{"id":10006640,"date":"2022-09-30T21:52:06","date_gmt":"2022-09-30T13:52:06","guid":{"rendered":"https:\/\/ityet.com\/?p=10006640"},"modified":"2022-09-30T21:52:06","modified_gmt":"2022-09-30T13:52:06","slug":"easycv%e5%b8%a6%e4%bd%a0%e5%a4%8d%e7%8e%b0%e6%9b%b4%e5%a5%bd%e6%9b%b4%e5%bf%ab%e7%9a%84%e8%87%aa%e7%9b%91%e7%9d%a3%e7%ae%97%e6%b3%95-fastconvmae","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10006640","title":{"rendered":"EasyCV\u5e26\u4f60\u590d\u73b0\u66f4\u597d\u66f4\u5feb\u7684\u81ea\u76d1\u7763\u7b97\u6cd5-FastConvMAE"},"content":{"rendered":"<p>\u7b80\u4ecb\uff1a\u00a0\u8fd1\u671fFastConvMAE\u5de5\u4f5c\u5728EasyCV\u6846\u67b6\u5185\u9996\u6b21\u5bf9\u5916\u5f00\u6e90\uff0c\u672c\u6587\u5c06\u91cd\u70b9\u4ecb\u7ecdConvMAE\u548cFastConvMAE\u7684\u4e3b\u8981\u5de5\u4f5c\uff0c\u4ee5\u53ca\u5bf9\u5e94\u7684\u4ee3\u7801\u5b9e\u73b0\uff0c\u6700\u540e\u63d0\u4f9b\u8be6\u7ec6\u7684\u6559\u7a0b\u793a\u4f8b\u5982\u4f55\u8fdb\u884cFastConvMAE\u7684\u9884\u8bad\u7ec3\u548c\u4e0b\u6e38\u4efb\u52a1\u7684finetune\u3002\u4f5c\u8005\uff1a \u00a0\u5915\u964c\u3001\u8c26\u8a00\u3001\u83ab\u7533\u7ae5\u3001\u4e34\u5728\u5bfc\u8bfb\u81ea\u76d1\u7763\u5b66\u4e60\uff08Self-Supervised Learning\uff09\u5229\u7528\u5927\u91cf\u65e0\u6807\u6ce8\u7684\u6570\u636e\u8fdb\u884c\u8868\u5f81\u5b66\u4e60\uff0c\u5728\u7279\u5b9a\u4e0b\u6e38\u4efb\u52a1\u4e0a\u5bf9\u53c2\u6570\u8fdb\u884c\u5fae\u8c03\uff0c\u6781\u5927\u964d\u4f4e\u4e86\u56fe\u50cf\u4efb\u52a1\u7e41\u91cd\u7684\u6807\u6ce8\u5de5\u4f5c\uff0c\u8282\u7701\u5927\u91cf\u4eba\u529b\u6210\u672c\u3002\u8fd1\u5e74\u6765\uff0c\u81ea\u76d1\u7763\u5b66\u4e60\u5728\u89c6\u89c9\u9886\u57df\u5927\u653e\u5f02\u5f69\uff0c\u53d7\u5230\u4e86\u8d8a\u6765\u8d8a\u591a\u7684\u5173\u6ce8\u3002\u5728CV\u9886\u57df\u6d8c\u73b0\u4e86\u5982SIMCLR\u3001MOCO\u3001SwAV\u3001DINO\u3001MoBY\u3001MAE\u7b49\u4e00\u7cfb\u5217\u5de5\u4f5c\u3002\u5176\u4e2dMAE\u7684\u8868\u73b0\u5c24\u4e3a\u60ca\u8273\uff0c\u5927\u5bb6\u90fd\u88abMAE\u7b80\u6d01\u9ad8\u6548\u7684\u6027\u80fd\u6240\u5438\u5f15\uff0c\u7eb7\u7eb7\u5728 MAE\u4e0a\u8fdb\u884c\u6539\u8fdb\uff0c\u4f8b\u5982MixMIM\uff0cVideoMAE\u7b49\u5de5\u4f5c\u3002MAE\u8be6\u89e3\u8bf7\u53c2\u8003\u5f80\u671f\u6587\u7ae0\uff1aMAE\u81ea\u76d1\u7763\u7b97\u6cd5\u4ecb\u7ecd\u548c\u57fa\u4e8eEasyCV\u7684\u590d\u73b0\u00a0\u00a0 \u3002ConvMAE\u662f\u7531\u4e0a\u6d77\u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4\u548cmmlab\u8054\u5408\u53d1\u8868\u5728NeurIPS2022\u7684\u4e00\u9879\u5de5\u4f5c\uff0c\u4e0eMAE\u76f8\u6bd4\uff0c\u8bad\u7ec3\u76f8\u540c\u7684epoch\u6570\uff0c ImageNet-1K \u6570\u636e\u96c6\u7684finetune\u51c6\u786e\u7387\u63d0\u9ad8\u4e86 1.4%\uff0cCOCO2017\u6570\u636e\u96c6\u4e0a\u5fae\u8c03 25 \u4e2a epoch\u76f8\u6bd4\u5fae\u8c03100 \u4e2a epoch \u7684 MAE \u00a0AP box\u63d0\u53472.9\uff0c AP mask\u63d0\u53472.2\uff0c \u8bed\u4e49\u5206\u5272\u4efb\u52a1\u4e0a\u76f8\u6bd4MAE mIOU\u63d0\u53473.6%\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86FastConvMAE\uff0c\u8fdb\u4e00\u6b65\u4f18\u5316\u4e86\u8bad\u7ec3\u6027\u80fd\uff0c\u4ec5\u9884\u8bad\u7ec350\u4e2aepoch\uff0cImageNet Finetuning\u7684\u7cbe\u5ea6\u5c31\u8d85\u8fc7MAE\u9884\u8bad\u7ec31600\u4e2aepoch\u7684\u7cbe\u5ea60.77\u4e2a\u70b9\uff0883.6\/84.37)\u3002\u5728\u68c0\u6d4b\u4efb\u52a1\u4e0a\uff0c\u7cbe\u5ea6\u4e5f\u8d85\u8fc7ViTDet\u548cSwin\u3002EasyCV\u662f\u963f\u91cc\u5df4\u5df4\u5f00\u6e90\u7684\u57fa\u4e8ePytorch\uff0c\u4ee5\u81ea\u76d1\u7763\u5b66\u4e60\u548cTransformer\u6280\u672f\u4e3a\u6838\u5fc3\u7684 all-in-one \u89c6\u89c9\u7b97\u6cd5\u5efa\u6a21\u5de5\u5177\uff0c\u8986\u76d6\u4e3b\u6d41\u7684\u89c6\u89c9\u5efa\u6a21\u4efb\u52a1\u4f8b\u5982\u56fe\u50cf\u5206\u7c7b\uff0c\u5ea6\u91cf\u5b66\u4e60\uff0c\u76ee\u6807\u68c0\u6d4b\uff0c\u5b9e\u4f8b\/\u8bed\u97f3\/\u5168\u666f\u5206\u5272\u3001\u5173\u952e\u70b9\u68c0\u6d4b\u7b49\u9886\u57df\uff0c\u5177\u6709\u8f83\u5f3a\u7684\u6613\u7528\u6027\u548c\u6269\u5c55\u6027\uff0c\u540c\u65f6\u6ce8\u91cd\u6027\u80fd\u8c03\u4f18\uff0c\u65e8\u5728\u4e3a\u793e\u533a\u5e26\u6765\u66f4\u591a\u66f4\u5feb\u66f4\u5f3a\u7684\u7b97\u6cd5\u3002\u00a0 \u00a0 \u00a0 \u00a0\u8fd1\u671fFastConvMAE\u5de5\u4f5c\u5728EasyCV\u6846\u67b6\u5185\u9996\u6b21\u5bf9\u5916\u5f00\u6e90\uff0c\u672c\u6587\u5c06\u91cd\u70b9\u4ecb\u7ecdConvMAE\u548cFastConvMAE\u7684\u4e3b\u8981\u5de5\u4f5c\uff0c\u4ee5\u53ca\u5bf9\u5e94\u7684\u4ee3\u7801\u5b9e\u73b0\uff0c\u6700\u540e\u63d0\u4f9b\u8be6\u7ec6\u7684\u6559\u7a0b\u793a\u4f8b\u5982\u4f55\u8fdb\u884cFastConvMAE\u7684\u9884\u8bad\u7ec3\u548c\u4e0b\u6e38\u4efb\u52a1\u7684finetune\u3002ConvMAEConvMAE\u662f\u7531\u4e0a\u6d77\u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4\u548cmmlab\u8054\u5408\u53d1\u8868\u5728NeurIPS2022\u91cc\u7684\u4e00\u9879\u5de5\u4f5c\uff0cConvMAE\u7684\u63d0\u51fa\u8bc1\u660e\u4e86\u4f7f\u7528\u5c40\u90e8\u5f52\u7eb3\u504f\u7f6e\u548c\u591a\u5c3a\u5ea6\u7684\u91d1\u5b57\u5854\u7ed3\u6784\uff0c\u901a\u8fc7MAE\u7684\u8bad\u7ec3\u65b9\u5f0f\u53ef\u4ee5\u5b66\u4e60\u5230\u66f4\u597d\u7684\u7279\u5f81\u8868\u793a\u3002\u8be5\u5de5\u4f5c\u63d0\u51fa\uff1a\u4f7f\u7528block-wise mask\u7b56\u7565\u6765\u786e\u4fdd\u8ba1\u7b97\u6548\u7387\u3002\u8f93\u51fa\u7f16\u7801\u5668\u7684\u591a\u5c3a\u5ea6\u7279\u5f81\uff0c\u540c\u65f6\u6355\u83b7\u7ec6\u7c92\u5ea6\u548c\u7c97\u7c92\u5ea6\u56fe\u50cf\u4fe1\u606f\u3002\u539f\u6587\u53c2\u8003\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=G6BguHIEkUzKHd2U7oKc6A%3D%3D.6KuhJk0yahGUQBZXO0aUBB70yJbXTODCp9IY%2B27360NCkiQaTiM2Foq%2BChQSA3kH\" rel=\"nofollow\">https:\/\/arxiv.org\/abs\/2205.03892<\/a>\u00a0 \u00a0 \u00a0 \u00a0\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u4e0a\u8ff0\u4e24\u9879\u7b56\u7565\u662f\u7b80\u6d01\u800c\u6709\u6548\u7684\uff0c\u4f7f\u5f97ConvMAE\u5728\u591a\u4e2a\u89c6\u89c9\u4efb\u52a1\u4e2d\u76f8\u6bd4MAE\u83b7\u5f97\u4e86\u660e\u663e\u63d0\u5347\u3002\u4ee5ConvMAE-Base\u548cMAE-Base\u76f8\u6bd4\u4e3a\u4f8b\uff1a\u5728\u56fe\u50cf\u5206\u7c7b\u4efb\u52a1\u4e0a\uff0c ImageNet-1K \u6570\u636e\u96c6\u7684\u5fae\u8c03\u51c6\u786e\u7387\u63d0\u9ad8\u4e86 1.4%\uff1b\u5728\u76ee\u6807\u68c0\u6d4b\u4efb\u52a1\u4e0a\uff0cCOCO2017\u5fae\u8c03 25 \u4e2a epoch \u7684AP box\u8fbe\u523053.2%\uff0cAP mask\u8fbe\u523047.1%\uff0c\u4e0e\u5fae\u8c03100 \u4e2a epoch \u7684 MAE-Base\u76f8\u6bd4\u5206\u522b\u63d0\u53472.9% \u548c 2.2% \uff1b\u5728\u8bed\u4e49\u5206\u5272\u4efb\u52a1\u4e0a\uff0c\u4f7f\u7528UperNet\u7f51\u7edc\u5934\uff0cConvMAE-Base\u5728ADE20K\u4e0a\u7684mIoU\u8fbe\u523051.7%\uff0c\u76f8\u6bd4MAE-Base\u63d0\u53473.6%\u3002<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042566830\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u00a0 \u00a0ConvMAE\u7684\u603b\u4f53\u6d41\u7a0b\u00a0\u00a0 \u00a0 \u00a0 \u00a0\u4e0eMAE\u4e0d\u540c\u7684\u662f\uff0cConvMAE\u7684\u7f16\u7801\u5668\u5c06\u8f93\u5165\u56fe\u50cf\u9010\u6b65\u62bd\u8c61\u4e3a\u591a\u5c3a\u5ea6token embedding\uff0c\u800c\u89e3\u7801\u5668\u5219\u91cd\u5efa\u88abmask\u6389\u7684tokens\u5bf9\u5e94\u7684\u50cf\u7d20\u3002\u5bf9\u4e8e\u524d\u9762stage\u90e8\u5206\u7684\u9ad8\u5206\u8fa8\u7387token embedding\uff0c\u91c7\u7528\u5377\u79ef\u5757\u5bf9\u5c40\u90e8\u8fdb\u884c\u7f16\u7801\uff0c\u5bf9\u4e8e\u540e\u9762\u7684\u4f4e\u5206\u8fa8\u7387token embedding\uff0c\u5219\u4f7f\u7528transformer\u6765\u805a\u5408\u5168\u5c40\u4fe1\u606f\u3002\u56e0\u6b64\uff0cConvMAE\u7684\u7f16\u7801\u5668\u5728\u4e0d\u540c\u9636\u6bb5\u53ef\u4ee5\u540c\u65f6\u83b7\u5f97\u5c40\u90e8\u548c\u5168\u5c40\u4fe1\u606f\uff0c\u5e76\u751f\u6210\u591a\u5c3a\u5ea6\u7279\u5f81\u3002\u00a0 \u00a0 \u00a0 \u00a0\u5f53\u524d\u7684masked auto encoding\u6846\u67b6\uff0c\u5982BEiT\uff0cSimMIM\uff0c\u6240\u91c7\u7528\u7684mask\u7b56\u7565\u4e0d\u80fd\u76f4\u63a5\u7528\u4e8eConvMAE\uff0c\u56e0\u4e3a\u5728\u540e\u9762\u7684transformer\u9636\u6bb5\uff0c\u6240\u6709\u7684tokens\u90fd\u9700\u8981\u4fdd\u7559\u3002\u8fd9\u5bfc\u81f4\u5bf9\u5927\u6a21\u578b\u8fdb\u884c\u9884\u8bad\u7ec3\u7684\u8ba1\u7b97\u6210\u672c\u8fc7\u9ad8\uff0c\u5931\u53bb\u4e86MAE\u5728transformer\u7f16\u7801\u5668\u4e2d\u7701\u53bbmasked tokens\u7684\u6548\u7387\u4f18\u52bf\u3002\u6b64\u5916\uff0c\u76f4\u63a5\u4f7f\u7528convolution-transformer\u7ed3\u6784\u7684\u7f16\u7801\u5668\u8fdb\u884c\u9884\u8bad\u7ec3\u4f1a\u5bfc\u81f4\u5377\u79ef\u90e8\u5206\u56e0\u4e3a\u968f\u673a\u7684mask\u800c\u9020\u6210\u9884\u8bad\u7ec3\u7684\u4fe1\u606f\u6cc4\u9732\uff0c\u56e0\u800c\u4e5f\u4f1a\u964d\u4f4e\u9884\u8bad\u7ec3\u6240\u5f97\u6a21\u578b\u7684\u8d28\u91cf\u3002\u00a0 \u00a0 \u00a0 \u00a0 \u9488\u5bf9\u8fd9\u4e9b\u95ee\u9898\uff0cConvMAE\u63d0\u51fa\u4e86\u6df7\u5408convolution-transformer\u67b6\u6784\u3002ConvMAE\u91c7\u7528\u5206\u5757mask\u7b56\u7565 (block-wise masking strategy)\uff1a\uff0c\u9996\u5148\u968f\u673a\u5728\u540e\u671f\u7684\u83b7\u53d6transformer token\u4e2d\u751f\u6210\u540e\u671f\u7684mask\uff0c\u7136\u540e\u5bf9mask\u56fa\u5b9a\u4f4d\u7f6e\u9010\u6b65\u8fdb\u884c\u4e0a\u91c7\u6837\u5230\u65e9\u671f\u5377\u79ef\u9636\u6bb5\u7684\u9ad8\u5206\u8fa8\u7387\u3002\u8fd9\u6837\uff0c\u540e\u671f\u5904\u7406\u7684token\u53ef\u4ee5\u5b8c\u5168\u5206\u79bb\u4e3amasked tokens\u548cvisible tokens\uff0c\u4ece\u800c\u5e76\u7ee7\u627f\u4e86MAE\u4f7f\u7528\u7a00\u758fencoder\u7684\u8ba1\u7b97\u6548\u7387\u3002\u00a0 \u00a0 \u00a0 \u00a0 \u4e0b\u9762\u5c06\u5206\u522b\u9488\u5bf9encoder\u3001mask\u7b56\u7565\u4ee5\u53cadecoder\u90e8\u5206\u5c55\u5f00\u4ecb\u7ecd\u3002Encoder\u5982\u603b\u4f53\u6d41\u7a0b\u56fe\u6240\u793a\uff0cencoder\u5305\u62ec3 \u4e2a\u9636\u6bb5\uff0c\u6bcf\u4e2a\u9636\u6bb5\u8f93\u51fa\u7684\u7279\u5f81\u7ef4\u5ea6\u5206\u522b\u662f\uff1aH\/4 \u00d7 W\/4, H\/8 \u00d7 W\/8, H\/16 \u00d7 W\/16\uff0c\u5176\u4e2dH \u00d7 W\u4e3a\u8f93\u5165\u56fe\u50cf\u5206\u8fa8\u7387\u3002\u524d\u4e24\u4e2a\u662f\u5377\u79ef\u9636\u6bb5\uff0c\u4f7f\u7528\u5377\u79ef\u6a21\u5757\u5c06\u8f93\u5165\u8f6c\u6362\u4e3atoken embeddings\u00a0E1 \u2208 R^(H\/4 \u00d7 W\/4 \u00d7C1) and E2 \u2208 R^(H\/8 \u00d7 W\/8 \u00d7C2) \u3002\u5176\u4e2d\u5377\u79ef\u6a21\u5757\u75285 \u00d7 5\u7684\u5377\u79ef\u4ee3\u66ffself-attention\u64cd\u4f5c\u3002\u524d\u4e24\u4e2a\u9636\u6bb5\u7684\u611f\u53d7\u91ce\u8f83\u5c0f\u4e3b\u8981\u6355\u6349\u56fe\u50cf\u7684\u5c40\u90e8\u7279\u5f81\uff0c\u7b2c\u4e09\u4e2a\u9636\u6bb5\u4f7f\u7528transformer\u6a21\u5757\uff0c\u5c06\u7c97\u7c92\u5ea6\u7279\u5f81\u878d\u5408, \u5e76\u5c06\u611f\u53d7\u91ce\u6269\u5c55\u5230\u6574\u4e2a\u56fe\u50cf\uff0c\u83b7\u5f97token embeddings E3 \u2208 R(H\/16 \u00d7 W\/16 \u00d7C3)\u3002\u5728\u6bcf\u4e2a\u9636\u6bb5\u4e4b\u95f4\uff0c\u4f7f\u7528stride\u4e3a2\u7684\u5377\u79ef\u5bf9tokens\u8fdb\u884c\u4e0b\u91c7\u6837\u3002\u5176\u4ed6\u5305\u542btransformer\u7684\u7ed3\u6784\uff0c\u5982CPT\u3001Container\u3001Uniformer\u3001CMT\u3001Swin\u7b49\uff0c\u5728\u7b2c\u4e00\u9636\u6bb5\u7684\u8f93\u5165\u7528\u76f8\u5bf9\u4f4d\u7f6e\u7f16\u7801\u6216\u96f6\u586b\u5145\u5377\u79ef\u66ff\u4ee3\u7edd\u5bf9\u4f4d\u7f6e\u7f16\u7801\uff0c\u800c\u4f5c\u8005\u53d1\u73b0\u5728\u7b2c3\u4e2atransformer stage\u4e2d\u4f7f\u7528\u7edd\u5bf9\u4f4d\u7f6e\u7f16\u7801\u53ef\u83b7\u5f97\u6700\u4f18\u6027\u80fd\u3002class token\u4e5f\u4ece\u7f16\u7801\u5668\u4e2d\u79fb\u9664\u3002Mask\u7b56\u7565MAE\u3001BEiT\u7b49\uff0c\u5bf9\u8f93\u5165patch\u91c7\u7528\u968f\u673amask\u3002\u4f46\u540c\u6837\u7684\u7b56\u7565\u4e0d\u80fd\u76f4\u63a5\u5e94\u7528\u4e8eConvMAE\u7f16\u7801\u5668\uff1a\u5982\u679c\u72ec\u7acb\u5730\u4ecestage-1\u7684H\/4 \u00d7 W\/4\u4e2atokens\u4e2d\u968f\u673a\u62bd\u53d6mask\uff0c\u5c06\u5bfc\u81f4\u964d\u91c7\u6837\u540e\u7684stage-3\u7684\u51e0\u4e4e\u6240\u6709token\u90fd\u6709\u90e8\u5206\u53ef\u89c1\u4fe1\u606f\uff0c\u4f7f\u5f97\u7f16\u7801\u5668\u4e0d\u518d\u7a00\u758f\u3002\u56e0\u6b64\u4f5c\u8005\u63d0\u51fa\uff0c\u4ecestage-3\u7684\u8f93\u5165tokens\u4e2d\u4ee5\u540c\u6837\u6bd4\u4f8b \uff08\u4f8b\u598275%\uff09\u751f\u6210mask\uff0c\u518d\u5bf9mask\u4e0a\u91c7\u68372\u500d\u548c4\u500d\uff0c\u5206\u522b\u4f5c\u4e3astage-2\u548cstage-1\u7684mask\u3002\u8fd9\u6837\uff0cConvMAE\u57283\u4e2a\u9636\u6bb5\u90fd\u53ea\u542b\u6709\u5f88\u5c11\u7684\uff08\u4f8b\u598225%\uff09\u53ef\u89c1token\uff0c\u4ece\u800c\u4f7f\u5f97\u9884\u8bad\u7ec3\u65f6\u7f16\u7801\u5668\u7684\u6548\u7387\u4e0d\u53d7\u5f71\u54cd\u3002\u800c\u89e3\u7801\u5668\u7684\u4efb\u52a1e\u5219\u4fdd\u6301\u76f8\u540c\uff0c\u5373\u91cd\u5efa\u7f16\u7801\u8fc7\u7a0b\u4e2d\u88abmask\u6389\u7684tokens\u3002\u540c\u65f6\uff0c\u524d2\u4e2a\u9636\u6bb5\u76845X5\u5377\u79ef\u64cd\u4f5c\u4f1a\u5728masked patches\u7684\u8fb9\u7f18\u5904\u6cc4\u6f0f\u4e0d\u53ef\u89c1token\u7684\u91cd\u5efa\u7b54\u6848\u3002\u4e3a\u4e86\u907f\u514d\u8fd9\u79cd\u60c5\u51b5\u4fdd\u8bc1\u9884\u8bad\u7ec3\u7684\u8d28\u91cf\uff0c\u4f5c\u8005\u5728\u524d\u4e24\u4e2a\u9636\u6bb5\u91c7\u7528\u4e86masked convolution, \u4f7f\u88abmask\u6389\u7684\u533a\u57df\u4e0d\u53c2\u4e0e\u7f16\u7801\u8fc7\u7a0b\u3002Decoder\u539f\u59cbMAE\u7684decoder\u7684\u8f93\u5165\u4ee5\u7f16\u7801\u5668\u7684\u8f93\u51fa\u548cmask\u6389\u7684tokens\u4f5c\u4e3a\u8f93\u5165\uff0c\u7136\u540e\u901a\u8fc7\u5806\u53e0\u7684transformer blocks\u8fdb\u884c\u56fe\u50cf\u91cd\u5efa\u3002ConvMAE\u7f16\u7801\u5668\u83b7\u5f97\u591a\u5c3a\u5ea6\u7279\u5f81E1\u3001E2\u3001E3\uff0c\u540c\u65f6\u6355\u83b7\u7ec6\u7c92\u5ea6\u548c\u7c97\u7c92\u5ea6\u56fe\u50cf\u4fe1\u606f\u3002\u4e3a\u4e86\u66f4\u597d\u5730\u7684\u9884\u8bad\u7ec3\uff0c\u4f5c\u8005\u901a\u8fc7stride-4\u548cstride-2\u5377\u79ef\u5c06E1\u548cE2\u4e0b\u91c7\u6837\u5230E3\u7684\u76f8\u540c\u5927\u5c0f\uff0c\u5e76\u8fdb\u884c\u591a\u5c3a\u5ea6\u7279\u5f81\u878d\u5408\uff0c\u518d\u901a\u8fc7\u4e00\u4e2alinear\u5c42\u5f97\u5230\u6700\u7ec8\u8981\u8f93\u5165\u7ed9 decoder \u7684\u53ef\u89c1token\u3002\u76ee\u6807\u51fd\u6570\u548cMAE\u76f8\u540c\uff0c\u4ec5\u91c7\u7528MSE\u4f5c\u4e3a\u635f\u5931\u51fd\u6570\uff0c\u8ba1\u7b97\u9884\u6d4b\u5411\u91cf\u548c\u88abmask\u6389\u50cf\u7d20\u503c\u4e4b\u524d\u7684MSE loss\uff0c\u5373\u53ea\u8003\u8651mask\u6389\u7684patches\u7684\u91cd\u5efa\u3002\u4e0b\u6e38\u4efb\u52a1\u00a0<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042566831\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u9884\u8bad\u7ec3\u4e4b\u540e\uff0cConvMAE\u53ef\u4ee5\u8f93\u51fa\u591a\u5c3a\u5ea6\u7684\u7279\u5f81\u7528\u4e8e\u68c0\u6d4b\u5206\u5272\u4efb\u52a1\u3002\u68c0\u6d4b\u4efb\u52a1\u4e2d\uff0c\u5148\u5c06\u7b2cstage-3\u7684\u8f93\u51fa\u7279\u5f81E3\u901a\u8fc72&#215;2\u6700\u5927\u6c60\u5316\u83b7\u5f97E4\u3002\u7531\u4e8eConvMAE stage-3\u670911\u4e2aself-attention\u5c42(ConvMAE-base)\uff0c\u8ba1\u7b97\u6210\u672c\u8fc7\u9ad8\uff0c\u4f5c\u8005\u53c2\u8003ViT\u7684benchmark\u5c06stage-3\u4e2d\u9664\u7b2c1\u30014\u30017\u300111\u4e4b\u5916\u7684\u6240\u6709global self-attention layers\u66ff\u6362\u4e3a\u4e86Window size7\u00d77 \u7684 \u00a0local self-attention \u5c42\u3002\u4fee\u6539\u540e\u7684local self-attention\u4ecd\u7136\u7531\u9884\u8bad\u7ec3\u7684global self-attention\u8fdb\u884c\u521d\u59cb\u5316\u3002global transformer blocks\u4e4b\u95f4\u5171\u4eabglobal relative position bias\uff0clocal transformer blocks\u4e4b\u95f4\u5171\u4eablocal relative position bias\uff0c\u8fd9\u6837\u5c31\u5927\u5927\u51cf\u8f7b\u4e86stage-3\u7684\u8ba1\u7b97\u548cGPU\u5185\u5b58\u5f00\u9500\u3002\u7136\u540e\u5c06\u591a\u5c3a\u5ea6\u7279\u5f81E1\u3001E2\u3001E3\u3001E4\u9001\u5165MaskRCNN head\u8fdb\u884c\u76ee\u6807\u68c0\u6d4b\u3002\u800c\u5206\u5272\u4efb\u52a1\u4fdd\u7559\u4e86stage-3\u7684\u7ed3\u6784\u3002Benchmark\u56fe\u50cf\u5206\u7c7bConvMAE\u57fa\u4e8eImageNet-1K\uff0cmask\u638925%\u7684input token\u505a\u9884\u8bad\u7ec3\uff0cDecoder\u90e8\u5206\u662f\u4e00\u4e2a8\u5c42\u7684transformer\uff0cembedding \u7ef4\u5ea6\u662f512\uff0chead\u662f12\u4e2a\u3002\u9884\u8bad\u7ec3\u53c2\u6570\u548c\u5206\u7c7bfinetuning\u7ed3\u679c\u5982\u4e0b\uff1a<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042566832\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>BEiT\u9884\u8bad\u7ec3300\u4e2aepoch\uff0cfinetune\u7684\u7cbe\u5ea6\u8fbe\u523083.0%\uff0clinear-prob\u7684\u7cbe\u5ea6\u662f37.6%\u3002\u4e0eBEiT\u76f8\u6bd4\uff0cConVMAE\u4ec5\u9700\u898125%\u7684token\u548c\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684decoder finetune\u53ef\u8fbe\u523085%\uff0clinear-prob\u53ef\u4ee5\u8fbe\u523070.9%\u3002\u4e0e\u539f\u6765\u7684MAE\u76f8\u6bd4\uff0c\u9884\u8bad\u7ec3\u76f8\u540c\u76841600\u4e2aepoch\uff0cConVMAE\u6bd4MAE\u63d0\u53471.4\u4e2a\u70b9\u3002\u4e0eSimMIM\uff08backbone\u4f7f\u7528Swin-B\uff09\u76f8\u6bd4\u63d0\u5347\u4e861\u4e2a\u70b9\u3002\u68c0\u6d4b\u4f5c\u8005\u7528ConvMAE\u66ff\u6362Mask-RCNN\u7684backbone\uff0c\u52a0\u8f7dConvMAE\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u8bad\u7ec3COCO\u6570\u636e\u96c6\u3002<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042566833\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042566834\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u4e0eViT\u5728COCO\u6570\u636e\u96c6\u4e0afinetune100\u4e2aepoch\u7684\u7ed3\u679c\u76f8\u6bd4\uff0cConVMAE\u4ec5finetune 25\u4e2aepoch\u5728APbox\u548cAPmask\u5c31\u63d0\u5347\u4e862.9\u548c2.2\u4e2a\u70b9\u3002\u4e0eViTDet\u548cMIMDet\u76f8\u6bd4\uff0cConvMAE finetune epoch\u66f4\u5c11\u3001\u53c2\u6570\u66f4\u5c11\uff0c\u5206\u522b\u8d85\u8fc7\u4e86\u5b83\u4eec2.0%\u548c1.7%\u3002\u4e0eSwin\u548cMViTv2\u76f8\u6bd4\uff0c\u5728APbox\/APmask\uff0c\u5176\u6027\u80fd\u5206\u522b\u9ad8\u51fa4.0%\/3.6%\u548c2.2%\/1.4%\u3002\u5206\u5272\u4f5c\u8005\u7528ConvMAE\u66ff\u6362UperNet\u7684backbone\uff0c\u52a0\u8f7dConvMAE\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u8bad\u7ec3ADE20K\u6570\u636e\u96c6\u3002<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042566835\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u4ece\u7ed3\u679c\u4e2d\u53ef\u4ee5\u770b\u51fa\uff0c\u76f8\u6bd4\u4e0eDeiT, Swin\uff0cMoCo-v3\u7b49\u7f51\u7edcConvMAE\u53d6\u5f97\u4e86\u66f4\u9ad8\u7684\u6027\u80fd(51.7%)\u3002\u8868\u660eConvMAE\u7684\u591a\u5c3a\u5ea6\u7279\u5f81\u5927\u5927\u7f29\u5c0f\u4e86\u9884\u8bad\u7ec3Backbone \u548c\u4e0b\u6e38\u7f51\u7edc\u4e4b\u95f4\u7684\u4f20\u8f93\u5dee\u8ddd\u3002Fast ConvMAEConvMAE\u867d\u7136\u5728\u5206\u7c7b\u3001\u68c0\u6d4b\u3001\u5206\u5272\u7b49\u4e0b\u6e38\u4efb\u52a1\u4e2d\u6709\u4e86\u7cbe\u5ea6\u63d0\u5347\uff0c\u5e76\u89e3\u51b3\u4e86pretraining-finetuning \u7684\u5dee\u5f02\u95ee\u9898\uff0c\u4f46\u662f\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u4f9d\u7136\u8017\u65f6\uff0cConvMAE\u7684\u7ed3\u679c\u4e2d\uff0c\u6a21\u578b\u9884\u8bad\u7ec3\u4e861600\u4e2aepoch\uff0c\u56e0\u6b64\u4f5c\u8005\u53c8\u5728ConvMAE\u7684\u57fa\u7840\u4e4b\u4e0a\u505a\u4e86\u8fdb\u4e00\u6b65\u7684\u6027\u80fd\u4f18\u5316\uff0c\u63d0\u51fa\u4e86Fast ConvMAE\uff0cFastConvMAE\u63d0\u51fa\u4e86mask\u4e92\u8865\u548cdeocder\u878d\u5408\u7684\u65b9\u6848\uff0c\u6765\u5b9e\u73b0\u5feb\u901f\u7684mask\u5efa\u6a21\u65b9\u6848\uff0c\u8fdb\u4e00\u6b65\u7f29\u77ed\u4e86\u9884\u8bad\u7ec3\u7684\u65f6\u95f4\uff0c\u4ece\u539f\u6765\u9884\u8bad\u7ec3\u76841600epoch\u7f29\u77ed\u5230\u4e8650epoch\u3002FastConvMAE\u7684\u6b63\u5f0f\u8bba\u6587\u4f5c\u8005\u4f1a\u5728\u672a\u6765\u53d1\u51fa\u3002<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042566836\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u9996\u5148\uff0cFastConvMAE\u521b\u65b0\u5730\u8bbe\u8ba1\u51fadecoder\u4e92\u76f8\u878d\u5408\u7684Mixture of Reconstructor (MoR)\uff0c\u53ef\u4ee5\u8ba9masked patches\u4ece\u4e0d\u540c\u7684tokenizer\u4e2d\u5b66\u4e60\u5230\u4e92\u8865\u7684\u4fe1\u606f\uff0c\u5305\u62ecEMA \u7684self-ensembling\u6027\u8d28\uff0cDINO\u7684similarity-discrimination\u80fd\u529b,\u4ee5\u53caCLIP\u7684multimodal\u77e5\u8bc6\u3002MoR\u4e3b\u8981\u5305\u62ec\u4e24\u4e2a\u90e8\u5206\uff0cPartially-Shared Decoder(PS-Decoder)\u548cMixture of Tokenizer(MoT), PS-Decoder\u53ef\u4ee5\u907f\u514d\u4e0d\u540ctokenizer\u7684\u4e0d\u540c\u77e5\u8bc6\u4e4b\u95f4\u4f1a\u4ea7\u751f\u68af\u5ea6\u7684\u51b2\u7a81\uff0cMoT\u662f\u7528\u6765\u751f\u6210\u4e0d\u540c\u7684token\u4f5c\u4e3amasked patches\u7684target\u3002\u540c\u65f6Mask\u90e8\u5206\u91c7\u7528\u4e86\u4e92\u8865\u7b56\u7565\uff0c\u539f\u6765\u7684mask\u6bcf\u6b21\u53ea\u4f1a\u4fdd\u7559\u4f8b\u598225%\u7684tokens\uff0cFastConvMAE\u5c06mask\u5206\u6210\u4e864\u4efd\uff0c\u6bcf\u4e00\u4efd\u90fd\u4fdd\u755925%\uff0c4\u4efdmask\u4e4b\u95f4\u4e92\u8865\u3002\u8fd9\u6837\uff0c\u76f8\u5f53\u4e8e1\u5f20\u56fe\u7247\u88ab\u5206\u6210\u4e864\u5f20\u56fe\u7247\u8fdb\u884c\u5b66\u4e60\uff0c\u7406\u8bba\u4e0a\u8fbe\u5230\u4e864\u500d\u7684\u5b66\u4e60\u6548\u679c\u3002\u00a0    def random_masking(self, x, mask_ratio=None):<\/p>\n<pre><code>    &quot;&quot;&quot;\n    Perform per-sample random masking by per-sample shuffling.\n    Per-sample shuffling is done by argsort random noise.\n    x: [N, L, D], sequence\n    &quot;&quot;&quot;\n    N = x.shape[0]\n    L = self.num_patches\n    len_keep = int(L * (1 - mask_ratio))\n    noise = torch.rand(N, L, device=x.device)  # noise in [0, 1]\n\n    # sort noise for each sample\n    ids_shuffle = torch.argsort(\n        noise, dim=1)  # ascend: small is keep, large is remove\n    ids_restore = torch.argsort(ids_shuffle, dim=1)\n\n    # keep the first subset\n    ids_keep1 = ids_shuffle[:, :len_keep]\n    ids_keep2 = ids_shuffle[:, len_keep:2 * len_keep]\n    ids_keep3 = ids_shuffle[:, 2 * len_keep:3 * len_keep]\n    ids_keep4 = ids_shuffle[:, 3 * len_keep:]\n\n    # generate the binary mask: 0 is keep, 1 is remove\n    mask1 = torch.ones([N, L], device=x.device)\n    mask1[:, :len_keep] = 0\n    # unshuffle to get the binary mask\n    mask1 = torch.gather(mask1, dim=1, index=ids_restore)\n\n    mask2 = torch.ones([N, L], device=x.device)\n    mask2[:, len_keep:2 * len_keep] = 0\n    # unshuffle to get the binary mask\n    mask2 = torch.gather(mask2, dim=1, index=ids_restore)\n\n    mask3 = torch.ones([N, L], device=x.device)\n    mask3[:, 2 * len_keep:3 * len_keep] = 0\n    # unshuffle to get the binary mask\n    mask3 = torch.gather(mask3, dim=1, index=ids_restore)\n\n    mask4 = torch.ones([N, L], device=x.device)\n    mask4[:, 3 * len_keep:4 * len_keep] = 0\n    # unshuffle to get the binary mask\n    mask4 = torch.gather(mask4, dim=1, index=ids_restore)\n\n    return [ids_keep1, ids_keep2, ids_keep3,\n            ids_keep4], [mask1, mask2, mask3, mask4], ids_restore\u524d\u4e24\u4e2a\u5377\u79ef\u9636\u6bb5\u5c06\u8f93\u5165\u8f6c\u6362\u4e3aembeddings tokens E1\u548cE2\u3002\u7136\u540eE1\u548cE2\u5206\u522b\u4ece4\u4efdmask\u4e2d\u83b7\u53d64\u4efd\u53ef\u89c1\u7684tokens\u5e76\u8fdb\u884c\u62fc\u63a5\uff0c\u4f5c\u4e3adecoder\u7684\u8f93\u5165\uff0cDecoder\u5904\u7406\u7684\u662f\u62fc\u63a5\u540e\u7684tokens\u3002\u4ee3\u7801\u53c2\u8003\u5982\u4e0b\uff1a   def encoder_forward(self, x, mask_ratio):\n    # embed patches\n    ids_keep, masks, ids_restore = self.random_masking(x, mask_ratio)\n    mask_for_patch1 = [\n      1 - mask.reshape(-1, 14, 14).unsqueeze(-1).repeat(\n        1, 1, 1, 16).reshape(-1, 14, 14, 4, 4).permute(\n        0, 1, 3, 2, 4).reshape(x.shape[0], 56, 56).unsqueeze(1)\n      for mask in masks\n    ]\n    mask_for_patch2 = [\n      1 - mask.reshape(-1, 14, 14).unsqueeze(-1).repeat(\n        1, 1, 1, 4).reshape(-1, 14, 14, 2, 2).permute(\n        0, 1, 3, 2, 4).reshape(x.shape[0], 28, 28).unsqueeze(1)\n      for mask in masks\n    ]\n\n    s1 = self.patch_embed1(x)\n    s1 = self.pos_drop(s1)\n    for blk in self.blocks1:\n        s1 = blk(s1, mask_for_patch1)\n\n    s2 = self.patch_embed2(s1)\n    for blk in self.blocks2:\n        s2 = blk(s2, mask_for_patch2)\n\n    stage1_embed = self.stage1_output_decode(s1).flatten(2).permute(0, 2, 1)\n    stage2_embed = self.stage2_output_decode(s2).flatten(2).permute(0, 2, 1)\n    stage1_embed_1 = torch.gather(\n      stage1_embed,\n      dim=1,\n      index=ids_keep[0].unsqueeze(-1).repeat(1, 1, stage1_embed.shape[-1]))\n    stage2_embed_1 = torch.gather(\n      stage2_embed,\n      dim=1,\n      index=ids_keep[0].unsqueeze(-1).repeat(1, 1, stage2_embed.shape[-1]))\n    stage1_embed_2 = torch.gather(\n      stage1_embed,\n      dim=1,\n      index=ids_keep[1].unsqueeze(-1).repeat(1, 1, stage1_embed.shape[-1]))\n    stage2_embed_2 = torch.gather(\n      stage2_embed,\n      dim=1,\n      index=ids_keep[1].unsqueeze(-1).repeat(1, 1, stage2_embed.shape[-1]))\n    stage1_embed_3 = torch.gather(\n      stage1_embed,\n      dim=1,\n      index=ids_keep[2].unsqueeze(-1).repeat(1, 1, stage1_embed.shape[-1]))\n    stage2_embed_3 = torch.gather(\n      stage2_embed,\n      dim=1,\n      index=ids_keep[2].unsqueeze(-1).repeat(1, 1, stage2_embed.shape[-1]))\n    stage1_embed_4 = torch.gather(\n      stage1_embed,\n      dim=1,\n      index=ids_keep[3].unsqueeze(-1).repeat(1, 1, stage1_embed.shape[-1]))\n    stage2_embed_4 = torch.gather(\n      stage2_embed,\n      dim=1,\n      index=ids_keep[3].unsqueeze(-1).repeat(1, 1, stage2_embed.shape[-1]))\n    stage1_embed = torch.cat([\n      stage1_embed_1, stage1_embed_2, stage1_embed_3, stage1_embed_4\n    ])\n    stage2_embed = torch.cat([\n      stage2_embed_1, stage2_embed_2, stage2_embed_3, stage2_embed_4\n    ])\n\n    x = self.patch_embed3(s2)\n    x = x.flatten(2).permute(0, 2, 1)\n    x = self.patch_embed4(x)\n    # add pos embed w\/o cls token\n    x = x + self.pos_embed\n\n    x1 = torch.gather(x, dim=1, index=ids_keep[0].unsqueeze(-1).repeat(1, 1, x.shape[-1]))\n    x2 = torch.gather(x, dim=1, index=ids_keep[1].unsqueeze(-1).repeat(1, 1, x.shape[-1]))\n    x3 = torch.gather(x, dim=1, index=ids_keep[2].unsqueeze(-1).repeat(1, 1, x.shape[-1]))\n    x4 = torch.gather(x, dim=1, index=ids_keep[3].unsqueeze(-1).repeat(1, 1, x.shape[-1]))\n    x = torch.cat([x1, x2, x3, x4])\n\n    # apply Transformer blocks\n    for blk in self.blocks3:\n        x = blk(x)\n\n    x = x + stage1_embed + stage2_embed\n    x = self.norm(x)\n    mask = torch.cat([masks[0], masks[1], masks[2], masks[3]])\n    return x, mask, ids_restoreBenchmarkEasyCV\u590d\u73b0\u7684\u7ed3\u679c\u5982\u4e0b\uff1aImageNet PretrainedConfigEpochsDownloadfast_convmae_vit_base_patch16_8xb64_50e50model-\u00a0log\u00a0ImageNet FinetuningAlgorithmFintune ConfigPretrained ConfigTop-1DownloadFast ConvMAE\uff08EasyCV)fast_convmae_vit_base_patch16_8xb64_100e_fintunefast_convmae_vit_base_patch16_8xb64_50e84.4%fintune model-\u00a0logFast ConvMAE\uff08\u5b98\u65b9\uff09\u00a0\u00a084.4%\u00a0\u00a0Object DetectionAlgorithmEval ConfigPretrained ConfigmAP (Box)mAP (Mask)DownloadFast ConvMAE\uff08EasyCV)mask_rcnn_conv_vitdet_50e_cocofast_convmae_vit_base_patch16_8xb64_50e51.3%45.6%finetune modelFast ConvMAE\uff08\u5b98\u65b9\uff09\u00a0\u00a051.0%45.4%\u00a0\u00a0\u4ece\u7ed3\u679c\u53ef\u4ee5\u770b\u51fa\uff0c\u4ec5\u9884\u8bad\u7ec350\u4e2aepoch\uff0cImageNet Finetuning\u7684\u7cbe\u5ea6\u5c31\u8d85\u8fc7MAE\u9884\u8bad\u7ec31600\u4e2aepoch\u7684\u7cbe\u5ea60.77\u4e2a\u70b9\uff0883.6\/84.37)\u3002\u5728\u68c0\u6d4b\u4efb\u52a1\u4e0a\uff0c\u7cbe\u5ea6\u4e5f\u8d85\u8fc7ViTDet\u548cSwin\u3002FastConvMAE\u7684\u66f4\u591a\u5b98\u65b9\u7ed3\u679c\u8bf7\u53c2\u8003\uff1ahttps:\/\/github.com\/Alpha-VL\/FastConvMAE\u00a0\u3002Tutorial\u4e00\u3001\u5b89\u88c5\u4f9d\u8d56\u5305\u5982\u679c\u662f\u5728\u672c\u5730\u5f00\u53d1\u73af\u5883\u8fd0\u884c\uff0c\u53ef\u4ee5\u53c2\u8003\u8be5\u94fe\u63a5\u5b89\u88c5\u73af\u5883\u3002\u82e5\u4f7f\u7528PAI-DSW\u8fdb\u884c\u5b9e\u9a8c\u5219\u65e0\u9700\u5b89\u88c5\u76f8\u5173\u4f9d\u8d56\uff0c\u5728PAI-DSW docker\u4e2d\u5df2\u5185\u7f6e\u76f8\u5173\u73af\u5883\u3002\u4e8c\u3001\u6570\u636e\u51c6\u5907\u6570\u636e\u51c6\u5907\u8bf7\u53c2\u8003\u6587\u6863\uff1ahttps:\/\/github.com\/alibaba\/EasyCV\/blob\/master\/docs\/source\/prepare_data.md\u4e09\u3001\u6a21\u578b\u9884\u8bad\u7ec3FastConvMAE\u5360\u7528\u663e\u5b58\u8f83\u5927\uff0c\u5efa\u8bae\u4f7f\u7528A100\u8d44\u6e90\u3002\uff08FastConvMAE\u4e00\u6b21forward-backward\u7b49\u4ef7\u4e8eConvMAE forward-backward 4\u6b21\uff09\u5728EasyCV\u4e2d\uff0c\u4f7f\u7528\u914d\u7f6e\u6587\u4ef6\u7684\u5f62\u5f0f\u6765\u5b9e\u73b0\u5bf9\u6a21\u578b\u53c2\u6570\u3001\u6570\u636e\u8f93\u5165\u53ca\u589e\u5e7f\u65b9\u5f0f\u3001\u8bad\u7ec3\u7b56\u7565\u7684\u914d\u7f6e\uff0c\u4ec5\u901a\u8fc7\u4fee\u6539\u914d\u7f6e\u6587\u4ef6\u4e2d\u7684\u53c2\u6570\u8bbe\u7f6e\uff0c\u5c31\u53ef\u4ee5\u5b8c\u6210\u5b9e\u9a8c\u914d\u7f6e\u8fdb\u884c\u8bad\u7ec3\u3002\u914d\u7f6eEasyCV\u8def\u5f84# \u67e5\u770beasycv\u5b89\u88c5\u4f4d\u7f6e<\/code><\/pre>\n<p>import easycv<br \/>print(easycv.__file__)$ export PYTHONPATH=$PYTHONPATH:${your EasyCV root path}\u4e0b\u6e38\u4efb\u52a1finetune\u4e0b\u8f7d\u9884\u8bad\u7ec3\u6a21\u578b$ wget\u00a0<a href=\"https:\/\/link.segmentfault.com\/?enc=A6SiMwQJubZ%2BVlCMoFCfDA%3D%3D.adZ5szOYBeHNBXq9KqLP%2FKRptiVq3LzILDTeJfDg5MPAlP8a4KGl%2B198azNSY98AJQCkuOoJ9fyrDm1J8uwUVBNDVqS8OFWEe8htAERFDbgAJWCAOC1Ovunbr4t8VuRZ%2F0m0bPpXgxLb3Ov5TwCmty179IwW%2FsrHBYmyMD3H8tU%3D\" rel=\"nofollow\">http:\/\/pai-vision-data-hz.oss&#8230;<\/a>\u5355\u5361$ python tools\/train.py \/<br \/>${CONFIG_FILE} \/<br \/>&#8211;work_dir .\/work_dir \/<br \/>&#8211;load_from=.\/epoch_50.pth\u591a\u5361$ python -m torch.distributed.launch &#8211;nproc_per_node=8 &#8211;master_port=29930 \/<br \/>tools\/train.py \/<br \/>${CONFIG_FILE} \/<br \/>&#8211;work_dir .\/work_dir \/<br \/>&#8211;launcher pytorch \/<br \/>&#8211;load_from=.\/epoch_50.pth\u5206\u7c7b\u4efb\u52a1 CONFIG_FILE\u00a0\u8bf7\u53c2\u8003\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=u44F34YLPoar3sh8giQRDQ%3D%3D.guoFJ2DOaF4lZl2X60f7tfGhYBYH2P6cv0nGonxgpS0ZtftTchXYekgVOjYSCK40HJpjBknM4fh36D683lvfmoWbCtuZF7yXKvUgqgawAzC%2BczhZHb2fZwjAvsLqx%2BS0j9rFVPPFleCMf6eOTGNPls9%2BjTwglzdzFkGvNPqUB%2B12u7CWcfJpsCSoQ09qkRhn\" rel=\"nofollow\">https:\/\/github.com\/alibaba\/Ea&#8230;<\/a>\u5206\u7c7b\u4efb\u52a1 CONFIG_FILE\u00a0\u8bf7\u53c2\u8003\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=ZwMrShrgnb7oPpC%2F7wF03g%3D%3D.wQtXbCFOaNjmOk0MMv4OAHCMyjlm%2FpJ9HU0xsXTPAh596chfBD93vrjomgqFQHDO%2BDOQCigxKynBHtNS2xk1SKx7uku4odd6ACNJKpBcSJphsuVDpL%2FdDfXDtcuMjQho7ovuoEns6uvOFSUdcY42gDAPpAr0TspyjC8G2%2BS5uyrn%2B%2B0xdgotW9DYNyg1syZi\" rel=\"nofollow\">https:\/\/github.com\/alibaba\/Ea&#8230;<\/a>\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=4%2FJ%2B8IKUMJP0fOckw%2BX5wA%3D%3D.cRA8JHOqocD2%2BS7RcjQhaWeUQIwIDyvX8JOFkyw4phycPYzWF3AT94Fk%2Fn2F6khuNw4F579IJQz6Jo0C40Vc0oTWfOqo1TC7CN3aHPCrVBJ1EhkZm2zTqhw%2BOlZNQlP8\" rel=\"nofollow\">https:\/\/github.com\/alibaba\/Ea&#8230;<\/a>\u5f80\u671f\u5206\u4eabYOLOX-PAI:\u52a0\u901fYOLOX,\u6bd4YOLOV6\u66f4\u5feb\u66f4\u5f3a\u57fa\u4e8eEasyCV\u590d\u73b0DETR\u548cDAB-DETR\uff0cObject Query\u7684\u6b63\u786e\u6253\u5f00\u65b9\u5f0f\u57fa\u4e8eEasyCV\u590d\u73b0ViTDet\uff1a\u5355\u5c42\u7279\u5f81\u8d85\u8d8aFPNMAE\u81ea\u76d1\u7763\u7b97\u6cd5\u4ecb\u7ecd\u548c\u57fa\u4e8eEasyCV\u7684\u590d\u73b0EasyCV\u5f00\u6e90\uff5c\u5f00\u7bb1\u5373\u7528\u7684\u89c6\u89c9\u81ea\u76d1\u7763+Transformer\u7b97\u6cd5\u5e93\u539f\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=AxbaFstoUXEbayIjwXvxqg%3D%3D.nB1cKV4k0KNSTSmVJlq1%2F1UhJbR%2Fy4SdkrvUpDX7VfHjSQuqyTxS5sFYhVKvGwaE\" rel=\"nofollow\">https:\/\/click.aliyun.com\/m\/10&#8230;<\/a>\u672c\u6587\u4e3a\u963f\u91cc\u4e91\u539f\u521b\u5185\u5bb9\uff0c\u672a\u7ecf\u5141\u8bb8\u4e0d\u5f97\u8f6c\u8f7d\u3002<\/p>\n<p><a href=\"https:\/\/segmentfault.com\/a\/1190000042566828\">https:\/\/segmentfault.com\/a\/1190000042566828<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7b80\u4ecb\uff1a\u00a0\u8fd1\u671fFastConvMAE\u5de5\u4f5c\u5728EasyCV\u6846\u67b6\u5185\u9996\u6b21\u5bf9\u5916\u5f00\u6e90\uff0c\u672c\u6587\u5c06\u91cd\u70b9\u4ecb\u7ecdConvMAE\u548cFast [&hellip;]<\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[],"tags":[],"class_list":["post-10006640","post","type-post","status-publish","format-standard","hentry"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006640","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10006640"}],"version-history":[{"count":0,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006640\/revisions"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10006640"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10006640"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10006640"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}