![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260202054924-69803ae457a04.jpg\")
\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260202054927-69803ae711292.jpg\")
\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260202054929-69803ae9a21ee.jpg\")
<\/p>\n\u968f\u7740\u7535\u5b50\u5546\u52a1\u7684\u8fc5\u731b\u53d1\u5c55，\u5feb\u9012\u884c\u4e1a\u4e5f\u968f\u4e4b\u84ec\u52c3\u5174\u8d77，\u5feb\u9012\u5305\u88f9\u7684\u6570\u91cf\u65e5\u76ca\u589e\u52a0\u3002\u6839\u636e\u7edf\u8ba1\u6570\u636e，\u5168\u7403\u5feb\u9012\u5e02\u573a\u7684\u5e74\u589e\u957f\u7387\u5df2\u8d85\u8fc710%，\u9884\u8ba1\u5728\u672a\u6765\u51e0\u5e74\u5185\u5c06\u7ee7\u7eed\u4fdd\u6301\u8fd9\u4e00\u589e\u957f\u8d8b\u52bf\u3002\u5728\u8fd9\u4e00\u80cc\u666f\u4e0b，\u5feb\u9012\u5305\u88f9\u7684\u7ba1\u7406\u548c\u5904\u7406\u6548\u7387\u6210\u4e3a\u4e86\u884c\u4e1a\u5185\u4e9f\u5f85\u89e3\u51b3\u7684\u91cd\u8981\u95ee\u9898\u3002\u4f20\u7edf\u7684\u5feb\u9012\u5305\u88f9\u68c0\u6d4b\u65b9\u6cd5\u5f80\u5f80\u4f9d\u8d56\u4eba\u5de5\u64cd\u4f5c，\u4e0d\u4ec5\u6548\u7387\u4f4e\u4e0b，\u800c\u4e14\u5bb9\u6613\u53d7\u5230\u4eba\u4e3a\u56e0\u7d20\u7684\u5f71\u54cd，\u5bfc\u81f4\u9519\u8bef\u7387\u8f83\u9ad8\u3002\u56e0\u6b64，\u57fa\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u81ea\u52a8\u5316\u68c0\u6d4b\u7cfb\u7edf\u5e94\u8fd0\u800c\u751f，\u6210\u4e3a\u63d0\u5347\u5feb\u9012\u5904\u7406\u6548\u7387\u548c\u51c6\u786e\u6027\u7684\u91cd\u8981\u624b\u6bb5\u3002<\/p>\n
\u5728\u4f17\u591a\u8ba1\u7b97\u673a\u89c6\u89c9\u6280\u672f\u4e2d，\u76ee\u6807\u68c0\u6d4b\u6280\u672f\u56e0\u5176\u5728\u5b9e\u65f6\u6027\u548c\u51c6\u786e\u6027\u65b9\u9762\u7684\u4f18\u52bf，\u9010\u6e10\u6210\u4e3a\u5feb\u9012\u5305\u88f9\u68c0\u6d4b\u7684\u4e3b\u6d41\u9009\u62e9\u3002YOLO（You Only Look Once）\u7cfb\u5217\u6a21\u578b\u4ee5\u5176\u9ad8\u6548\u7684\u7279\u6027\u53d7\u5230\u5e7f\u6cdb\u5173\u6ce8\u3002YOLOv8\u4f5c\u4e3a\u8be5\u7cfb\u5217\u7684\u6700\u65b0\u7248\u672c，\u7ed3\u5408\u4e86\u6df1\u5ea6\u5b66\u4e60\u7684\u5148\u8fdb\u6280\u672f，\u80fd\u591f\u5728\u4fdd\u6301\u9ad8\u7cbe\u5ea6\u7684\u540c\u65f6\u5b9e\u73b0\u5feb\u901f\u7684\u76ee\u6807\u68c0\u6d4b\u3002\u7136\u800c，\u5c3d\u7ba1YOLOv8\u5728\u591a\u79cd\u5e94\u7528\u573a\u666f\u4e2d\u8868\u73b0\u51fa\u8272，\u4f46\u5728\u7279\u5b9a\u9886\u57df\u5982\u5feb\u9012\u76d2\u68c0\u6d4b\u4e2d，\u4ecd\u7136\u5b58\u5728\u4e00\u4e9b\u6311\u6218，\u4f8b\u5982\u5bf9\u4e0d\u540c\u5c3a\u5bf8\u3001\u5f62\u72b6\u548c\u989c\u8272\u7684\u5feb\u9012\u76d2\u7684\u8bc6\u522b\u80fd\u529b，\u4ee5\u53ca\u5728\u590d\u6742\u80cc\u666f\u4e0b\u7684\u68c0\u6d4b\u51c6\u786e\u6027\u7b49\u3002\u56e0\u6b64，\u9488\u5bf9\u5feb\u9012\u76d2\u68c0\u6d4b\u4efb\u52a1\u5bf9YOLOv8\u8fdb\u884c\u6539\u8fdb\u548c\u4f18\u5316，\u5177\u6709\u91cd\u8981\u7684\u7814\u7a76\u610f\u4e49\u3002<\/p>\n
\u672c\u7814\u7a76\u6240\u4f7f\u7528\u7684\u6570\u636e\u96c6\u5305\u542b7815\u5f20\u56fe\u50cf，\u6db5\u76d6\u4e86\u56db\u4e2a\u7c7b\u522b\u7684\u5feb\u9012\u76d2，\u5206\u522b\u4e3a0\u30011\u300115\u548cbox\u3002\u8fd9\u4e00\u6570\u636e\u96c6\u7684\u6784\u5efa\u4e3a\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u8bc4\u4f30\u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u6837\u672c，\u80fd\u591f\u6709\u6548\u652f\u6301\u76ee\u6807\u68c0\u6d4b\u7b97\u6cd5\u7684\u4f18\u5316\u4e0e\u6539\u8fdb\u3002\u901a\u8fc7\u5bf9\u6570\u636e\u96c6\u7684\u6df1\u5165\u5206\u6790，\u6211\u4eec\u53ef\u4ee5\u8bc6\u522b\u51fa\u5feb\u9012\u76d2\u5728\u4e0d\u540c\u73af\u5883\u4e0b\u7684\u7279\u5f81，\u8fdb\u800c\u9488\u5bf9\u6027\u5730\u8c03\u6574YOLOv8\u7684\u7f51\u7edc\u7ed3\u6784\u548c\u53c2\u6570\u8bbe\u7f6e，\u4ee5\u63d0\u9ad8\u5176\u5728\u5feb\u9012\u76d2\u68c0\u6d4b\u4e2d\u7684\u8868\u73b0\u3002\u6b64\u5916，\u6570\u636e\u96c6\u7684\u591a\u6837\u6027\u4e5f\u4e3a\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u63d0\u4f9b\u4e86\u4fdd\u969c，\u4f7f\u5176\u80fd\u591f\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u9002\u5e94\u4e0d\u540c\u7684\u5feb\u9012\u573a\u666f\u3002<\/p>\n
\u672c\u7814\u7a76\u7684\u610f\u4e49\u4e0d\u4ec5\u5728\u4e8e\u63d0\u5347\u5feb\u9012\u76d2\u7684\u68c0\u6d4b\u7cbe\u5ea6\u548c\u6548\u7387，\u66f4\u5728\u4e8e\u63a8\u52a8\u5feb\u9012\u884c\u4e1a\u7684\u667a\u80fd\u5316\u8fdb\u7a0b\u3002\u901a\u8fc7\u5f15\u5165\u5148\u8fdb\u7684\u76ee\u6807\u68c0\u6d4b\u6280\u672f，\u6211\u4eec\u53ef\u4ee5\u5b9e\u73b0\u5feb\u9012\u5305\u88f9\u7684\u81ea\u52a8\u5316\u7ba1\u7406，\u964d\u4f4e\u4eba\u5de5\u6210\u672c，\u63d0\u9ad8\u5de5\u4f5c\u6548\u7387，\u8fdb\u800c\u4e3a\u5feb\u9012\u884c\u4e1a\u7684\u53ef\u6301\u7eed\u53d1\u5c55\u63d0\u4f9b\u6709\u529b\u652f\u6301\u3002\u540c\u65f6，\u7814\u7a76\u6210\u679c\u4e5f\u5c06\u4e3a\u5176\u4ed6\u9886\u57df\u7684\u76ee\u6807\u68c0\u6d4b\u4efb\u52a1\u63d0\u4f9b\u501f\u9274，\u63a8\u52a8\u8ba1\u7b97\u673a\u89c6\u89c9\u6280\u672f\u7684\u5e7f\u6cdb\u5e94\u7528\u3002<\/p>\n
\u7efc\u4e0a\u6240\u8ff0，\u57fa\u4e8e\u6539\u8fdbYOLOv8\u7684\u5feb\u9012\u76d2\u68c0\u6d4b\u7cfb\u7edf\u7684\u7814\u7a76，\u4e0d\u4ec5\u5177\u6709\u91cd\u8981\u7684\u7406\u8bba\u4ef7\u503c，\u4e5f\u5177\u5907\u663e\u8457\u7684\u5b9e\u9645\u5e94\u7528\u524d\u666f\u3002\u901a\u8fc7\u6df1\u5165\u63a2\u8ba8\u548c\u4f18\u5316\u76ee\u6807\u68c0\u6d4b\u6280\u672f，\u6211\u4eec\u671f\u5f85\u80fd\u591f\u4e3a\u5feb\u9012\u884c\u4e1a\u7684\u667a\u80fd\u5316\u8f6c\u578b\u8d21\u732e\u4e00\u4efd\u529b\u91cf，\u540c\u65f6\u4e3a\u76f8\u5173\u9886\u57df\u7684\u7814\u7a76\u63d0\u4f9b\u65b0\u7684\u601d\u8def\u548c\u65b9\u6cd5\u3002<\/p>\n
\n
\n<\/p>\n
\u5728\u672c\u7814\u7a76\u4e2d，\u6211\u4eec\u4f7f\u7528\u4e86\u540d\u4e3a\u201c1_box\u201d\u7684\u6570\u636e\u96c6，\u4ee5\u6539\u8fdbYOLOv8\u7684\u5feb\u9012\u76d2\u68c0\u6d4b\u7cfb\u7edf\u3002\u8be5\u6570\u636e\u96c6\u4e13\u95e8\u4e3a\u5feb\u9012\u76d2\u7684\u68c0\u6d4b\u4efb\u52a1\u800c\u8bbe\u8ba1，\u5305\u542b\u4e86\u4e30\u5bcc\u7684\u6837\u672c\u548c\u591a\u6837\u7684\u573a\u666f，\u65e8\u5728\u63d0\u9ad8\u6a21\u578b\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u51c6\u786e\u6027\u548c\u9c81\u68d2\u6027\u3002\u6570\u636e\u96c6\u7684\u7c7b\u522b\u6570\u91cf\u4e3a\u56db\u4e2a，\u5177\u4f53\u7c7b\u522b\u5305\u62ec\u201c0\u201d\u3001\u201c1\u201d\u3001\u201c15\u201d\u548c\u201cbox\u201d\u3002\u8fd9\u4e9b\u7c7b\u522b\u4ee3\u8868\u4e86\u4e0d\u540c\u7c7b\u578b\u7684\u5feb\u9012\u76d2，\u6db5\u76d6\u4e86\u4ece\u5e38\u89c1\u7684\u6807\u51c6\u5feb\u9012\u76d2\u5230\u7279\u6b8a\u5f62\u72b6\u548c\u5c3a\u5bf8\u7684\u5305\u88c5，\u786e\u4fdd\u4e86\u6a21\u578b\u5728\u591a\u79cd\u60c5\u51b5\u4e0b\u7684\u6709\u6548\u6027\u3002<\/p>\n
\u201c1_box\u201d\u6570\u636e\u96c6\u7684\u6784\u5efa\u7ecf\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1，\u5305\u542b\u4e86\u5927\u91cf\u7684\u6807\u6ce8\u56fe\u50cf，\u6db5\u76d6\u4e86\u5404\u79cd\u5149\u7167\u6761\u4ef6\u3001\u80cc\u666f\u73af\u5883\u548c\u62cd\u6444\u89d2\u5ea6\u3002\u8fd9\u4e9b\u56fe\u50cf\u4e0d\u4ec5\u5c55\u793a\u4e86\u5feb\u9012\u76d2\u7684\u4e0d\u540c\u5916\u89c2，\u8fd8\u8003\u8651\u4e86\u4e0d\u540c\u7684\u906e\u6321\u60c5\u51b5\u548c\u590d\u6742\u7684\u80cc\u666f，\u4ee5\u589e\u5f3a\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002\u6570\u636e\u96c6\u4e2d\u7684\u6bcf\u4e00\u5f20\u56fe\u50cf\u90fd\u7ecf\u8fc7\u4e13\u4e1a\u6807\u6ce8，\u786e\u4fdd\u4e86\u76ee\u6807\u7269\u4f53\u7684\u51c6\u786e\u5b9a\u4f4d\u548c\u7c7b\u522b\u8bc6\u522b\u3002\u8fd9\u79cd\u9ad8\u8d28\u91cf\u7684\u6807\u6ce8\u6570\u636e\u4e3aYOLOv8\u6a21\u578b\u7684\u8bad\u7ec3\u63d0\u4f9b\u4e86\u575a\u5b9e\u7684\u57fa\u7840，\u4f7f\u5176\u80fd\u591f\u5728\u5feb\u9012\u76d2\u68c0\u6d4b\u4efb\u52a1\u4e2d\u5b9e\u73b0\u66f4\u9ad8\u7684\u51c6\u786e\u7387\u3002<\/p>\n
\u5728\u6570\u636e\u96c6\u7684\u4f7f\u7528\u8fc7\u7a0b\u4e2d，\u6211\u4eec\u5c06\u5176\u5212\u5206\u4e3a\u8bad\u7ec3\u96c6\u548c\u9a8c\u8bc1\u96c6，\u4ee5\u4fbf\u4e8e\u5bf9\u6a21\u578b\u7684\u6027\u80fd\u8fdb\u884c\u8bc4\u4f30\u548c\u8c03\u4f18\u3002\u8bad\u7ec3\u96c6\u5305\u542b\u4e86\u5927\u90e8\u5206\u6837\u672c，\u7528\u4e8e\u6a21\u578b\u7684\u5b66\u4e60\u548c\u53c2\u6570\u4f18\u5316，\u800c\u9a8c\u8bc1\u96c6\u5219\u7528\u4e8e\u5b9e\u65f6\u76d1\u6d4b\u6a21\u578b\u5728\u672a\u89c1\u6570\u636e\u4e0a\u7684\u8868\u73b0\u3002\u901a\u8fc7\u8fd9\u79cd\u5212\u5206，\u6211\u4eec\u80fd\u591f\u6709\u6548\u5730\u907f\u514d\u8fc7\u62df\u5408\u73b0\u8c61，\u786e\u4fdd\u6a21\u578b\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u53ef\u9760\u6027\u3002<\/p>\n
\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u5347\u6a21\u578b\u7684\u6027\u80fd，\u6211\u4eec\u8fd8\u5bf9\u6570\u636e\u96c6\u8fdb\u884c\u4e86\u6570\u636e\u589e\u5f3a\u5904\u7406\u3002\u8fd9\u5305\u62ec\u56fe\u50cf\u7684\u65cb\u8f6c\u3001\u7f29\u653e\u3001\u88c1\u526a\u3001\u7ffb\u8f6c\u4ee5\u53ca\u989c\u8272\u8c03\u6574\u7b49\u64cd\u4f5c，\u4ee5\u589e\u52a0\u6570\u636e\u7684\u591a\u6837\u6027\u548c\u4e30\u5bcc\u6027\u3002\u8fd9\u79cd\u589e\u5f3a\u7b56\u7565\u4e0d\u4ec5\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u9c81\u68d2\u6027，\u8fd8\u5e2e\u52a9\u6a21\u578b\u66f4\u597d\u5730\u9002\u5e94\u4e0d\u540c\u7684\u5e94\u7528\u573a\u666f，\u4f8b\u5982\u5728\u5feb\u9012\u6295\u9012\u8fc7\u7a0b\u4e2d\u53ef\u80fd\u9047\u5230\u7684\u5404\u79cd\u590d\u6742\u73af\u5883\u3002<\/p>\n
\u6b64\u5916，\u6570\u636e\u96c6\u7684\u8bbe\u8ba1\u4e5f\u8003\u8651\u5230\u4e86\u5feb\u9012\u884c\u4e1a\u7684\u5b9e\u9645\u9700\u6c42\u3002\u968f\u7740\u7535\u5b50\u5546\u52a1\u7684\u5feb\u901f\u53d1\u5c55，\u5feb\u9012\u76d2\u7684\u79cd\u7c7b\u548c\u5f62\u72b6\u65e5\u76ca\u591a\u6837\u5316，\u56e0\u6b64\u201c1_box\u201d\u6570\u636e\u96c6\u7684\u7c7b\u522b\u8bbe\u7f6e\u6db5\u76d6\u4e86\u5e38\u89c1\u7684\u5feb\u9012\u76d2\u7c7b\u578b，\u4f7f\u5f97\u8bad\u7ec3\u51fa\u7684\u6a21\u578b\u80fd\u591f\u6709\u6548\u5e94\u5bf9\u5e02\u573a\u4e0a\u5404\u79cd\u5feb\u9012\u76d2\u7684\u68c0\u6d4b\u9700\u6c42\u3002\u8fd9\u4e00\u7279\u6027\u4f7f\u5f97\u6211\u4eec\u7684\u5feb\u9012\u76d2\u68c0\u6d4b\u7cfb\u7edf\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u5177\u6709\u8f83\u9ad8\u7684\u5b9e\u7528\u4ef7\u503c\u3002<\/p>\n
\u603b\u4e4b，\u201c1_box\u201d\u6570\u636e\u96c6\u4e3a\u6539\u8fdbYOLOv8\u7684\u5feb\u9012\u76d2\u68c0\u6d4b\u7cfb\u7edf\u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u6837\u672c\u548c\u9ad8\u8d28\u91cf\u7684\u6807\u6ce8，\u786e\u4fdd\u4e86\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u8bc4\u4f30\u8fc7\u7a0b\u7684\u79d1\u5b66\u6027\u548c\u6709\u6548\u6027\u3002\u901a\u8fc7\u5bf9\u6570\u636e\u96c6\u7684\u5408\u7406\u5229\u7528，\u6211\u4eec\u671f\u671b\u80fd\u591f\u63d0\u5347\u5feb\u9012\u76d2\u68c0\u6d4b\u7684\u51c6\u786e\u6027\u548c\u6548\u7387，\u4e3a\u5feb\u9012\u884c\u4e1a\u7684\u667a\u80fd\u5316\u53d1\u5c55\u8d21\u732e\u529b\u91cf\u3002
\n![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260202054932-69803aece79c1.jpg\")
\n![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260202054933-69803aed1ab4f.jpg\")
\n![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260202054933-69803aed57bb7.jpg\")
\n![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260202054933-69803aed715fe.jpg\")
\n![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260202054933-69803aed87d89.jpg\")
<\/p>\n
```python
\n# Ultralytics YOLO 🚀, AGPL-3.0 license<\/p>\n
from pathlib import Path
\nimport numpy as np
\nimport torch
\nfrom ultralytics.models.yolo.detect import DetectionValidator
\nfrom ultralytics.utils import LOGGER, ops
\nfrom ultralytics.utils.metrics import PoseMetrics, box_iou, kpt_iou
\nfrom ultralytics.utils.plotting import output_to_target, plot_images<\/p>\n
class PoseValidator(DetectionValidator):
\n """
\n PoseValidator\u7c7b\u6269\u5c55\u4e86DetectionValidator\u7c7b，\u7528\u4e8e\u57fa\u4e8e\u59ff\u6001\u6a21\u578b\u7684\u9a8c\u8bc1\u3002
\n """<\/p>\n
def __init__(self, dataloader=None, save_dir=None, pbar=None, args=None, _callbacks=None):
\n """\u521d\u59cb\u5316PoseValidator\u5bf9\u8c61，\u8bbe\u7f6e\u81ea\u5b9a\u4e49\u53c2\u6570\u548c\u5c5e\u6027\u3002"""
\n super().__init__(dataloader, save_dir, pbar, args, _callbacks)
\n self.sigma = None # \u7528\u4e8e\u8ba1\u7b97\u5173\u952e\u70b9\u7684\u6807\u51c6\u5dee
\n self.kpt_shape = None # \u5173\u952e\u70b9\u7684\u5f62\u72b6
\n self.args.task = 'pose' # \u8bbe\u7f6e\u4efb\u52a1\u7c7b\u578b\u4e3a\u59ff\u6001\u4f30\u8ba1
\n self.metrics = PoseMetrics(save_dir=self.save_dir, on_plot=self.on_plot) # \u521d\u59cb\u5316\u59ff\u6001\u5ea6\u91cf
\n if isinstance(self.args.device, str) and self.args.device.lower() == 'mps':
\n LOGGER.warning("WARNING \u26a0\ufe0f Apple MPS known Pose bug. Recommend 'device=cpu' for Pose models.")<\/p>\n
def preprocess(self, batch):
\n """\u9884\u5904\u7406\u6279\u6b21\u6570\u636e，\u5c06\u5173\u952e\u70b9\u6570\u636e\u8f6c\u6362\u4e3a\u6d6e\u70b9\u6570\u5e76\u79fb\u52a8\u5230\u8bbe\u5907\u4e0a\u3002"""
\n batch = super().preprocess(batch) # \u8c03\u7528\u7236\u7c7b\u7684\u9884\u5904\u7406\u65b9\u6cd5
\n batch['keypoints'] = batch['keypoints'].to(self.device).float() # \u8f6c\u6362\u5173\u952e\u70b9\u6570\u636e\u7c7b\u578b
\n return batch<\/p>\n
def postprocess(self, preds):
\n """\u5e94\u7528\u975e\u6700\u5927\u6291\u5236，\u8fd4\u56de\u9ad8\u7f6e\u4fe1\u5ea6\u7684\u68c0\u6d4b\u7ed3\u679c\u3002"""
\n return ops.non_max_suppression(preds,
\n self.args.conf,
\n self.args.iou,
\n labels=self.lb,
\n multi_label=True,
\n agnostic=self.args.single_cls,
\n max_det=self.args.max_det,
\n nc=self.nc)<\/p>\n
def init_metrics(self, model):
\n """\u521d\u59cb\u5316YOLO\u6a21\u578b\u7684\u59ff\u6001\u4f30\u8ba1\u5ea6\u91cf\u3002"""
\n super().init_metrics(model) # \u8c03\u7528\u7236\u7c7b\u7684\u521d\u59cb\u5316\u65b9\u6cd5
\n self.kpt_shape = self.data['kpt_shape'] # \u83b7\u53d6\u5173\u952e\u70b9\u5f62\u72b6
\n is_pose = self.kpt_shape == [17, 3] # \u5224\u65ad\u662f\u5426\u4e3a\u59ff\u6001\u6a21\u578b
\n nkpt = self.kpt_shape[0] # \u5173\u952e\u70b9\u6570\u91cf
\n self.sigma = OKS_SIGMA if is_pose else np.ones(nkpt) \/ nkpt # \u8bbe\u7f6e\u6807\u51c6\u5dee<\/p>\n
def update_metrics(self, preds, batch):
\n """\u66f4\u65b0\u5ea6\u91cf\u6307\u6807\u3002"""
\n for si, pred in enumerate(preds):
\n idx = batch['batch_idx'] == si # \u83b7\u53d6\u5f53\u524d\u6279\u6b21\u7684\u7d22\u5f15
\n cls = batch['cls'][idx] # \u83b7\u53d6\u5f53\u524d\u6279\u6b21\u7684\u7c7b\u522b
\n bbox = batch['bboxes'][idx] # \u83b7\u53d6\u5f53\u524d\u6279\u6b21\u7684\u8fb9\u754c\u6846
\n kpts = batch['keypoints'][idx] # \u83b7\u53d6\u5f53\u524d\u6279\u6b21\u7684\u5173\u952e\u70b9
\n nl, npr = cls.shape[0], pred.shape[0] # \u6807\u7b7e\u6570\u91cf\u548c\u9884\u6d4b\u6570\u91cf
\n nk = kpts.shape[1] # \u5173\u952e\u70b9\u6570\u91cf
\n shape = batch['ori_shape'][si] # \u539f\u59cb\u56fe\u50cf\u5f62\u72b6
\n correct_kpts = torch.zeros(npr, self.niou, dtype=torch.bool, device=self.device) # \u521d\u59cb\u5316\u6b63\u786e\u5173\u952e\u70b9
\n correct_bboxes = torch.zeros(npr, self.niou, dtype=torch.bool, device=self.device) # \u521d\u59cb\u5316\u6b63\u786e\u8fb9\u754c\u6846
\n self.seen += 1 # \u589e\u52a0\u5df2\u5904\u7406\u7684\u6279\u6b21\u6570<\/p>\n
if npr == 0: # \u5982\u679c\u6ca1\u6709\u9884\u6d4b
\n if nl:
\n self.stats.append((correct_bboxes, correct_kpts, *torch.zeros((2, 0), device=self.device), cls.squeeze(-1)))
\n continue<\/p>\n
# \u5904\u7406\u9884\u6d4b
\n if self.args.single_cls:
\n pred[:, 5] = 0 # \u5982\u679c\u662f\u5355\u7c7b，\u8bbe\u7f6e\u7c7b\u522b\u4e3a0
\n predn = pred.clone() # \u514b\u9686\u9884\u6d4b\u7ed3\u679c
\n ops.scale_boxes(batch['img'][si].shape[1:], predn[:, :4], shape, ratio_pad=batch['ratio_pad'][si]) # \u7f29\u653e\u8fb9\u754c\u6846
\n pred_kpts = predn[:, 6:].view(npr, nk, -1) # \u91cd\u5851\u5173\u952e\u70b9
\n ops.scale_coords(batch['img'][si].shape[1:], pred_kpts, shape, ratio_pad=batch['ratio_pad'][si]) # \u7f29\u653e\u5173\u952e\u70b9\u5750\u6807<\/p>\n
# \u8bc4\u4f30
\n if nl:
\n height, width = batch['img'].shape[2:] # \u83b7\u53d6\u56fe\u50cf\u9ad8\u5ea6\u548c\u5bbd\u5ea6
\n tbox = ops.xywh2xyxy(bbox) * torch.tensor((width, height, width, height), device=self.device) # \u76ee\u6807\u8fb9\u754c\u6846
\n ops.scale_boxes(batch['img'][si].shape[1:], tbox, shape, ratio_pad=batch['ratio_pad'][si]) # \u7f29\u653e\u76ee\u6807\u8fb9\u754c\u6846
\n tkpts = kpts.clone() # \u514b\u9686\u5173\u952e\u70b9
\n tkpts[…, 0] *= width # \u7f29\u653ex\u5750\u6807
\n tkpts[…, 1] *= height # \u7f29\u653ey\u5750\u6807
\n tkpts = ops.scale_coords(batch['img'][si].shape[1:], tkpts, shape, ratio_pad=batch['ratio_pad'][si]) # \u7f29\u653e\u5173\u952e\u70b9\u5750\u6807
\n labelsn = torch.cat((cls, tbox), 1) # \u5408\u5e76\u7c7b\u522b\u548c\u8fb9\u754c\u6846
\n correct_bboxes = self._process_batch(predn[:, :6], labelsn) # \u5904\u7406\u8fb9\u754c\u6846
\n correct_kpts = self._process_batch(predn[:, :6], labelsn, pred_kpts, tkpts) # \u5904\u7406\u5173\u952e\u70b9<\/p>\n
# \u8bb0\u5f55\u7edf\u8ba1\u4fe1\u606f
\n self.stats.append((correct_bboxes, correct_kpts, pred[:, 4], pred[:, 5], cls.squeeze(-1)))<\/p>\n
def _process_batch(self, detections, labels, pred_kpts=None, gt_kpts=None):
\n """
\n \u8fd4\u56de\u6b63\u786e\u7684\u9884\u6d4b\u77e9\u9635\u3002<\/p>\n
\u53c2\u6570:
\n detections (torch.Tensor): \u5f62\u72b6\u4e3a[N, 6]\u7684\u68c0\u6d4b\u5f20\u91cf\u3002
\n labels (torch.Tensor): \u5f62\u72b6\u4e3a[M, 5]\u7684\u6807\u7b7e\u5f20\u91cf\u3002
\n pred_kpts (torch.Tensor, \u53ef\u9009): \u5f62\u72b6\u4e3a[N, 51]\u7684\u9884\u6d4b\u5173\u952e\u70b9\u5f20\u91cf\u3002
\n gt_kpts (torch.Tensor, \u53ef\u9009): \u5f62\u72b6\u4e3a[N, 51]\u7684\u771f\u5b9e\u5173\u952e\u70b9\u5f20\u91cf\u3002<\/p>\n
\u8fd4\u56de:
\n torch.Tensor: \u5f62\u72b6\u4e3a[N, 10]\u7684\u6b63\u786e\u9884\u6d4b\u77e9\u9635\u3002
\n """
\n if pred_kpts is not None and gt_kpts is not None:
\n area = ops.xyxy2xywh(labels[:, 1:])[:, 2:].prod(1) * 0.53 # \u8ba1\u7b97\u9762\u79ef
\n iou = kpt_iou(gt_kpts, pred_kpts, sigma=self.sigma, area=area) # \u8ba1\u7b97\u5173\u952e\u70b9\u7684IoU
\n else: # \u5904\u7406\u8fb9\u754c\u6846
\n iou = box_iou(labels[:, 1:], detections[:, :4]) # \u8ba1\u7b97\u8fb9\u754c\u6846\u7684IoU<\/p>\n
return self.match_predictions(detections[:, 5], labels[:, 0], iou) # \u5339\u914d\u9884\u6d4b<\/p>\n
def plot_val_samples(self, batch, ni):
\n """\u7ed8\u5236\u5e76\u4fdd\u5b58\u9a8c\u8bc1\u96c6\u6837\u672c\u53ca\u5176\u9884\u6d4b\u7684\u8fb9\u754c\u6846\u548c\u5173\u952e\u70b9\u3002"""
\n plot_images(batch['img'],
\n batch['batch_idx'],
\n batch['cls'].squeeze(-1),
\n batch['bboxes'],
\n kpts=batch['keypoints'],
\n paths=batch['im_file'],
\n fname=self.save_dir \/ f'val_batch{ni}_labels.jpg',
\n names=self.names,
\n on_plot=self.on_plot)<\/p>\n
def plot_predictions(self, batch, preds, ni):
\n """\u7ed8\u5236YOLO\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c\u3002"""
\n pred_kpts = torch.cat([p[:, 6:].view(-1, *self.kpt_shape) for p in preds], 0) # \u5408\u5e76\u6240\u6709\u9884\u6d4b\u7684\u5173\u952e\u70b9
\n plot_images(batch['img'],
\n *output_to_target(preds, max_det=self.args.max_det),
\n kpts=pred_kpts,
\n paths=batch['im_file'],
\n fname=self.save_dir \/ f'val_batch{ni}_pred.jpg',
\n names=self.names,
\n on_plot=self.on_plot) # \u7ed8\u5236\u9884\u6d4b\u7ed3\u679c<\/p>\n
def pred_to_json(self, predn, filename):
\n """\u5c06YOLO\u9884\u6d4b\u7ed3\u679c\u8f6c\u6362\u4e3aCOCO JSON\u683c\u5f0f\u3002"""
\n stem = Path(filename).stem
\n image_id = int(stem) if stem.isnumeric() else stem # \u83b7\u53d6\u56fe\u50cfID
\n box = ops.xyxy2xywh(predn[:, :4]) # \u8f6c\u6362\u4e3axywh\u683c\u5f0f
\n box[:, :2] -= box[:, 2:] \/ 2 # \u5c06\u4e2d\u5fc3\u5750\u6807\u8f6c\u6362\u4e3a\u5de6\u4e0a\u89d2\u5750\u6807
\n for p, b in zip(predn.tolist(), box.tolist()):
\n self.jdict.append({
\n 'image_id': image_id,
\n 'category_id': self.class_map[int(p[5])],
\n 'bbox': [round(x, 3) for x in b],
\n 'keypoints': p[6:],
\n 'score': round(p[4], 5)})<\/p>\n
def eval_json(self, stats):
\n """\u4f7f\u7528COCO JSON\u683c\u5f0f\u8bc4\u4f30\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u3002"""
\n if self.args.save_json and self.is_coco and len(self.jdict):
\n anno_json = self.data['path'] \/ 'annotations\/person_keypoints_val2017.json' # \u6ce8\u91ca\u6587\u4ef6\u8def\u5f84
\n pred_json = self.save_dir \/ 'predictions.json' # \u9884\u6d4b\u7ed3\u679c\u6587\u4ef6\u8def\u5f84
\n LOGGER.info(f'\\\\nEvaluating pycocotools mAP using {pred_json} and {anno_json}…')
\n try:
\n check_requirements('pycocotools>=2.0.6') # \u68c0\u67e5pycocotools\u7684\u7248\u672c
\n from pycocotools.coco import COCO # \u5bfc\u5165COCO API
\n from pycocotools.cocoeval import COCOeval # \u5bfc\u5165COCO\u8bc4\u4f30API<\/p>\n
for x in anno_json, pred_json:
\n assert x.is_file(), f'{x} file not found' # \u786e\u4fdd\u6587\u4ef6\u5b58\u5728
\n anno = COCO(str(anno_json)) # \u521d\u59cb\u5316\u6ce8\u91caAPI
\n pred = anno.loadRes(str(pred_json)) # \u521d\u59cb\u5316\u9884\u6d4bAPI
\n for i, eval in enumerate([COCOeval(anno, pred, 'bbox'), COCOeval(anno, pred, 'keypoints')]):
\n if self.is_coco:
\n eval.params.imgIds = [int(Path(x).stem) for x in self.dataloader.dataset.im_files] # \u8bbe\u7f6e\u8bc4\u4f30\u56fe\u50cfID
\n eval.evaluate() # \u8bc4\u4f30
\n eval.accumulate() # \u7d2f\u79ef\u7ed3\u679c
\n eval.summarize() # \u603b\u7ed3\u7ed3\u679c
\n idx = i * 4 + 2
\n stats[self.metrics.keys[idx + 1]], stats[self.metrics.keys[idx]] = eval.stats[:2] # \u66f4\u65b0mAP
\n except Exception as e:
\n LOGGER.warning(f'pycocotools unable to run: {e}') # \u6355\u83b7\u5f02\u5e38\u5e76\u8bb0\u5f55\u8b66\u544a
\n return stats # \u8fd4\u56de\u7edf\u8ba1\u4fe1\u606f<\/p>\n
\u8fd9\u4e9b\u6838\u5fc3\u90e8\u5206\u5171\u540c\u6784\u6210\u4e86\u59ff\u6001\u4f30\u8ba1\u6a21\u578b\u9a8c\u8bc1\u7684\u4e3b\u8981\u6d41\u7a0b，\u6db5\u76d6\u4e86\u6570\u636e\u5904\u7406\u3001\u6a21\u578b\u8bc4\u4f30\u548c\u7ed3\u679c\u53ef\u89c6\u5316\u7b49\u5173\u952e\u73af\u8282\u3002```
\n\u8be5\u6587\u4ef6\u662f\u4e00\u4e2a\u7528\u4e8eYOLOv8\u6a21\u578b\u59ff\u6001\u4f30\u8ba1\u7684\u9a8c\u8bc1\u5668\u7c7bPoseValidator\u7684\u5b9e\u73b0，\u7ee7\u627f\u81eaDetectionValidator\u7c7b\u3002\u8be5\u7c7b\u4e3b\u8981\u7528\u4e8e\u5728\u59ff\u6001\u4f30\u8ba1\u4efb\u52a1\u4e2d\u8bc4\u4f30\u6a21\u578b\u7684\u6027\u80fd\u3002\u4ee3\u7801\u4e2d\u5305\u542b\u4e86\u591a\u4e2a\u65b9\u6cd5\u548c\u5c5e\u6027，\u4e0b\u9762\u662f\u5bf9\u5176\u529f\u80fd\u7684\u9010\u6b65\u5206\u6790\u3002<\/p>\n
\u9996\u5148，\u7c7b\u7684\u6784\u9020\u51fd\u6570__init__\u521d\u59cb\u5316\u4e86\u4e00\u4e9b\u57fa\u672c\u53c2\u6570，\u5305\u62ec\u6570\u636e\u52a0\u8f7d\u5668\u3001\u4fdd\u5b58\u76ee\u5f55\u3001\u8fdb\u5ea6\u6761\u3001\u53c2\u6570\u548c\u56de\u8c03\u51fd\u6570\u3002\u5b83\u8fd8\u8bbe\u7f6e\u4e86\u4efb\u52a1\u7c7b\u578b\u4e3a\u201cpose\u201d，\u5e76\u521d\u59cb\u5316\u4e86\u59ff\u6001\u4f30\u8ba1\u7684\u5ea6\u91cf\u6307\u6807\u3002\u82e5\u8bbe\u5907\u4e3aApple\u7684MPS，\u7cfb\u7edf\u4f1a\u53d1\u51fa\u8b66\u544a，\u5efa\u8bae\u4f7f\u7528CPU\u8fdb\u884c\u59ff\u6001\u6a21\u578b\u7684\u8bad\u7ec3\u3002<\/p>\n
\u5728preprocess\u65b9\u6cd5\u4e2d，\u8f93\u5165\u7684\u6279\u6b21\u6570\u636e\u88ab\u5904\u7406，\u5c06\u5173\u952e\u70b9\u6570\u636e\u8f6c\u6362\u4e3a\u6d6e\u70b9\u6570\u5e76\u79fb\u52a8\u5230\u6307\u5b9a\u8bbe\u5907\u4e0a\u3002get_desc\u65b9\u6cd5\u8fd4\u56de\u4e00\u4e2a\u5b57\u7b26\u4e32，\u63cf\u8ff0\u4e86\u8bc4\u4f30\u6307\u6807\u7684\u683c\u5f0f\u3002<\/p>\n
postprocess\u65b9\u6cd5\u5e94\u7528\u975e\u6781\u5927\u503c\u6291\u5236，\u8fd4\u56de\u5177\u6709\u9ad8\u7f6e\u4fe1\u5ea6\u5206\u6570\u7684\u68c0\u6d4b\u7ed3\u679c\u3002init_metrics\u65b9\u6cd5\u521d\u59cb\u5316\u59ff\u6001\u4f30\u8ba1\u7684\u5ea6\u91cf\u6307\u6807，\u68c0\u67e5\u5173\u952e\u70b9\u7684\u5f62\u72b6\u5e76\u8bbe\u7f6e\u76f8\u5e94\u7684sigma\u503c\u3002<\/p>\n
update_metrics\u65b9\u6cd5\u7528\u4e8e\u66f4\u65b0\u6a21\u578b\u7684\u8bc4\u4f30\u6307\u6807\u3002\u5b83\u5904\u7406\u6bcf\u4e2a\u9884\u6d4b\u7ed3\u679c，\u8ba1\u7b97\u6b63\u786e\u7684\u5173\u952e\u70b9\u548c\u8fb9\u754c\u6846，\u5e76\u5c06\u7ed3\u679c\u6dfb\u52a0\u5230\u7edf\u8ba1\u4fe1\u606f\u4e2d\u3002\u5982\u679c\u9700\u8981，\u8fd8\u53ef\u4ee5\u5c06\u9884\u6d4b\u7ed3\u679c\u4fdd\u5b58\u4e3aJSON\u683c\u5f0f\u3002<\/p>\n
_process_batch\u65b9\u6cd5\u7528\u4e8e\u5904\u7406\u68c0\u6d4b\u548c\u6807\u7b7e，\u8ba1\u7b97IoU（\u4ea4\u5e76\u6bd4）\u5e76\u8fd4\u56de\u6b63\u786e\u7684\u9884\u6d4b\u77e9\u9635\u3002\u8be5\u65b9\u6cd5\u652f\u6301\u5173\u952e\u70b9\u548c\u8fb9\u754c\u6846\u7684\u5904\u7406\u3002<\/p>\n
plot_val_samples\u548cplot_predictions\u65b9\u6cd5\u7528\u4e8e\u7ed8\u5236\u9a8c\u8bc1\u96c6\u6837\u672c\u548c\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c，\u4fdd\u5b58\u4e3a\u56fe\u50cf\u6587\u4ef6\u3002pred_to_json\u65b9\u6cd5\u5c06YOLO\u7684\u9884\u6d4b\u7ed3\u679c\u8f6c\u6362\u4e3aCOCO\u683c\u5f0f\u7684JSON，\u4ee5\u4fbf\u540e\u7eed\u8bc4\u4f30\u3002<\/p>\n
\u6700\u540e，eval_json\u65b9\u6cd5\u4f7f\u7528COCO\u683c\u5f0f\u7684JSON\u8bc4\u4f30\u5bf9\u8c61\u68c0\u6d4b\u6a21\u578b\u7684\u6027\u80fd，\u8ba1\u7b97mAP（\u5e73\u5747\u7cbe\u5ea6\u5747\u503c）\u7b49\u6307\u6807，\u5e76\u8f93\u51fa\u8bc4\u4f30\u7ed3\u679c\u3002<\/p>\n
\u6574\u4f53\u6765\u770b，\u8be5\u6587\u4ef6\u5b9e\u73b0\u4e86YOLOv8\u59ff\u6001\u4f30\u8ba1\u6a21\u578b\u7684\u9a8c\u8bc1\u8fc7\u7a0b，\u6db5\u76d6\u4e86\u6570\u636e\u9884\u5904\u7406\u3001\u6307\u6807\u8ba1\u7b97\u3001\u7ed3\u679c\u53ef\u89c6\u5316\u548c\u8bc4\u4f30\u7b49\u591a\u4e2a\u65b9\u9762，\u9002\u7528\u4e8e\u59ff\u6001\u4f30\u8ba1\u4efb\u52a1\u7684\u6a21\u578b\u6027\u80fd\u8bc4\u4f30\u3002<\/p>\n
```python
\nimport torch
\nimport torch.nn as nn
\nimport torch.nn.functional as F
\nimport numpy as np<\/p>\n
# \u5b9a\u4e49\u4e00\u4e2a\u540d\u4e3aOREPA\u7684\u5377\u79ef\u6a21\u5757
\nclass OREPA(nn.Module):
\n def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=None, groups=1, dilation=1, act=True):
\n super(OREPA, self).__init__()<\/p>\n
# \u521d\u59cb\u5316\u8f93\u5165\u8f93\u51fa\u901a\u9053\u3001\u5377\u79ef\u6838\u5927\u5c0f\u3001\u6b65\u5e45\u3001\u586b\u5145\u3001\u5206\u7ec4\u548c\u6269\u5f20
\n self.in_channels = in_channels
\n self.out_channels = out_channels
\n self.kernel_size = kernel_size
\n self.stride = stride
\n self.padding = padding if padding is not None else (kernel_size \/\/ 2)
\n self.groups = groups
\n self.dilation = dilation<\/p>\n
# \u6fc0\u6d3b\u51fd\u6570\u7684\u9009\u62e9
\n self.nonlinear = nn.ReLU() if act else nn.Identity()<\/p>\n
# \u5b9a\u4e49\u5377\u79ef\u5c42\u7684\u6743\u91cd\u53c2\u6570
\n self.weight_orepa_origin = nn.Parameter(torch.Tensor(out_channels, in_channels \/\/ groups, kernel_size, kernel_size))
\n nn.init.kaiming_uniform_(self.weight_orepa_origin) # \u4f7f\u7528He\u521d\u59cb\u5316<\/p>\n
# \u5b9a\u4e49\u5176\u4ed6\u5377\u79ef\u5c42\u7684\u6743\u91cd\u53c2\u6570
\n self.weight_orepa_avg_conv = nn.Parameter(torch.Tensor(out_channels, in_channels \/\/ groups, 1, 1))
\n nn.init.kaiming_uniform_(self.weight_orepa_avg_conv)<\/p>\n
self.weight_orepa_1x1 = nn.Parameter(torch.Tensor(out_channels, in_channels \/\/ groups, 1, 1))
\n nn.init.kaiming_uniform_(self.weight_orepa_1x1)<\/p>\n
# \u5176\u4ed6\u521d\u59cb\u5316\u4ee3\u7801\u7701\u7565…<\/p>\n
def weight_gen(self):
\n # \u751f\u6210\u6700\u7ec8\u7684\u5377\u79ef\u6743\u91cd
\n weight_orepa_origin = self.weight_orepa_origin # \u539f\u59cb\u5377\u79ef\u6743\u91cd
\n weight_orepa_avg = self.weight_orepa_avg_conv # \u5e73\u5747\u5377\u79ef\u6743\u91cd
\n weight_orepa_1x1 = self.weight_orepa_1x1 # 1×1\u5377\u79ef\u6743\u91cd<\/p>\n
# \u5c06\u6240\u6709\u6743\u91cd\u7ed3\u5408
\n weight = weight_orepa_origin + weight_orepa_avg + weight_orepa_1x1
\n return weight<\/p>\n
def forward(self, inputs):
\n # \u524d\u5411\u4f20\u64ad
\n weight = self.weight_gen() # \u751f\u6210\u6743\u91cd
\n out = F.conv2d(inputs, weight, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups) # \u5377\u79ef\u64cd\u4f5c
\n return self.nonlinear(out) # \u6fc0\u6d3b\u51fd\u6570<\/p>\n
# \u5176\u4ed6\u7c7b\u7684\u5b9a\u4e49\u7701\u7565…<\/p>\n
\u4ee5\u4e0a\u662f\u4ee3\u7801\u7684\u6838\u5fc3\u90e8\u5206\u53ca\u5176\u8be6\u7ec6\u6ce8\u91ca，\u5e2e\u52a9\u7406\u89e3\u8be5\u6a21\u5757\u7684\u7ed3\u6784\u548c\u529f\u80fd\u3002```
\n\u8fd9\u4e2a\u7a0b\u5e8f\u6587\u4ef6\u5305\u542b\u4e86\u591a\u4e2a\u7c7b\u548c\u51fd\u6570，\u4e3b\u8981\u7528\u4e8e\u5b9e\u73b0\u4e00\u79cd\u540d\u4e3aOREPA（Optimized Reparameterization for Efficient Convolution）\u7684\u5377\u79ef\u6a21\u5757，\u65e8\u5728\u63d0\u9ad8\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u7684\u6027\u80fd\u548c\u6548\u7387\u3002\u4ee5\u4e0b\u662f\u5bf9\u4ee3\u7801\u7684\u8be6\u7ec6\u8bf4\u660e\u3002<\/p>\n
\u9996\u5148，\u5bfc\u5165\u4e86\u5fc5\u8981\u7684\u5e93，\u5305\u62ecPyTorch\u7684\u6838\u5fc3\u5e93\u3001\u6570\u5b66\u5e93\u3001NumPy\u4ee5\u53ca\u4e00\u4e9b\u81ea\u5b9a\u4e49\u7684\u6a21\u5757\u3002transI_fusebn\u548ctransVI_multiscale\u662f\u4e24\u4e2a\u8f85\u52a9\u51fd\u6570，\u524d\u8005\u7528\u4e8e\u5c06\u5377\u79ef\u6838\u548c\u6279\u5f52\u4e00\u5316\u5c42\u7684\u53c2\u6570\u878d\u5408，\u540e\u8005\u7528\u4e8e\u5bf9\u5377\u79ef\u6838\u8fdb\u884c\u591a\u5c3a\u5ea6\u586b\u5145\u3002<\/p>\n
OREPA\u7c7b\u662f\u6838\u5fc3\u7c7b\u4e4b\u4e00，\u7ee7\u627f\u81eann.Module\u3002\u5728\u521d\u59cb\u5316\u65b9\u6cd5\u4e2d，\u5b9a\u4e49\u4e86\u8f93\u5165\u548c\u8f93\u51fa\u901a\u9053\u3001\u5377\u79ef\u6838\u5927\u5c0f\u3001\u6b65\u5e45\u3001\u586b\u5145\u3001\u5206\u7ec4\u3001\u6269\u5f20\u7b49\u53c2\u6570\u3002\u6839\u636e\u662f\u5426\u5904\u4e8e\u90e8\u7f72\u6a21\u5f0f，\u521d\u59cb\u5316\u4e0d\u540c\u7684\u5377\u79ef\u5c42\u548c\u53c2\u6570\u3002\u8be5\u7c7b\u7684\u8bbe\u8ba1\u5141\u8bb8\u901a\u8fc7\u4e0d\u540c\u7684\u5206\u652f\u7ec4\u5408\u751f\u6210\u5377\u79ef\u6743\u91cd，\u5229\u7528\u4e86\u591a\u4e2a\u5377\u79ef\u6838\u548c\u6279\u5f52\u4e00\u5316\u7684\u7ec4\u5408，\u4ee5\u63d0\u9ad8\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\u3002<\/p>\n
weight_gen\u65b9\u6cd5\u7528\u4e8e\u751f\u6210\u6700\u7ec8\u7684\u5377\u79ef\u6743\u91cd，\u901a\u8fc7\u5bf9\u4e0d\u540c\u5206\u652f\u7684\u6743\u91cd\u8fdb\u884c\u52a0\u6743\u6c42\u548c，\u5f62\u6210\u6700\u7ec8\u7684\u5377\u79ef\u6838\u3002forward\u65b9\u6cd5\u5b9e\u73b0\u4e86\u524d\u5411\u4f20\u64ad，\u8ba1\u7b97\u8f93\u5165\u6570\u636e\u7684\u5377\u79ef\u7ed3\u679c，\u5e76\u901a\u8fc7\u975e\u7ebf\u6027\u6fc0\u6d3b\u51fd\u6570\u548c\u6279\u5f52\u4e00\u5316\u5c42\u8fdb\u884c\u5904\u7406\u3002<\/p>\n
OREPA_LargeConv\u7c7b\u5b9e\u73b0\u4e86\u4e00\u4e2a\u5927\u578b\u5377\u79ef\u6a21\u5757，\u652f\u6301\u66f4\u5927\u7684\u5377\u79ef\u6838\u3002\u5b83\u7684\u7ed3\u6784\u4e0eOREPA\u7c7b\u4f3c，\u4f46\u589e\u52a0\u4e86\u591a\u4e2a\u5377\u79ef\u5c42\u7684\u7ec4\u5408，\u80fd\u591f\u5904\u7406\u66f4\u590d\u6742\u7684\u7279\u5f81\u63d0\u53d6\u4efb\u52a1\u3002<\/p>\n
ConvBN\u7c7b\u5219\u662f\u4e00\u4e2a\u7b80\u5355\u7684\u5377\u79ef\u52a0\u6279\u5f52\u4e00\u5316\u6a21\u5757，\u652f\u6301\u5728\u90e8\u7f72\u6a21\u5f0f\u4e0b\u76f4\u63a5\u4f7f\u7528\u878d\u5408\u540e\u7684\u5377\u79ef\u6838\u3002<\/p>\n
OREPA_3x3_RepVGG\u7c7b\u662f\u4e00\u4e2a\u7279\u5b9a\u7684OREPA\u6a21\u5757，\u4e13\u95e8\u7528\u4e8e3×3\u5377\u79ef\u6838\u7684\u60c5\u51b5，\u652f\u6301\u4e0d\u540c\u7684\u5206\u652f\u7ec4\u5408\u4ee5\u751f\u6210\u5377\u79ef\u6743\u91cd\u3002<\/p>\n
\u6700\u540e，RepVGGBlock_OREPA\u7c7b\u5b9e\u73b0\u4e86\u4e00\u4e2a\u5b8c\u6574\u7684\u5757\u7ed3\u6784，\u7ed3\u5408\u4e86OREPA\u6a21\u5757\u548c1×1\u5377\u79ef\u5c42，\u80fd\u591f\u5904\u7406\u8f93\u5165\u6570\u636e\u5e76\u8f93\u51fa\u7279\u5f81\u56fe\u3002\u8be5\u7c7b\u8fd8\u652f\u6301SE（Squeeze-and-Excitation）\u6ce8\u610f\u529b\u673a\u5236\u7684\u96c6\u6210，\u4ee5\u8fdb\u4e00\u6b65\u589e\u5f3a\u7279\u5f81\u8868\u793a\u80fd\u529b\u3002<\/p>\n
\u603b\u4f53\u6765\u8bf4，\u8fd9\u4e2a\u6587\u4ef6\u5b9e\u73b0\u4e86\u4e00\u79cd\u9ad8\u6548\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u6a21\u5757，\u80fd\u591f\u901a\u8fc7\u4e0d\u540c\u7684\u53c2\u6570\u7ec4\u5408\u548c\u7ed3\u6784\u8bbe\u8ba1，\u63d0\u5347\u6a21\u578b\u7684\u6027\u80fd\u548c\u7075\u6d3b\u6027，\u9002\u7528\u4e8e\u5404\u79cd\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u3002<\/p>\n
```python
\nimport sys
\nimport subprocess<\/p>\n
def run_script(script_path):
\n """
\n \u4f7f\u7528\u5f53\u524d Python \u73af\u5883\u8fd0\u884c\u6307\u5b9a\u7684\u811a\u672c\u3002<\/p>\n
Args:
\n script_path (str): \u8981\u8fd0\u884c\u7684\u811a\u672c\u8def\u5f84<\/p>\n
Returns:
\n None
\n """
\n # \u83b7\u53d6\u5f53\u524d Python \u89e3\u91ca\u5668\u7684\u8def\u5f84
\n python_path = sys.executable<\/p>\n
# \u6784\u5efa\u8fd0\u884c\u547d\u4ee4，\u4f7f\u7528 streamlit \u8fd0\u884c\u6307\u5b9a\u7684\u811a\u672c
\n command = f'"{python_path}" -m streamlit run "{script_path}"'<\/p>\n
# \u6267\u884c\u547d\u4ee4
\n result = subprocess.run(command, shell=True)
\n # \u68c0\u67e5\u547d\u4ee4\u6267\u884c\u7684\u8fd4\u56de\u7801，\u5982\u679c\u4e0d\u4e3a0，\u8868\u793a\u6267\u884c\u51fa\u9519
\n if result.returncode != 0:
\n print("\u811a\u672c\u8fd0\u884c\u51fa\u9519\u3002")<\/p>\n
# \u5b9e\u4f8b\u5316\u5e76\u8fd0\u884c\u5e94\u7528
\nif __name__ == "__main__":
\n # \u6307\u5b9a\u8981\u8fd0\u884c\u7684\u811a\u672c\u8def\u5f84
\n script_path = "web.py" # \u5047\u8bbe\u811a\u672c\u5728\u5f53\u524d\u76ee\u5f55\u4e0b<\/p>\n
# \u8c03\u7528\u51fd\u6570\u8fd0\u884c\u811a\u672c
\n run_script(script_path)<\/p>\n
\u5bfc\u5165\u6a21\u5757：<\/p>\n
run_script \u51fd\u6570：<\/p>\n
\u4e3b\u7a0b\u5e8f\u5757：<\/p>\n
\u5728 run_script \u51fd\u6570\u4e2d，\u7a0b\u5e8f\u63a5\u53d7\u4e00\u4e2a\u53c2\u6570 script_path，\u8be5\u53c2\u6570\u662f\u8981\u8fd0\u884c\u7684\u811a\u672c\u7684\u8def\u5f84\u3002\u51fd\u6570\u9996\u5148\u83b7\u53d6\u5f53\u524d Python \u89e3\u91ca\u5668\u7684\u8def\u5f84，\u8fd9\u901a\u8fc7 sys.executable \u5b9e\u73b0\u3002\u63a5\u7740，\u7a0b\u5e8f\u6784\u5efa\u4e86\u4e00\u4e2a\u547d\u4ee4\u5b57\u7b26\u4e32，\u8be5\u547d\u4ee4\u7528\u4e8e\u8fd0\u884c streamlit，\u8fd9\u662f\u4e00\u4e2a\u7528\u4e8e\u6784\u5efa\u6570\u636e\u5e94\u7528\u7684\u5e93\u3002\u547d\u4ee4\u7684\u683c\u5f0f\u662f\u5c06 Python \u89e3\u91ca\u5668\u4e0e -m streamlit run \u7ed3\u5408\u4f7f\u7528，\u5e76\u4f20\u5165\u8981\u8fd0\u884c\u7684\u811a\u672c\u8def\u5f84\u3002<\/p>\n
\u968f\u540e，\u7a0b\u5e8f\u4f7f\u7528 subprocess.run \u65b9\u6cd5\u6267\u884c\u6784\u5efa\u597d\u7684\u547d\u4ee4\u3002\u8fd9\u4e2a\u65b9\u6cd5\u4f1a\u5728\u65b0\u7684\u5b50\u8fdb\u7a0b\u4e2d\u8fd0\u884c\u547d\u4ee4，\u5e76\u7b49\u5f85\u5176\u5b8c\u6210\u3002\u5982\u679c\u547d\u4ee4\u6267\u884c\u7684\u8fd4\u56de\u7801\u4e0d\u4e3a\u96f6，\u8868\u793a\u811a\u672c\u8fd0\u884c\u51fa\u9519，\u7a0b\u5e8f\u4f1a\u8f93\u51fa\u4e00\u6761\u9519\u8bef\u4fe1\u606f\u3002<\/p>\n
\u5728\u6587\u4ef6\u7684\u6700\u540e\u90e8\u5206，\u7a0b\u5e8f\u901a\u8fc7 if __name__ == "__main__": \u8bed\u53e5\u6765\u786e\u4fdd\u53ea\u6709\u5728\u76f4\u63a5\u8fd0\u884c\u8be5\u6587\u4ef6\u65f6\u624d\u4f1a\u6267\u884c\u4ee5\u4e0b\u4ee3\u7801\u3002\u5b83\u6307\u5b9a\u4e86\u8981\u8fd0\u884c\u7684\u811a\u672c\u8def\u5f84\u4e3a web.py，\u5e76\u8c03\u7528 run_script \u51fd\u6570\u6765\u6267\u884c\u8fd9\u4e2a\u811a\u672c\u3002<\/p>\n
\u603b\u4f53\u6765\u8bf4，\u8fd9\u4e2a\u7a0b\u5e8f\u7684\u76ee\u7684\u662f\u63d0\u4f9b\u4e00\u4e2a\u7b80\u5355\u7684\u63a5\u53e3\u6765\u8fd0\u884c\u4e00\u4e2a\u7279\u5b9a\u7684 Python \u811a\u672c，\u5e76\u5728\u8fd0\u884c\u8fc7\u7a0b\u4e2d\u5904\u7406\u53ef\u80fd\u51fa\u73b0\u7684\u9519\u8bef\u3002<\/p>\n
```python
\nimport random
\nimport numpy as np
\nimport torch.nn as nn
\nfrom ultralytics.data import build_dataloader, build_yolo_dataset
\nfrom ultralytics.engine.trainer import BaseTrainer
\nfrom ultralytics.models import yolo
\nfrom ultralytics.nn.tasks import DetectionModel
\nfrom ultralytics.utils import LOGGER, RANK
\nfrom ultralytics.utils.torch_utils import de_parallel, torch_distributed_zero_first<\/p>\n
class DetectionTrainer(BaseTrainer):
\n """
\n \u6269\u5c55\u81ea BaseTrainer \u7c7b，\u7528\u4e8e\u57fa\u4e8e\u68c0\u6d4b\u6a21\u578b\u7684\u8bad\u7ec3\u3002
\n """<\/p>\n
def build_dataset(self, img_path, mode="train", batch=None):
\n """
\n \u6784\u5efa YOLO \u6570\u636e\u96c6\u3002<\/p>\n
\u53c2\u6570:
\n img_path (str): \u5305\u542b\u56fe\u50cf\u7684\u6587\u4ef6\u5939\u8def\u5f84\u3002
\n mode (str): \u6a21\u5f0f\u4e3a `train` \u6216 `val`，\u7528\u6237\u53ef\u4ee5\u4e3a\u6bcf\u79cd\u6a21\u5f0f\u81ea\u5b9a\u4e49\u4e0d\u540c\u7684\u589e\u5f3a\u3002
\n batch (int, optional): \u6279\u6b21\u5927\u5c0f，\u4ec5\u7528\u4e8e `rect` \u6a21\u5f0f\u3002\u9ed8\u8ba4\u4e3a None\u3002
\n """
\n gs = max(int(de_parallel(self.model).stride.max() if self.model else 0), 32) # \u83b7\u53d6\u6a21\u578b\u7684\u6700\u5927\u6b65\u5e45
\n return build_yolo_dataset(self.args, img_path, batch, self.data, mode=mode, rect=mode == "val", stride=gs)<\/p>\n
def get_dataloader(self, dataset_path, batch_size=16, rank=0, mode="train"):
\n """\u6784\u9020\u5e76\u8fd4\u56de\u6570\u636e\u52a0\u8f7d\u5668\u3002"""
\n assert mode in ["train", "val"] # \u786e\u4fdd\u6a21\u5f0f\u6709\u6548
\n with torch_distributed_zero_first(rank): # \u4ec5\u5728 DDP \u65f6\u521d\u59cb\u5316\u6570\u636e\u96c6 *.cache \u4e00\u6b21
\n dataset = self.build_dataset(dataset_path, mode, batch_size) # \u6784\u5efa\u6570\u636e\u96c6
\n shuffle = mode == "train" # \u8bad\u7ec3\u6a21\u5f0f\u4e0b\u6253\u4e71\u6570\u636e
\n workers = self.args.workers if mode == "train" else self.args.workers * 2 # \u6839\u636e\u6a21\u5f0f\u8bbe\u7f6e\u5de5\u4f5c\u7ebf\u7a0b\u6570
\n return build_dataloader(dataset, batch_size, workers, shuffle, rank) # \u8fd4\u56de\u6570\u636e\u52a0\u8f7d\u5668<\/p>\n
def preprocess_batch(self, batch):
\n """\u5bf9\u4e00\u6279\u56fe\u50cf\u8fdb\u884c\u9884\u5904\u7406，\u5305\u62ec\u7f29\u653e\u548c\u8f6c\u6362\u4e3a\u6d6e\u70b9\u6570\u3002"""
\n batch["img"] = batch["img"].to(self.device, non_blocking=True).float() \/ 255 # \u8f6c\u6362\u4e3a\u6d6e\u70b9\u6570\u5e76\u5f52\u4e00\u5316
\n if self.args.multi_scale: # \u5982\u679c\u542f\u7528\u591a\u5c3a\u5ea6
\n imgs = batch["img"]
\n sz = (
\n random.randrange(self.args.imgsz * 0.5, self.args.imgsz * 1.5 + self.stride)
\n \/\/ self.stride
\n * self.stride
\n ) # \u968f\u673a\u9009\u62e9\u56fe\u50cf\u5927\u5c0f
\n sf = sz \/ max(imgs.shape[2:]) # \u8ba1\u7b97\u7f29\u653e\u56e0\u5b50
\n if sf != 1:
\n ns = [
\n math.ceil(x * sf \/ self.stride) * self.stride for x in imgs.shape[2:]
\n ] # \u8ba1\u7b97\u65b0\u7684\u5f62\u72b6
\n imgs = nn.functional.interpolate(imgs, size=ns, mode="bilinear", align_corners=False) # \u8fdb\u884c\u63d2\u503c\u7f29\u653e
\n batch["img"] = imgs # \u66f4\u65b0\u56fe\u50cf
\n return batch<\/p>\n
def get_model(self, cfg=None, weights=None, verbose=True):
\n """\u8fd4\u56de YOLO \u68c0\u6d4b\u6a21\u578b\u3002"""
\n model = DetectionModel(cfg, nc=self.data["nc"], verbose=verbose and RANK == -1) # \u521b\u5efa\u68c0\u6d4b\u6a21\u578b
\n if weights:
\n model.load(weights) # \u52a0\u8f7d\u6743\u91cd
\n return model<\/p>\n
def plot_training_samples(self, batch, ni):
\n """\u7ed8\u5236\u5e26\u6709\u6ce8\u91ca\u7684\u8bad\u7ec3\u6837\u672c\u3002"""
\n plot_images(
\n images=batch["img"],
\n batch_idx=batch["batch_idx"],
\n cls=batch["cls"].squeeze(-1),
\n bboxes=batch["bboxes"],
\n paths=batch["im_file"],
\n fname=self.save_dir \/ f"train_batch{ni}.jpg",
\n on_plot=self.on_plot,
\n )<\/p>\n
def plot_metrics(self):
\n """\u4ece CSV \u6587\u4ef6\u4e2d\u7ed8\u5236\u6307\u6807\u3002"""
\n plot_results(file=self.csv, on_plot=self.on_plot) # \u4fdd\u5b58\u7ed3\u679c\u56fe<\/p>\n
\u9996\u5148，\u6587\u4ef6\u4e2d\u5bfc\u5165\u4e86\u4e00\u4e9b\u5fc5\u8981\u7684\u5e93\u548c\u6a21\u5757，\u5305\u62ec\u6570\u5b66\u8fd0\u7b97\u3001\u968f\u673a\u6570\u751f\u6210\u3001\u6df1\u5ea6\u5b66\u4e60\u76f8\u5173\u7684\u5e93（\u5982 PyTorch），\u4ee5\u53ca\u4e00\u4e9b\u6765\u81ea ultralytics \u7684\u6a21\u5757，\u8fd9\u4e9b\u6a21\u5757\u63d0\u4f9b\u4e86\u6570\u636e\u5904\u7406\u3001\u6a21\u578b\u6784\u5efa\u548c\u8bad\u7ec3\u7684\u529f\u80fd\u3002<\/p>\n
DetectionTrainer \u7c7b\u662f\u8be5\u6587\u4ef6\u7684\u6838\u5fc3，\u8d1f\u8d23\u7ba1\u7406\u8bad\u7ec3\u8fc7\u7a0b\u3002\u5b83\u5305\u542b\u591a\u4e2a\u65b9\u6cd5，\u5176\u4e2d build_dataset \u65b9\u6cd5\u7528\u4e8e\u6784\u5efa YOLO \u6570\u636e\u96c6，\u652f\u6301\u8bad\u7ec3\u548c\u9a8c\u8bc1\u6a21\u5f0f，\u5e76\u5141\u8bb8\u7528\u6237\u81ea\u5b9a\u4e49\u4e0d\u540c\u7684\u589e\u5f3a\u65b9\u5f0f\u3002get_dataloader \u65b9\u6cd5\u5219\u6784\u5efa\u5e76\u8fd4\u56de\u6570\u636e\u52a0\u8f7d\u5668，\u786e\u4fdd\u5728\u5206\u5e03\u5f0f\u8bad\u7ec3\u65f6\u53ea\u521d\u59cb\u5316\u4e00\u6b21\u6570\u636e\u96c6，\u5e76\u6839\u636e\u6a21\u5f0f\u9009\u62e9\u662f\u5426\u6253\u4e71\u6570\u636e\u3002<\/p>\n
\u5728\u56fe\u50cf\u9884\u5904\u7406\u65b9\u9762，preprocess_batch \u65b9\u6cd5\u8d1f\u8d23\u5c06\u56fe\u50cf\u7f29\u653e\u5e76\u8f6c\u6362\u4e3a\u6d6e\u70b9\u6570\u683c\u5f0f，\u652f\u6301\u591a\u5c3a\u5ea6\u8bad\u7ec3，\u786e\u4fdd\u8f93\u5165\u56fe\u50cf\u7684\u5c3a\u5bf8\u9002\u5408\u6a21\u578b\u3002set_model_attributes \u65b9\u6cd5\u5219\u7528\u4e8e\u8bbe\u7f6e\u6a21\u578b\u7684\u5c5e\u6027，\u5982\u7c7b\u522b\u6570\u91cf\u548c\u7c7b\u522b\u540d\u79f0\u3002<\/p>\n
get_model \u65b9\u6cd5\u8fd4\u56de\u4e00\u4e2a YOLO \u68c0\u6d4b\u6a21\u578b，\u5e76\u53ef\u9009\u62e9\u52a0\u8f7d\u9884\u8bad\u7ec3\u6743\u91cd\u3002get_validator \u65b9\u6cd5\u8fd4\u56de\u4e00\u4e2a\u7528\u4e8e\u6a21\u578b\u9a8c\u8bc1\u7684\u5bf9\u8c61，\u5e2e\u52a9\u8bc4\u4f30\u6a21\u578b\u6027\u80fd\u3002<\/p>\n
\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d，label_loss_items \u65b9\u6cd5\u7528\u4e8e\u8fd4\u56de\u5e26\u6807\u7b7e\u7684\u635f\u5931\u5b57\u5178，progress_string \u65b9\u6cd5\u683c\u5f0f\u5316\u8bad\u7ec3\u8fdb\u5ea6\u4fe1\u606f，\u4fbf\u4e8e\u76d1\u63a7\u8bad\u7ec3\u72b6\u6001\u3002plot_training_samples \u548c plot_training_labels \u65b9\u6cd5\u5219\u7528\u4e8e\u53ef\u89c6\u5316\u8bad\u7ec3\u6837\u672c\u53ca\u5176\u6807\u6ce8，\u5e2e\u52a9\u7528\u6237\u76f4\u89c2\u4e86\u89e3\u8bad\u7ec3\u6570\u636e\u7684\u60c5\u51b5\u3002<\/p>\n
\u6700\u540e，plot_metrics \u65b9\u6cd5\u4ece CSV \u6587\u4ef6\u4e2d\u7ed8\u5236\u8bad\u7ec3\u6307\u6807，\u63d0\u4f9b\u4e86\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u6027\u80fd\u53cd\u9988\u3002\u6574\u4f53\u6765\u770b，\u8fd9\u4e2a\u6587\u4ef6\u5b9e\u73b0\u4e86\u4e00\u4e2a\u5b8c\u6574\u7684\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u8bad\u7ec3\u6846\u67b6，\u6db5\u76d6\u4e86\u6570\u636e\u5904\u7406\u3001\u6a21\u578b\u8bad\u7ec3\u548c\u7ed3\u679c\u53ef\u89c6\u5316\u7b49\u591a\u4e2a\u65b9\u9762\u3002<\/p>\n
```python
\nimport torch.nn as nn
\nimport torch<\/p>\n
class Conv2d_BN(torch.nn.Sequential):
\n """
\n \u5b9a\u4e49\u4e00\u4e2a\u5305\u542b\u5377\u79ef\u5c42\u548c\u6279\u5f52\u4e00\u5316\u5c42\u7684\u6a21\u5757
\n """
\n def __init__(self, a, b, ks=1, stride=1, pad=0, dilation=1,
\n groups=1, bn_weight_init=1):
\n super().__init__()
\n # \u6dfb\u52a0\u5377\u79ef\u5c42
\n self.add_module('c', torch.nn.Conv2d(
\n a, b, ks, stride, pad, dilation, groups, bias=False))
\n # \u6dfb\u52a0\u6279\u5f52\u4e00\u5316\u5c42
\n self.add_module('bn', torch.nn.BatchNorm2d(b))
\n # \u521d\u59cb\u5316\u6279\u5f52\u4e00\u5316\u5c42\u7684\u6743\u91cd
\n torch.nn.init.constant_(self.bn.weight, bn_weight_init)
\n torch.nn.init.constant_(self.bn.bias, 0)<\/p>\n
@torch.no_grad()
\n def fuse_self(self):
\n """
\n \u5c06\u5377\u79ef\u5c42\u548c\u6279\u5f52\u4e00\u5316\u5c42\u878d\u5408\u4e3a\u4e00\u4e2a\u5377\u79ef\u5c42
\n """
\n c, bn = self._modules.values() # \u83b7\u53d6\u5377\u79ef\u5c42\u548c\u6279\u5f52\u4e00\u5316\u5c42
\n # \u8ba1\u7b97\u878d\u5408\u540e\u7684\u6743\u91cd\u548c\u504f\u7f6e
\n w = bn.weight \/ (bn.running_var + bn.eps)**0.5
\n w = c.weight * w[:, None, None, None]
\n b = bn.bias – bn.running_mean * bn.weight \/ (bn.running_var + bn.eps)**0.5
\n # \u521b\u5efa\u65b0\u7684\u5377\u79ef\u5c42
\n m = torch.nn.Conv2d(w.size(1) * self.c.groups, w.size(0), w.shape[2:],
\n stride=self.c.stride, padding=self.c.padding,
\n dilation=self.c.dilation, groups=self.c.groups,
\n device=c.weight.device)
\n m.weight.data.copy_(w) # \u590d\u5236\u6743\u91cd
\n m.bias.data.copy_(b) # \u590d\u5236\u504f\u7f6e
\n return m # \u8fd4\u56de\u878d\u5408\u540e\u7684\u5377\u79ef\u5c42<\/p>\n
class RepViTBlock(nn.Module):
\n """
\n \u5b9a\u4e49RepViT\u7684\u57fa\u672c\u5757
\n """
\n def __init__(self, inp, hidden_dim, oup, kernel_size, stride, use_se, use_hs):
\n super(RepViTBlock, self).__init__()
\n assert stride in [1, 2] # \u786e\u4fdd\u6b65\u5e45\u4e3a1\u62162
\n self.identity = stride == 1 and inp == oup # \u5224\u65ad\u662f\u5426\u9700\u8981\u8df3\u8dc3\u8fde\u63a5
\n assert(hidden_dim == 2 * inp) # \u9690\u85cf\u5c42\u7ef4\u5ea6\u4e3a\u8f93\u5165\u7ef4\u5ea6\u7684\u4e24\u500d<\/p>\n
if stride == 2:
\n # \u5982\u679c\u6b65\u5e45\u4e3a2，\u5b9a\u4e49token\u6df7\u5408\u548c\u901a\u9053\u6df7\u5408
\n self.token_mixer = nn.Sequential(
\n Conv2d_BN(inp, inp, kernel_size, stride, (kernel_size – 1) \/\/ 2, groups=inp),
\n nn.Identity() if not use_se else SqueezeExcite(inp, 0.25),
\n Conv2d_BN(inp, oup, ks=1, stride=1, pad=0)
\n )
\n self.channel_mixer = nn.Sequential(
\n Conv2d_BN(oup, 2 * oup, 1, 1, 0),
\n nn.GELU() if use_hs else nn.GELU(),
\n Conv2d_BN(2 * oup, oup, 1, 1, 0, bn_weight_init=0),
\n )
\n else:
\n assert(self.identity) # \u5982\u679c\u6b65\u5e45\u4e3a1，\u786e\u4fdd\u8f93\u5165\u548c\u8f93\u51fa\u901a\u9053\u76f8\u540c
\n self.token_mixer = nn.Sequential(
\n Conv2d_BN(inp, inp, 3, 1, 1, groups=inp),
\n nn.Identity() if not use_se else SqueezeExcite(inp, 0.25),
\n )
\n self.channel_mixer = nn.Sequential(
\n Conv2d_BN(inp, hidden_dim, 1, 1, 0),
\n nn.GELU() if use_hs else nn.GELU(),
\n Conv2d_BN(hidden_dim, oup, 1, 1, 0, bn_weight_init=0),
\n )<\/p>\n
def forward(self, x):
\n """
\n \u524d\u5411\u4f20\u64ad
\n """
\n return self.channel_mixer(self.token_mixer(x)) # \u5148\u8fdb\u884ctoken\u6df7\u5408，\u518d\u8fdb\u884c\u901a\u9053\u6df7\u5408<\/p>\n
class RepViT(nn.Module):
\n """
\n \u5b9a\u4e49RepViT\u6a21\u578b
\n """
\n def __init__(self, cfgs):
\n super(RepViT, self).__init__()
\n self.cfgs = cfgs # \u914d\u7f6e\u53cd\u5411\u6b8b\u5dee\u5757
\n input_channel = self.cfgs[0][2] # \u83b7\u53d6\u8f93\u5165\u901a\u9053\u6570
\n # \u6784\u5efa\u7b2c\u4e00\u4e2a\u5c42
\n patch_embed = nn.Sequential(
\n Conv2d_BN(3, input_channel \/\/ 2, 3, 2, 1),
\n nn.GELU(),
\n Conv2d_BN(input_channel \/\/ 2, input_channel, 3, 2, 1)
\n )
\n layers = [patch_embed] # \u5b58\u50a8\u6240\u6709\u5c42
\n # \u6784\u5efa\u53cd\u5411\u6b8b\u5dee\u5757
\n for k, t, c, use_se, use_hs, s in self.cfgs:
\n output_channel = _make_divisible(c, 8) # \u786e\u4fdd\u8f93\u51fa\u901a\u9053\u6570\u53ef\u88ab8\u6574\u9664
\n exp_size = _make_divisible(input_channel * t, 8) # \u786e\u4fdd\u6269\u5c55\u901a\u9053\u6570\u53ef\u88ab8\u6574\u9664
\n layers.append(RepViTBlock(input_channel, exp_size, output_channel, k, s, use_se, use_hs))
\n input_channel = output_channel # \u66f4\u65b0\u8f93\u5165\u901a\u9053\u6570
\n self.features = nn.ModuleList(layers) # \u5c06\u6240\u6709\u5c42\u5b58\u50a8\u4e3aModuleList<\/p>\n
def forward(self, x):
\n """
\n \u524d\u5411\u4f20\u64ad
\n """
\n features = [None, None, None, None] # \u5b58\u50a8\u7279\u5f81\u56fe
\n for f in self.features:
\n x = f(x) # \u901a\u8fc7\u6bcf\u4e00\u5c42
\n # \u6839\u636e\u8f93\u5165\u5927\u5c0f\u8bb0\u5f55\u7279\u5f81\u56fe
\n if x.size(2) in [x.size(2) \/\/ 4, x.size(2) \/\/ 8, x.size(2) \/\/ 16, x.size(2) \/\/ 32]:
\n features[x.size(2) \/\/ x.size(2)] = x
\n return features # \u8fd4\u56de\u7279\u5f81\u56fe<\/p>\n
def repvit_m0_9(weights=''):
\n """
\n \u6784\u5efaRepViT\u6a21\u578b\u7684\u4e00\u4e2a\u53d8\u4f53
\n """
\n cfgs = [
\n # k, t, c, SE, HS, s
\n [3, 2, 48, 1, 0, 1],
\n # \u5176\u4ed6\u914d\u7f6e…
\n ]
\n model = RepViT(cfgs) # \u521b\u5efa\u6a21\u578b
\n if weights:
\n model.load_state_dict(update_weight(model.state_dict(), torch.load(weights)['model'])) # \u52a0\u8f7d\u6743\u91cd
\n return model # \u8fd4\u56de\u6a21\u578b<\/p>\n
\u8fd9\u4e9b\u6838\u5fc3\u90e8\u5206\u662f\u6784\u5efaRepViT\u6a21\u578b\u7684\u57fa\u7840，\u8d1f\u8d23\u5b9a\u4e49\u7f51\u7edc\u7ed3\u6784\u548c\u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\u3002```
\n\u8fd9\u4e2a\u7a0b\u5e8f\u6587\u4ef6\u5b9a\u4e49\u4e86\u4e00\u4e2a\u57fa\u4e8eRepVGG\u67b6\u6784\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b，\u4e3b\u8981\u7528\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u3002\u6587\u4ef6\u4e2d\u5305\u542b\u4e86\u591a\u4e2a\u7c7b\u548c\u51fd\u6570，\u6784\u5efa\u4e86\u4e00\u4e2a\u53ef\u914d\u7f6e\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b，\u5177\u4f53\u5b9e\u73b0\u4e86RepViT（\u4ee3\u8868\u6027\u89c6\u89c9\u53d8\u6362\u5668）\u6a21\u578b\u3002<\/p>\n
\u9996\u5148，\u6587\u4ef6\u5f15\u5165\u4e86\u5fc5\u8981\u7684\u5e93，\u5305\u62ecPyTorch\u548c\u4e00\u4e9b\u7528\u4e8e\u6784\u5efa\u6a21\u578b\u7684\u6a21\u5757\u3002__all__\u5b9a\u4e49\u4e86\u53ef\u4ee5\u88ab\u5916\u90e8\u8c03\u7528\u7684\u6a21\u578b\u7248\u672c，\u5305\u62ec\u4e0d\u540c\u7684RepViT\u6a21\u578b\u3002<\/p>\n
replace_batchnorm\u51fd\u6570\u7528\u4e8e\u66ff\u6362\u7f51\u7edc\u4e2d\u7684BatchNorm\u5c42，\u5c06\u5176\u8f6c\u6362\u4e3aIdentity\u5c42，\u4ee5\u4fbf\u5728\u63a8\u7406\u65f6\u52a0\u901f\u8ba1\u7b97\u3002_make_divisible\u51fd\u6570\u786e\u4fdd\u7f51\u7edc\u4e2d\u6240\u6709\u5c42\u7684\u901a\u9053\u6570\u90fd\u662f8\u7684\u500d\u6570，\u8fd9\u662f\u4e3a\u4e86\u517c\u5bb9\u67d0\u4e9b\u786c\u4ef6\u52a0\u901f\u5668\u7684\u8981\u6c42\u3002<\/p>\n
Conv2d_BN\u7c7b\u662f\u4e00\u4e2a\u7ec4\u5408\u6a21\u5757，\u5305\u542b\u5377\u79ef\u5c42\u548cBatchNorm\u5c42，\u5e76\u5728\u521d\u59cb\u5316\u65f6\u8bbe\u7f6e\u4e86BatchNorm\u7684\u6743\u91cd\u548c\u504f\u7f6e\u3002\u5b83\u8fd8\u5b9e\u73b0\u4e86fuse_self\u65b9\u6cd5，\u7528\u4e8e\u5c06\u5377\u79ef\u548cBatchNorm\u5c42\u878d\u5408\u4e3a\u4e00\u4e2a\u5377\u79ef\u5c42，\u4ee5\u63d0\u9ad8\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n
Residual\u7c7b\u5b9e\u73b0\u4e86\u6b8b\u5dee\u8fde\u63a5，\u5141\u8bb8\u5728\u8bad\u7ec3\u65f6\u968f\u673a\u4e22\u5f03\u4e00\u4e9b\u8f93\u5165，\u4ee5\u589e\u5f3a\u6a21\u578b\u7684\u9c81\u68d2\u6027\u3002\u5b83\u540c\u6837\u652f\u6301\u5c42\u7684\u878d\u5408\u3002<\/p>\n
RepVGGDW\u7c7b\u5b9e\u73b0\u4e86\u4e00\u4e2a\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\u6a21\u5757，\u7ed3\u5408\u4e86\u5377\u79ef\u548cBatchNorm，\u5e76\u652f\u6301\u878d\u5408\u64cd\u4f5c\u3002<\/p>\n
RepViTBlock\u7c7b\u662fRepViT\u6a21\u578b\u7684\u57fa\u672c\u6784\u5efa\u5757，\u5305\u542b\u4e86\u901a\u9053\u6df7\u5408\u548c\u4ee4\u724c\u6df7\u5408\u7684\u64cd\u4f5c\u3002\u6839\u636e\u6b65\u5e45\u7684\u4e0d\u540c，\u6784\u5efa\u4e86\u4e0d\u540c\u7684\u7f51\u7edc\u7ed3\u6784\u3002<\/p>\n
RepViT\u7c7b\u662f\u6574\u4e2a\u6a21\u578b\u7684\u4e3b\u4f53，\u63a5\u6536\u914d\u7f6e\u53c2\u6570\u5e76\u6784\u5efa\u7f51\u7edc\u3002\u5b83\u7684forward\u65b9\u6cd5\u5b9a\u4e49\u4e86\u524d\u5411\u4f20\u64ad\u7684\u8fc7\u7a0b，\u5e76\u5728\u8f93\u5165\u5c3a\u5bf8\u53d8\u5316\u65f6\u63d0\u53d6\u7279\u5f81\u3002<\/p>\n
update_weight\u51fd\u6570\u7528\u4e8e\u66f4\u65b0\u6a21\u578b\u7684\u6743\u91cd，\u5c06\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u6743\u91cd\u52a0\u8f7d\u5230\u5f53\u524d\u6a21\u578b\u4e2d\u3002<\/p>\n
\u6700\u540e，\u5b9a\u4e49\u4e86\u591a\u4e2a\u6a21\u578b\u6784\u9020\u51fd\u6570，\u5982repvit_m0_9\u3001repvit_m1_0\u7b49，\u8fd9\u4e9b\u51fd\u6570\u6839\u636e\u4e0d\u540c\u7684\u914d\u7f6e\u53c2\u6570\u6784\u5efa\u4e0d\u540c\u7248\u672c\u7684RepViT\u6a21\u578b，\u5e76\u53ef\u9009\u62e9\u6027\u5730\u52a0\u8f7d\u9884\u8bad\u7ec3\u6743\u91cd\u3002<\/p>\n
\u5728\u6587\u4ef6\u7684\u6700\u540e\u90e8\u5206，\u63d0\u4f9b\u4e86\u4e00\u4e2a\u793a\u4f8b\u4ee3\u7801，\u521b\u5efa\u4e86\u4e00\u4e2aRepViT\u6a21\u578b\u5b9e\u4f8b，\u5e76\u5bf9\u968f\u673a\u8f93\u5165\u8fdb\u884c\u4e86\u524d\u5411\u4f20\u64ad，\u8f93\u51fa\u4e86\u5404\u5c42\u7684\u7279\u5f81\u56fe\u5c3a\u5bf8\u3002<\/p>\n
\u6574\u4f53\u800c\u8a00，\u8fd9\u4e2a\u6587\u4ef6\u5b9e\u73b0\u4e86\u4e00\u4e2a\u7075\u6d3b\u4e14\u9ad8\u6548\u7684\u89c6\u89c9\u6a21\u578b，\u9002\u7528\u4e8e\u5404\u79cd\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1，\u5e76\u63d0\u4f9b\u4e86\u826f\u597d\u7684\u6269\u5c55\u6027\u548c\u53ef\u914d\u7f6e\u6027\u3002<\/p>\n
```python
\nimport torch
\nimport torch.nn.functional as F
\nfrom torch.autograd import Function
\nfrom torch.cuda.amp import custom_bwd, custom_fwd<\/p>\n
class DCNv3Function(Function):
\n @staticmethod
\n @custom_fwd
\n def forward(ctx, input, offset, mask, kernel_h, kernel_w, stride_h, stride_w, pad_h, pad_w, dilation_h, dilation_w, group, group_channels, offset_scale, im2col_step, remove_center):
\n # \u4fdd\u5b58\u5377\u79ef\u53c2\u6570\u5230\u4e0a\u4e0b\u6587\u4e2d，\u4ee5\u4fbf\u5728\u53cd\u5411\u4f20\u64ad\u65f6\u4f7f\u7528
\n ctx.kernel_h = kernel_h
\n ctx.kernel_w = kernel_w
\n ctx.stride_h = stride_h
\n ctx.stride_w = stride_w
\n ctx.pad_h = pad_h
\n ctx.pad_w = pad_w
\n ctx.dilation_h = dilation_h
\n ctx.dilation_w = dilation_w
\n ctx.group = group
\n ctx.group_channels = group_channels
\n ctx.offset_scale = offset_scale
\n ctx.im2col_step = im2col_step
\n ctx.remove_center = remove_center<\/p>\n
# \u8c03\u7528DCNv3\u7684\u524d\u5411\u8ba1\u7b97\u51fd\u6570
\n output = DCNv3.dcnv3_forward(input, offset, mask, kernel_h, kernel_w, stride_h, stride_w, pad_h, pad_w, dilation_h, dilation_w, group, group_channels, offset_scale, ctx.im2col_step)
\n ctx.save_for_backward(input, offset, mask) # \u4fdd\u5b58\u8f93\u5165\u5f20\u91cf\u4ee5\u4fbf\u53cd\u5411\u4f20\u64ad\u4f7f\u7528<\/p>\n
return output<\/p>\n
@staticmethod
\n @once_differentiable
\n @custom_bwd
\n def backward(ctx, grad_output):
\n # \u4ece\u4e0a\u4e0b\u6587\u4e2d\u6062\u590d\u4fdd\u5b58\u7684\u5f20\u91cf
\n input, offset, mask = ctx.saved_tensors<\/p>\n
# \u8c03\u7528DCNv3\u7684\u53cd\u5411\u8ba1\u7b97\u51fd\u6570
\n grad_input, grad_offset, grad_mask = DCNv3.dcnv3_backward(input, offset, mask, ctx.kernel_h, ctx.kernel_w, ctx.stride_h, ctx.stride_w, ctx.pad_h, ctx.pad_w, ctx.dilation_h, ctx.dilation_w, ctx.group, ctx.group_channels, ctx.offset_scale, grad_output.contiguous(), ctx.im2col_step)<\/p>\n
return grad_input, grad_offset, grad_mask, None, None, None, None, None, None, None, None, None, None, None, None, None<\/p>\n
def dcnv3_core_pytorch(input, offset, mask, kernel_h, kernel_w, stride_h, stride_w, pad_h, pad_w, dilation_h, dilation_w, group, group_channels, offset_scale, remove_center):
\n # \u8f93\u5165\u5f20\u91cf\u8fdb\u884c\u586b\u5145
\n input = F.pad(input, [0, 0, pad_h, pad_h, pad_w, pad_w])
\n N_, H_in, W_in, _ = input.shape # \u83b7\u53d6\u8f93\u5165\u7684\u5f62\u72b6
\n _, H_out, W_out, _ = offset.shape # \u83b7\u53d6\u504f\u79fb\u91cf\u7684\u5f62\u72b6<\/p>\n
# \u8ba1\u7b97\u53c2\u8003\u70b9\u548c\u7f51\u683c
\n ref = _get_reference_points(input.shape, input.device, kernel_h, kernel_w, dilation_h, dilation_w, pad_h, pad_w, stride_h, stride_w)
\n grid = _generate_dilation_grids(input.shape, kernel_h, kernel_w, dilation_h, dilation_w, group, input.device)<\/p>\n
# \u8ba1\u7b97\u91c7\u6837\u4f4d\u7f6e
\n sampling_locations = (ref + grid * offset_scale).repeat(N_, 1, 1, 1, 1)
\n if remove_center:
\n sampling_locations = remove_center_sampling_locations(sampling_locations, kernel_w=kernel_w, kernel_h=kernel_h)
\n sampling_locations = sampling_locations.flatten(3, 4) + offset * offset_scale \/ torch.tensor([W_in, H_in]).reshape(1, 1, 1, 2).to(input.device)<\/p>\n
# \u4f7f\u7528grid_sample\u8fdb\u884c\u91c7\u6837
\n sampling_input_ = F.grid_sample(input.view(N_, H_in * W_in, group * group_channels).transpose(1, 2).reshape(N_ * group, group_channels, H_in, W_in), sampling_locations, mode='bilinear', padding_mode='zeros', align_corners=False)<\/p>\n
# \u8ba1\u7b97\u8f93\u51fa
\n output = (sampling_input_ * mask.view(N_, H_out * W_out, group, -1).transpose(1, 2).reshape(N_ * group, 1, H_out * W_out, -1)).sum(-1).view(N_, group * group_channels, H_out * W_out)<\/p>\n
return output.transpose(1, 2).reshape(N_, H_out, W_out, -1).contiguous()<\/p>\n
DCNv3Function\u7c7b：\u8fd9\u662f\u4e00\u4e2a\u81ea\u5b9a\u4e49\u7684PyTorch\u51fd\u6570，\u5305\u542b\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u7684\u5b9e\u73b0\u3002<\/p>\n
dcnv3_core_pytorch\u51fd\u6570：\u5b9e\u73b0\u4e86DCNv3\u7684\u6838\u5fc3\u8ba1\u7b97\u903b\u8f91\u3002<\/p>\n
\u8f85\u52a9\u51fd\u6570：\u5982_get_reference_points\u548c_generate_dilation_grids\u7b49，\u8d1f\u8d23\u751f\u6210\u53c2\u8003\u70b9\u548c\u81a8\u80c0\u7f51\u683c，\u8fd9\u4e9b\u90fd\u662f\u5b9e\u73b0DCNv3\u6240\u9700\u7684\u5173\u952e\u6b65\u9aa4\u3002```
\n\u8fd9\u4e2a\u7a0b\u5e8f\u6587\u4ef6\u662f\u4e00\u4e2a\u5b9e\u73b0\u6df1\u5ea6\u5b66\u4e60\u4e2d\u53ef\u53d8\u5f62\u5377\u79ef（Deformable Convolution）V3\u7248\u672c\u7684PyTorch\u51fd\u6570\u7684\u4ee3\u7801\u3002\u8be5\u5b9e\u73b0\u4e3b\u8981\u4f9d\u8d56\u4e8e\u4e00\u4e2a\u540d\u4e3aDCNv3\u7684\u5916\u90e8\u5e93，\u5e76\u63d0\u4f9b\u4e86\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u7684\u529f\u80fd\u3002<\/p>\n<\/li>\n
\u9996\u5148，\u6587\u4ef6\u4e2d\u5bfc\u5165\u4e86\u4e00\u4e9b\u5fc5\u8981\u7684\u5e93，\u5305\u62ecPyTorch\u548c\u76f8\u5173\u7684\u529f\u80fd\u6a21\u5757\u3002\u63a5\u7740，\u5c1d\u8bd5\u5bfc\u5165DCNv3\u5e93\u5e76\u83b7\u53d6\u5176\u7248\u672c\u4fe1\u606f，\u4ee5\u4fbf\u5728\u540e\u7eed\u7684\u64cd\u4f5c\u4e2d\u4f7f\u7528\u3002<\/p>\n
\u6838\u5fc3\u7c7b\u662fDCNv3Function，\u5b83\u7ee7\u627f\u81eatorch.autograd.Function，\u5e76\u5b9e\u73b0\u4e86\u4e09\u4e2a\u9759\u6001\u65b9\u6cd5：forward\u3001backward\u548csymbolic\u3002forward\u65b9\u6cd5\u7528\u4e8e\u6267\u884c\u524d\u5411\u4f20\u64ad，\u63a5\u6536\u8f93\u5165\u5f20\u91cf\u3001\u504f\u79fb\u91cf\u3001\u63a9\u7801\u4ee5\u53ca\u5377\u79ef\u7684\u5404\u79cd\u53c2\u6570（\u5982\u6838\u7684\u5927\u5c0f\u3001\u6b65\u5e45\u3001\u586b\u5145\u7b49），\u5e76\u8c03\u7528DCNv3\u5e93\u4e2d\u7684\u524d\u5411\u51fd\u6570\u6765\u8ba1\u7b97\u8f93\u51fa\u3002\u8be5\u65b9\u6cd5\u8fd8\u4f1a\u4fdd\u5b58\u8f93\u5165\u3001\u504f\u79fb\u91cf\u548c\u63a9\u7801，\u4ee5\u4fbf\u5728\u53cd\u5411\u4f20\u64ad\u65f6\u4f7f\u7528\u3002<\/p>\n
backward\u65b9\u6cd5\u5b9e\u73b0\u4e86\u53cd\u5411\u4f20\u64ad\u7684\u529f\u80fd，\u8ba1\u7b97\u8f93\u5165\u3001\u504f\u79fb\u91cf\u548c\u63a9\u7801\u7684\u68af\u5ea6\u3002\u5b83\u4f7f\u7528\u4e86\u4fdd\u5b58\u7684\u5f20\u91cf\u548c\u4e4b\u524d\u4f20\u5165\u7684\u53c2\u6570\u6765\u8c03\u7528DCNv3\u5e93\u4e2d\u7684\u53cd\u5411\u51fd\u6570\u3002<\/p>\n
symbolic\u65b9\u6cd5\u7528\u4e8eONNX（\u5f00\u653e\u795e\u7ecf\u7f51\u7edc\u4ea4\u6362）\u6a21\u578b\u7684\u7b26\u53f7\u8868\u793a，\u5141\u8bb8\u5c06\u8fd9\u4e2a\u81ea\u5b9a\u4e49\u64cd\u4f5c\u5bfc\u51fa\u4e3aONNX\u683c\u5f0f\u3002<\/p>\n
\u63a5\u4e0b\u6765，\u6587\u4ef6\u4e2d\u5b9a\u4e49\u4e86\u4e00\u4e9b\u8f85\u52a9\u51fd\u6570，\u5305\u62ec_get_reference_points\u3001_generate_dilation_grids\u3001remove_center_sampling_locations\u548cdcnv3_core_pytorch\u3002\u8fd9\u4e9b\u51fd\u6570\u5206\u522b\u7528\u4e8e\u8ba1\u7b97\u53c2\u8003\u70b9\u3001\u751f\u6210\u81a8\u80c0\u7f51\u683c\u3001\u79fb\u9664\u4e2d\u5fc3\u91c7\u6837\u4f4d\u7f6e\u4ee5\u53ca\u5b9e\u73b0\u53ef\u53d8\u5f62\u5377\u79ef\u7684\u6838\u5fc3\u903b\u8f91\u3002<\/p>\n
\u5728dcnv3_core_pytorch\u51fd\u6570\u4e2d，\u9996\u5148\u5bf9\u8f93\u5165\u8fdb\u884c\u586b\u5145，\u7136\u540e\u8ba1\u7b97\u53c2\u8003\u70b9\u548c\u81a8\u80c0\u7f51\u683c\u3002\u63a5\u7740，\u6839\u636e\u7ed9\u5b9a\u7684\u504f\u79fb\u91cf\u548c\u63a9\u7801\u8fdb\u884c\u91c7\u6837，\u6700\u540e\u5c06\u7ed3\u679c\u8fdb\u884c\u5904\u7406\u5e76\u8fd4\u56de\u3002<\/p>\n
\u603b\u4f53\u6765\u8bf4，\u8fd9\u6bb5\u4ee3\u7801\u5b9e\u73b0\u4e86\u53ef\u53d8\u5f62\u5377\u79ef\u7684\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b，\u80fd\u591f\u5728\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u4e2d\u7075\u6d3b\u5730\u5904\u7406\u4e0d\u540c\u5f62\u72b6\u548c\u5c3a\u5bf8\u7684\u8f93\u5165\u6570\u636e，\u63d0\u5347\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\u3002<\/p>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260202054933-69803aedaf499.png\")
<\/p>\n
\u6b22\u8fce\u5927\u5bb6\u70b9\u8d5e\u3001\u6536\u85cf\u3001\u5173\u6ce8\u3001\u8bc4\u8bba\u5566 \u3001\u67e5\u770b👇🏻\u83b7\u53d6\u8054\u7cfb\u65b9\u5f0f👇🏻<\/p>\n","protected":false},"excerpt":{"rendered":"
\u80cc\u666f\u610f\u4e49
\n\u968f\u7740\u7535\u5b50\u5546\u52a1\u7684\u8fc5\u731b\u53d1\u5c55，\u5feb\u9012\u884c\u4e1a\u4e5f\u968f\u4e4b\u84ec\u52c3\u5174\u8d77，\u5feb\u9012\u5305\u88f9\u7684\u6570\u91cf\u65e5\u76ca\u589e\u52a0\u3002\u6839\u636e\u7edf\u8ba1\u6570\u636e，\u5168\u7403\u5feb\u9012\u5e02\u573a\u7684\u5e74\u589e\u957f\u7387\u5df2\u8d85\u8fc710%，\u9884\u8ba1\u5728\u672a\u6765\u51e0\u5e74\u5185\u5c06\u7ee7\u7eed\u4fdd\u6301\u8fd9\u4e00\u589e\u957f\u8d8b\u52bf\u3002\u5728\u8fd9\u4e00\u80cc\u666f\u4e0b，\u5feb\u9012\u5305\u88f9\u7684\u7ba1\u7406\u548c\u5904\u7406\u6548\u7387\u6210\u4e3a\u4e86\u884c\u4e1a\u5185\u4e9f\u5f85\u89e3\u51b3\u7684\u91cd\u8981\u95ee\u9898\u3002\u4f20\u7edf\u7684\u5feb\u9012\u5305\u88f9\u68c0\u6d4b\u65b9\u6cd5\u5f80\u5f80\u4f9d\u8d56\u4eba\u5de5\u64cd\u4f5c，\u4e0d\u4ec5\u6548\u7387\u4f4e\u4e0b，\u800c\u4e14\u5bb9\u6613\u53d7\u5230\u4eba\u4e3a\u56e0\u7d20\u7684\u5f71\u54cd，\u5bfc\u81f4\u9519\u8bef\u7387\u8f83\u9ad8\u3002\u56e0\u6b64，\u57fa\u4e8e\u8ba1\u7b97<\/p>\n","protected":false},"author":2,"featured_media":70830,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[7445,156,87,5134,5281,523],"topic":[],"class_list":["post-70839","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-7445","tag-yolo","tag-87","tag-5134","tag-5281","tag-523"],"yoast_head":"\n