h1>\n\n\n\n
\u811a\u672c\uff1atools\/export_model.py<\/p>\n\n\n\n
* \u5bfc\u51faYOLOv3\u6a21\u578b\npython tools\/export_model.py -c configs\/yolov3\/yolov3_darknet53_270e_coco.yml --output_dir=.\/inference_model \\\n -o weights=weights\/yolov3_darknet53_270e_coco.pdparams<\/code><\/pre>\n\n\n\n
\u53c2\u6570\u8bf4\u660e<\/p>\n\n\n\n
FLAG<\/th>\u652f\u6301\u811a\u672c<\/th>\u7528\u9014<\/th>\u9ed8\u8ba4\u503c<\/th>\u5907\u6ce8<\/th><\/tr><\/thead>
–output_dir<\/td>export_model<\/td>\u9884\u6d4b\u540e\u7ed3\u679c\u6216\u5bfc\u51fa\u6a21\u578b\u4fdd\u5b58\u8def\u5f84<\/td>.\/output<\/code><\/td>\u4f8b\u5982--output_dir=output<\/code><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
output\n  \u251c\u2500\u2500 infer_cfg.yaml            # \u90e8\u7f72\u76f8\u5173\u7684\u914d\u7f6e\u6587\u4ef6\n  \u251c\u2500\u2500 model.pdiparams        # \u9759\u6001\u56fe\u6a21\u578b\u53c2\u6570\n  \u251c\u2500\u2500 model.pdiparams.info   # \u53c2\u6570\u989d\u5916\u4fe1\u606f\uff0c\u4e00\u822c\u65e0\u9700\u5173\u6ce8\n  \u2514\u2500\u2500 model.pdmodel          # \u9759\u6001\u56fe\u6a21\u578b\u6587\u4ef6<\/code><\/pre>\n\n\n\n
6.\u90e8\u7f72<\/h1>\n\n\n\n
\u5f62\u5f0f<\/th>\u8bed\u8a00<\/th>\u6559\u7a0b<\/th>\u8bbe\u5907\/\u5e73\u53f0<\/th><\/tr><\/thead>
PaddleInference<\/td>Python<\/td>\u5df2\u5b8c\u5584<\/td>linux(NV-Jetson)\u3001windows<\/td><\/tr>
PaddleInference<\/td>C++<\/td>\u5df2\u5b8c\u5584<\/td>linux(NV-Jetson)\u3001windows<\/td><\/tr>
PaddleServing<\/td>Python<\/td>\u5df2\u5b8c\u5584<\/td>linux(NV-Jetson)\u3001windows<\/td><\/tr>
PaddleLite<\/td>Python<\/td>\u5df2\u5b8c\u5584<\/td>Android\u3001IOS\u3001FPGA\u3001RK…<\/td><\/tr><\/tbody><\/table>
PaddleDetection\u652f\u6301TensorRT\u52a0\u901f\uff0c\u8bf7\u53c2\u8003\u6587\u6863TensorRT\u9884\u6d4b\u90e8\u7f72\u6559\u7a0b<\/a><\/figcaption><\/figure>\n","protected":false},"excerpt":{"rendered":"
\u6587\u4ef6\u7ed3\u6784 \u811a\u672c\u516c\u5171\u53c2\u6570 FLAG \u652f\u6301\u811a\u672c \u7528\u9014 \u9ed8\u8ba4\u503c \u5907\u6ce8 -c ALL \u6307\u5b9a\u914d\u7f6e\u6587\u4ef6 None \u5fc5\u9009\uff0c\u4f8b […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7,86],"tags":[],"class_list":["post-1825","post","type-post","status-publish","format-standard","hentry","category-ai","category-paddle"],"_links":{"self":[{"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/posts\/1825","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1825"}],"version-history":[{"count":16,"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/posts\/1825\/revisions"}],"predecessor-version":[{"id":1859,"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/posts\/1825\/revisions\/1859"}],"wp:attachment":[{"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1825"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1825"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/usei.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1825"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}