{"id":8319,"date":"2022-09-25T20:49:41","date_gmt":"2022-09-25T12:49:41","guid":{"rendered":"http:\/\/139.9.1.231\/?p=8319"},"modified":"2022-10-07T20:34:07","modified_gmt":"2022-10-07T12:34:07","slug":"unext","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/09\/25\/unext\/","title":{"rendered":"MICCAI 2022\uff1a\u57fa\u4e8e MLP \u7684\u5feb\u901f\u533b\u5b66\u56fe\u50cf\u5206\u5272\u7f51\u7edc UNeXt"},"content":{"rendered":"\n

\u8bba\u6587\u5730\u5740: https:\/\/arxiv.org\/abs\/2203.04967<\/a><\/strong><\/p>\n\n\n\n

github:https:\/\/github.com\/jeya-maria-jose\/UNeXt-pytorch<\/a><\/em><\/strong><\/p>\n\n\n\n

\"\"\/
UnetX \u7f51\u7edc\u7ed3\u6784<\/figcaption><\/figure>\n\n\n\n

Datasets<\/h4>\n\n\n\n
  1. ISIC 2018 – Link<\/a><\/li>
  2. BUSI – Link<\/a><\/li><\/ol>\n\n\n\n

    MICCAI 2022\uff1a\u57fa\u4e8e MLP \u7684\u5feb\u901f\u533b\u5b66\u56fe\u50cf\u5206\u5272\u7f51\u7edc UNeXt<\/h3>\n\n\n\n

    \u524d\u8a00<\/h2>\n\n\n\n

    \u6700\u8fd1 MICCAI 2022 \u7684\u8bba\u6587\u96c6\u5f00\u653e\u4e0b\u8f7d\u4e86\uff0c\u5730\u5740\uff1ahttps:\/\/link.springer.com\/book\/10.1007\/978-3-031-16443-9 \uff0c\u6bcf\u4e2a\u90e8\u5206\u7684\u5185\u5bb9\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n

    Part I: Brain development and atlases; DWI and tractography; functional brain networks; neuroimaging; heart and lung imaging; dermatology;

    Part II: Computational (integrative) pathology; computational anatomy and physiology; ophthalmology; fetal imaging;

    Part III: Breast imaging; colonoscopy; computer aided diagnosis;

    Part IV: Microscopic image analysis; positron emission tomography; ultrasound imaging; video data analysis; image segmentation I;

    Part V: Image segmentation II; integration of imaging with non-imaging biomarkers;

    Part VI: Image registration; image reconstruction;

    Part VII: Image-Guided interventions and surgery; outcome and disease prediction; surgical data science; surgical planning and simulation; machine learning \u2013 domain adaptation and generalization;

    Part VIII: Machine learning \u2013 weakly-supervised learning; machine learning \u2013 model interpretation; machine learning \u2013 uncertainty; machine learning theory and methodologies.<\/code><\/pre>\n\n\n\n
    \u5176\u4e2d\u5173\u4e8e\u5206\u5272\u6709\u4e24\u4e2a\u90e8\u5206\uff0cImage segmentation I \u5728 Part IV, \u800c Image segmentation II \u5728 Part V\u3002<\/p>\n\n\n\n
    \u968f\u7740\u533b\u5b66\u56fe\u50cf\u7684\u89e3\u51b3\u65b9\u6848\u53d8\u5f97\u8d8a\u6765\u8d8a\u9002\u7528\uff0c\u6211\u4eec\u66f4\u9700\u8981\u5173\u6ce8\u4f7f\u6df1\u5ea6\u7f51\u7edc\u8f7b\u91cf\u7ea7\u3001\u5feb\u901f\u4e14\u9ad8\u6548\u7684\u65b9\u6cd5\u3002\u5177\u6709\u9ad8\u63a8\u7406\u901f\u5ea6\u7684\u8f7b\u91cf\u7ea7\u7f51\u7edc\u53ef\u4ee5\u88ab\u90e8\u7f72\u5728\u624b\u673a\u7b49\u8bbe\u5907\u4e0a\uff0c\u4f8b\u5982 POCUS\uff08point-of-care ultrasound\uff09\u88ab\u7528\u4e8e\u68c0\u6d4b\u548c\u8bca\u65ad\u76ae\u80a4\u72b6\u51b5\u3002\u8fd9\u5c31\u662f UNeXt \u7684\u52a8\u673a\u3002<\/p>\n\n\n\n
    \u65b9\u6cd5\u6982\u8ff0<\/h2>\n\n\n\n
        \u4e4b\u524d\u6211\u4eec\u89e3\u8bfb\u8fc7\u57fa\u4e8e Transformer \u7684 U-Net \u53d8\u4f53\uff0c\u8fd1\u5e74\u6765\u4e00\u76f4\u662f\u9886\u5148\u7684\u533b\u5b66\u56fe\u50cf\u5206\u5272\u65b9\u6cd5\uff0c\u4f46\u662f\u53c2\u6570\u91cf\u5f80\u5f80\u4e0d\u4e50\u89c2\uff0c\u8ba1\u7b97\u590d\u6742\uff0c\u63a8\u7406\u7f13\u6162\u3002\u8fd9\u7bc7\u6587\u7ae0\u63d0\u51fa\u4e86\u57fa\u4e8e\u5377\u79ef\u591a\u5c42\u611f\u77e5\u5668\uff08MLP\uff09\u6539\u8fdb U \u578b\u67b6\u6784\u7684\u65b9\u6cd5\uff0c\u53ef\u4ee5\u7528\u4e8e\u56fe\u50cf\u5206\u5272\u3002\u8bbe\u8ba1\u4e86\u4e00\u4e2a tokenized MLP \u5757\u6709\u6548\u5730\u6807\u8bb0\u548c\u6295\u5f71\u5377\u79ef\u7279\u5f81\uff0c\u4f7f\u7528 MLPs \u6765\u5efa\u6a21\u8868\u793a\u3002\u8fd9\u4e2a\u7ed3\u6784\u88ab\u5e94\u7528\u5230 U \u578b\u67b6\u6784\u7684\u4e0b\u4e24\u5c42\u4e2d\uff08\u8fd9\u91cc\u6211\u4eec\u5047\u8bbe\u7eb5\u5411\u4e00\u5171\u4e94\u5c42\uff09\u3002\u6587\u7ae0\u4e2d\u63d0\u5230\uff0c\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6027\u80fd\uff0c\u5efa\u8bae\u5728\u8f93\u5165\u5230 MLP \u7684\u8fc7\u7a0b\u4e2d\u6539\u53d8\u8f93\u5165\u7684\u901a\u9053\uff0c\u4ee5\u4fbf\u4e13\u6ce8\u4e8e\u5b66\u4e60\u5c40\u90e8\u4f9d\u8d56\u5173\u7cfb\u7279\u5f81\u3002\u8fd8\u6709\u989d\u5916\u7684\u8bbe\u8ba1\u5c31\u662f\u8df3\u8dc3\u8fde\u63a5\u4e86\uff0c\u5e76\u4e0d\u662f\u6211\u4eec\u4e3b\u8981\u5173\u6ce8\u7684\u5730\u65b9\u3002\u6700\u7ec8\uff0cUNeXt \u5c06\u53c2\u6570\u6570\u91cf\u51cf\u5c11\u4e86 72 \u500d\uff0c\u8ba1\u7b97\u590d\u6742\u5ea6\u964d\u4f4e\u4e86 68 \u500d\uff0c\u63a8\u7406\u901f\u5ea6\u63d0\u9ad8\u4e86 10 \u500d\uff0c\u540c\u65f6\u8fd8\u83b7\u5f97\u4e86\u66f4\u597d\u7684\u5206\u5272\u6027\u80fd\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    UNeXt \u67b6\u6784<\/h2>\n\n\n\n
    UNeXt \u7684\u8bbe\u8ba1\u5982\u4e0b\u56fe\u6240\u793a\u3002\u7eb5\u5411\u6765\u770b\uff0c\u4e00\u5171\u6709\u4e24\u4e2a\u9636\u6bb5\uff0c\u666e\u901a\u7684\u5377\u79ef\u548c Tokenized MLP \u9636\u6bb5\u3002\u5176\u4e2d\uff0c\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u5206\u522b\u8bbe\u8ba1\u4e24\u4e2a Tokenized MLP \u5757\u3002\u6bcf\u4e2a\u7f16\u7801\u5668\u5c06\u5206\u8fa8\u7387\u964d\u4f4e\u4e24\u500d\uff0c\u89e3\u7801\u5668\u5de5\u4f5c\u76f8\u53cd\uff0c\u8fd8\u6709\u8df3\u8dc3\u8fde\u63a5\u7ed3\u6784\u3002\u6bcf\u4e2a\u5757\u7684\u901a\u9053\u6570\uff08C1-C5\uff09\u88ab\u8bbe\u8ba1\u6210\u8d85\u53c2\u6570\u4e3a\u4e86\u627e\u5230\u4e0d\u6389\u70b9\u60c5\u51b5\u4e0b\u6700\u5c0f\u53c2\u6570\u91cf\u7684\u7f51\u7edc\uff0c\u5bf9\u4e8e\u4f7f\u7528 UNeXt \u67b6\u6784\u7684\u5b9e\u9a8c\uff0c\u9075\u5faa C1 = 32\u3001C2 = 64\u3001C3 = 128\u3001C4 = 160 \u548c C5 = 256\u3002<\/p>\n\n\n\n
    TokMLP \u8bbe\u8ba1\u601d\u8def<\/h2>\n\n\n\n
        \u5173\u4e8e Convolutional Stage \u6211\u4eec\u4e0d\u505a\u8fc7\u591a\u4ecb\u7ecd\u4e86\uff0c\u5728\u8fd9\u4e00\u90e8\u5206\u91cd\u70b9\u4e13\u6ce8 Tokenized MLP Stage\u3002\u4ece\u4e0a\u4e00\u90e8\u5206\u7684\u56fe\u4e2d\uff0c\u53ef\u4ee5\u770b\u5230 Shifted MLP \u8fd9\u4e00\u64cd\u4f5c\uff0c\u5176\u5b9e\u601d\u8def\u7c7b\u4f3c\u4e8e Swin transformer\uff0c\u5f15\u5165\u57fa\u4e8e\u7a97\u53e3\u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5411\u5168\u5c40\u6a21\u578b\u4e2d\u6dfb\u52a0\u66f4\u591a\u7684\u5c40\u57df\u6027<\/strong>\u3002\u4e0b\u56fe\u7684\u610f\u601d\u662f\uff0cTokenized MLP \u5757\u6709 2 \u4e2a MLP\uff0c\u5728\u4e00\u4e2a MLP \u4e2d\u8de8\u8d8a\u5bbd\u5ea6\u79fb\u52a8\u7279\u5f81\uff0c\u5728\u53e6\u4e00\u4e2a MLP \u4e2d\u8de8\u8d8a\u9ad8\u5ea6\u79fb\u52a8\u7279\u5f81\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u7279\u5f81\u5728\u9ad8\u5ea6\u548c\u5bbd\u5ea6\u4e0a\u4f9d\u6b21\u79fb\u4f4d\u3002\u8bba\u6587\u4e2d\u662f\u8fd9\u4e48\u8bf4\u7684\uff1a\u201c\u6211\u4eec\u5c06\u7279\u5f81\u5206\u6210 h \u4e2a\u4e0d\u540c\u7684\u5206\u533a\uff0c\u5e76\u6839\u636e\u6307\u5b9a\u7684\u8f74\u7ebf\u5c06\u5b83\u4eec\u79fb\u5230 j=5 \u7684\u4f4d\u7f6e\u201d\u3002\u5176\u5b9e\u5c31\u662f\u521b\u5efa\u4e86\u968f\u673a\u7a97\u53e3\uff0c\u8fd9\u4e2a\u56fe\u53ef\u4ee5\u7406\u89e3\u4e3a\u7070\u8272\u662f\u7279\u5f81\u5757\u7684\u4f4d\u7f6e\uff0c\u767d\u8272\u662f\u79fb\u52a8\u4e4b\u540e\u7684 padding\u3002<\/p>\n\n\n\n
    \uff08\u8865\u5145\uff1aMLP\u62e5\u6709\u5927\u91cf\u53c2\u6570\uff0c\u8ba1\u7b97\u6210\u672c\u9ad8\u4e14\u5bb9\u6613\u8fc7\u5ea6\u62df\u5408\uff0c\u800c\u4e14\u56e0\u4e3a\u5c42\u4e4b\u95f4\u7684\u7ebf\u6027\u53d8\u6362\u603b\u662f\u5c06\u524d\u4e00\u5c42\u7684\u8f93\u51fa\u4f5c\u4e3a\u4e00\u4e2a\u6574\u4f53\uff0c\u6240\u4ee5MLP\u5728\u6355\u83b7\u8f93\u5165\u7279\u5f81\u56fe\u4e2d\u7684\u5c40\u90e8\u7279\u5f81\u7ed3\u6784\u7684\u80fd\u529b\u8f83\u5f31\u3002\u901a\u8fc7\u8f74\u5411\u79fb\u52a8\u7279\u5f81\u4fe1\u606f\uff0c  Shifted MLP\u53ef\u4ee5\u5f97\u5230\u4e0d\u540c\u65b9\u5411\u7684\u4fe1\u606f\u6d41\uff0c\u8fd9\u6709\u52a9\u4e8e\u6355\u83b7\u5c40\u90e8\u76f8\u5173\u6027\u3002\u8be5\u64cd\u4f5c\u4f7f\u5f97\u6211\u4eec\u91c7\u7528\u7eafMLP\u67b6\u6784\u5373\u53ef\u53d6\u5f97\u4e0eCNN\u76f8\u540c\u7684\u611f\u53d7\u91ce\u3002<\/strong>\uff09<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
       \u89e3\u91ca\u8fc7 Shifted MLP \u540e\uff0c\u6211\u4eec\u518d\u770b\u53e6\u4e00\u90e8\u5206\uff1atokenized MLP block\u3002\u9996\u5148\uff0c\u9700\u8981\u628a\u7279\u5f81\u8f6c\u6362\u4e3a tokens\uff08\u53ef\u4ee5\u7406\u89e3\u4e3a Patch Embedding \u7684\u8fc7\u7a0b\uff0c\u611f\u89c9\u8fd9\u4e2a\u5c31\u662f\u4e2a\u666e\u901a\u5377\u79ef\uff0c\u800c\u4e14\u4f5c\u8005\u4e3a\u4e86\u4fdd\u8bc1conv\u540e\u7684\u77e9\u9635\u51cf\u534a\uff0c\u8bbe\u7f6e\u6b65\u5e45\u4e3a2\uff0c\u603b\u4e4b\uff0c\u6709\u4e9b\u7f16\u6545\u4e8b\u7684\u610f\u601d\u4e86\uff09\u3002\u4e3a\u4e86\u5b9e\u73b0 tokenized \u5316\uff0c\u4f7f\u7528 kernel size \u4e3a 3 \u7684\u5377\u79ef\uff08patch_size=3, stride=2\uff09\uff0c\u8fd9\u6837\u4f1a\u4f7f\u5f97\u77e9\u9635H\u548cW\u51cf\u534a\uff0c\u5e76\u5c06\u901a\u9053\u7684\u6570\u91cf\u6539\u4e3a E\uff0cE \u662f embadding \u5d4c\u5165\u7ef4\u5ea6\uff08 token \u7684\u6570\u91cf\uff09\uff0c\u4e5f\u662f\u4e00\u4e2a\u8d85\u53c2\u6570\u3002\u7136\u540e\u628a\u8fd9\u4e9b token \u9001\u5230\u4e0a\u9762\u63d0\u5230\u7684\u7b2c\u4e00\u4e2a\u8de8\u8d8a\u5bbd\u5ea6\u7684 MLP \u4e2d\u3002<\/p>\n\n\n\n
    \u8fd9\u91cc\u4f1a\u4ea7\u751f\u4e86\u4e00\u4e2a\u7591\u95ee\uff0c\u5173\u4e8e kernel size \u4e3a 3 \u7684\u5377\u79ef\uff0c\u4f7f\u7528\u7684\u662f\u4ec0\u4e48\u6837\u7684\u5377\u79ef\u5c42\uff1f\u7b54\uff1a\u8fd9\u91cc\u8fd8\u662f\u666e\u901a\u7684\u5377\u79ef\uff0c\u6587\u7ae0\u4e2d\u63d0\u5230\u4e86 DWConv\uff08DepthWise Conv\uff09\uff0c\u662f\u540e\u9762\u7684\u7279\u5f81\u901a\u8fc7 DW-Conv \u4f20\u9012\u3002\u4f7f\u7528 DWConv \u6709\u4e24\u4e2a\u539f\u56e0\uff1a\uff081\uff09\u5b83\u6709\u52a9\u4e8e\u5bf9 MLP \u7279\u5f81\u7684\u4f4d\u7f6e\u4fe1\u606f\u8fdb\u884c\u7f16\u7801\u3002MLP \u5757\u4e2d\u7684\u5377\u79ef\u5c42\u8db3\u4ee5\u7f16\u7801\u4f4d\u7f6e\u4fe1\u606f\uff0c\u5b83\u5b9e\u9645\u4e0a\u6bd4\u6807\u51c6\u7684\u4f4d\u7f6e\u7f16\u7801\u8868\u73b0\u5f97\u66f4\u597d\u3002\u50cf ViT \u4e2d\u7684\u4f4d\u7f6e\u7f16\u7801\u6280\u672f\uff0c\u5f53\u6d4b\u8bd5\u548c\u8bad\u7ec3\u7684\u5206\u8fa8\u7387\u4e0d\u4e00\u6837\u65f6\uff0c\u9700\u8981\u8fdb\u884c\u63d2\u503c\uff0c\u5f80\u5f80\u4f1a\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d\u3002\uff082\uff09DWConv \u4f7f\u7528\u7684\u53c2\u6570\u6570\u91cf\u8f83\u5c11\u3002<\/p>\n\n\n\n
    \u8fd9\u65f6\u6211\u4eec\u5f97\u5230\u4e86 DW-Conv \u4f20\u9012\u8fc7\u6765\u7684\u7279\u5f81\uff0c\u7136\u540e\u4f7f\u7528 GELU \u5b8c\u6210\u6fc0\u6d3b\u3002\u63a5\u4e0b\u6765\uff0c\u901a\u8fc7\u53e6\u4e00\u4e2a MLP\uff08\u8de8\u8d8aheight\uff09\u4f20\u9012\u7279\u5f81\uff0c\u8be5 MLP \u628a\u8fdb\u4e00\u6b65\u6539\u53d8\u4e86\u7279\u5f81\u5c3a\u5bf8\u3002\u5728\u8fd9\u91cc\u8fd8\u4f7f\u7528\u4e00\u4e2a\u6b8b\u5dee\u8fde\u63a5\uff0c\u5c06\u539f\u59cb token \u6dfb\u52a0\u4e3a\u6b8b\u5dee\u3002\u7136\u540e\u6211\u4eec\u5229\u7528 Layer Norm\uff08LN\uff09\uff0c\u5c06\u8f93\u51fa\u7279\u5f81\u4f20\u9012\u5230\u4e0b\u4e00\u4e2a\u5757\u3002LN \u6bd4 BN \u66f4\u53ef\u53d6\uff0c\u56e0\u4e3a\u5b83\u662f\u6cbf\u7740 token \u8fdb\u884c\u89c4\u8303\u5316\uff0c\u800c\u4e0d\u662f\u5728 Tokenized MLP \u5757\u7684\u6574\u4e2a\u6279\u5904\u7406\u4e2d\u8fdb\u884c\u89c4\u8303\u5316\u3002\u4e0a\u9762\u8fd9\u4e9b\u5c31\u662f\u4e00\u4e2a tokenized MLP block \u7684\u8bbe\u8ba1\u601d\u8def\u3002<\/p>\n\n\n\n
    \u6b64\u5916\uff0c\u6587\u7ae0\u4e2d\u7ed9\u51fa\u4e86 tokenized MLP block \u6d89\u53ca\u7684\u8ba1\u7b97\u516c\u5f0f\uff1a<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u5176\u4e2d T \u8868\u793a tokens\uff0cH \u8868\u793a\u9ad8\u5ea6\uff0cW \u8868\u793a\u5bbd\u5ea6\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6240\u6709\u8fd9\u4e9b\u8ba1\u7b97\u90fd\u662f\u5728 embedding \u7ef4\u5ea6 H \u4e0a\u8fdb\u884c\u7684\uff0c\u5b83\u660e\u663e\u5c0f\u4e8e\u7279\u5f81\u56fe\u7684\u7ef4\u5ea6 HN\u00d7HN<\/code>\uff0c\u5176\u4e2d N \u53d6\u51b3\u4e8e block \u5927\u5c0f\u3002\u5728\u4e0b\u9762\u7684\u5b9e\u9a8c\u90e8\u5206\uff0c\u6587\u7ae0\u5c06 H \u8bbe\u7f6e\u4e3a 768\u3002<\/p>\n\n\n\n
    \u5b9e\u9a8c\u90e8\u5206<\/h2>\n\n\n\n
    \u5b9e\u9a8c\u5728 ISIC \u548c BUSI \u6570\u636e\u96c6\u4e0a\u8fdb\u884c\uff0c\u53ef\u4ee5\u770b\u5230\uff0c\u5728 GLOPs\u3001\u6027\u80fd\u548c\u63a8\u7406\u65f6\u95f4\u90fd\u4e0a\u8868\u73b0\u4e0d\u9519\u3002<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u4e0b\u9762\u662f\u53ef\u89c6\u5316\u548c\u6d88\u878d\u5b9e\u9a8c\u7684\u90e8\u5206\u3002\u53ef\u89c6\u5316\u56fe\u53ef\u4ee5\u53d1\u73b0\uff0cUNeXt \u5904\u7406\u7684\u66f4\u52a0\u5706\u6ed1\u548c\u63a5\u8fd1\u771f\u5b9e\u6807\u7b7e\u3002<\/p>\n\n\n\n
    \u6d88\u878d\u5b9e\u9a8c\u53ef\u4ee5\u53d1\u73b0\uff0c\u4ece\u539f\u59cb\u7684 UNet \u5f00\u59cb\uff0c\u7136\u540e\u53ea\u662f\u51cf\u5c11\u8fc7\u6ee4\u5668\u7684\u6570\u91cf\uff0c\u53d1\u73b0\u6027\u80fd\u4e0b\u964d\uff0c\u4f46\u53c2\u6570\u5e76\u6ca1\u6709\u51cf\u5c11\u592a\u591a\u3002\u63a5\u4e0b\u6765\uff0c\u4ec5\u4f7f\u7528 3 \u5c42\u6df1\u5ea6\u67b6\u6784\uff0c\u65e2 UNeXt \u7684 Conv \u9636\u6bb5\u3002\u663e\u7740\u51cf\u5c11\u4e86\u53c2\u6570\u7684\u6570\u91cf\u548c\u590d\u6742\u6027\uff0c\u4f46\u6027\u80fd\u964d\u4f4e\u4e86 4%\u3002\u52a0\u5165 tokenized MLP block \u540e\uff0c\u5b83\u663e\u7740\u63d0\u9ad8\u4e86\u6027\u80fd\uff0c\u540c\u65f6\u5c06\u590d\u6742\u5ea6\u548c\u53c2\u6570\u91cf\u662f\u4e00\u4e2a\u6700\u5c0f\u503c\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06 DWConv \u6dfb\u52a0\u5230 positional embedding\uff0c\u6027\u80fd\u53c8\u63d0\u9ad8\u4e86\u3002\u63a5\u4e0b\u6765\uff0c\u5728 MLP \u4e2d\u6dfb\u52a0  Shifted \u64cd\u4f5c\uff0c\u8868\u660e\u5728\u6807\u8bb0\u5316\u4e4b\u524d\u79fb\u4f4d\u7279\u5f81\u53ef\u4ee5\u63d0\u9ad8\u6027\u80fd\uff0c\u4f46\u662f\u4e0d\u4f1a\u589e\u52a0\u4efb\u4f55\u53c2\u6570\u6216\u590d\u6742\u6027\u3002\u6ce8\u610f\uff1aShifted MLP \u4e0d\u4f1a\u589e\u52a0 GLOPs\u3002<\/p>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \u4e00\u4e9b\u7406\u89e3\u548c\u603b\u7ed3<\/h2>\n\n\n\n
    \u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u6df1\u5ea6\u7f51\u7edc\u67b6\u6784 UNeXt\uff0c\u7528\u4e8e\u533b\u7597\u56fe\u50cf\u5206\u5272\uff0c\u4e13\u6ce8\u4e8e\u53c2\u6570\u91cf\u7684\u51cf\u5c0f\u3002UNeXt \u662f\u4e00\u79cd\u57fa\u4e8e\u5377\u79ef\u548c MLP \u7684\u67b6\u6784\uff0c\u5176\u4e2d\u6709\u4e00\u4e2a\u521d\u59cb\u7684 Conv \u9636\u6bb5\uff0c\u7136\u540e\u662f\u6df1\u5c42\u7a7a\u95f4\u4e2d\u7684 MLP\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u63d0\u51fa\u4e86\u4e00\u4e2a\u5e26\u6709\u79fb\u4f4d MLP \u7684\u6807\u8bb0\u5316 MLP \u5757\u3002\u5728\u591a\u4e2a\u6570\u636e\u96c6\u4e0a\u9a8c\u8bc1\u4e86 UNeXt\uff0c\u5b9e\u73b0\u4e86\u66f4\u5feb\u7684\u63a8\u7406\u3001\u66f4\u4f4e\u7684\u590d\u6742\u6027\u548c\u66f4\u5c11\u7684\u53c2\u6570\u6570\u91cf\uff0c\u540c\u65f6\u8fd8\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n
    \u53e6\u5916\uff0c\u4e2a\u4eba\u89c9\u5f97 \u5e26\u6709\u79fb\u4f4d MLP \u7684\u6807\u8bb0\u5316 MLP \u5757\u8fd9\u91cc\u5176\u5b9e\u6709\u70b9\u8bb2\u6545\u4e8b\u7684\u610f\u601d\u4e86\u3002<\/p>\n\n\n\n
    \u6211\u5728\u8bfb\u8fd9\u7bc7\u8bba\u6587\u7684\u65f6\u5019\uff0c\u76f4\u63a5\u6ce8\u610f\u5230\u4e86\u5b83\u7528\u7684\u6570\u636e\u96c6\u3002\u6211\u8ba4\u4e3a UNeXt \u53ef\u80fd\u53ea\u9002\u7528\u4e8e\u8fd9\u79cd\u7b80\u5355\u7684\u533b\u5b66\u56fe\u50cf\u5206\u5272\u4efb\u52a1\uff0c\u7c7b\u4f3c\u7684\u6709 Optic Disc and Cup Seg\uff0c\u5bf9\u4e8e\u66f4\u590d\u6742\u7684\uff0c\u6bd4\u5982\u8840\u7ba1\uff0c\u8f6f\u9aa8\uff0cLiver Tumor\uff0ckidney Seg \u8fd9\u4e9b\uff0c\u53ef\u80fd\u6548\u679c\u8fbe\u4e0d\u5230\u8fd9\u4e48\u597d\uff0c\u56e0\u4e3a\u8fd0\u7b97\u91cf\u88ab\u6781\u5927\u7684\u51cf\u5c11\u4e86\uff0c\u6bcf\u4e2a convolutional \u9636\u6bb5\u53ea\u6709\u4e00\u4e2a\u5377\u79ef\u5c42\u3002MLP \u9b54\u6539 U-Net \u4e5f\u7b97\u662f\u4e00\u4e2a\u5c1d\u8bd5\uff0c\u5728 Tokenized MLP block \u4e2d\u52a0\u5165 DWConv \u4e5f\u662f\u5f88\u5408\u7406\u7684\u8bbe\u8ba1\u3002<\/p>\n\n\n\n
    \u4ee3\u7801\u5b9e\u73b0\uff1a<\/h2>\n\n\n\n
    class shiftmlp(nn.Module):\n    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0., shift_size=5):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        self.dim = in_features\n        self.fc1 = nn.Linear(in_features, hidden_features)\n        self.dwconv = DWConv(hidden_features)\n        self.act = act_layer()\n        self.fc2 = nn.Linear(hidden_features, out_features)\n        self.drop = nn.Dropout(drop)\n\n        self.shift_size = shift_size\n        self.pad = shift_size \/\/ 2\n\n        \n        self.apply(self._init_weights)\n\n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            trunc_normal_(m.weight, std=.02)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n        elif isinstance(m, nn.Conv2d):\n            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels\n            fan_out \/\/= m.groups\n            m.weight.data.normal_(0, math.sqrt(2.0 \/ fan_out))\n            if m.bias is not None:\n                m.bias.data.zero_()\n    \n\n\n    def forward(self, x, H, W):\n        # pdb.set_trace()\n        B, N, C = x.shape\n\n        xn = x.transpose(1, 2).view(B, C, H, W).contiguous()\n        #pad\uff0c\u65b9\u4fbf\u540e\u9762\u7684torch.chunk\n        xn = F.pad(xn, (self.pad, self.pad, self.pad, self.pad) , \"constant\", 0)\n        #\u6309\u7167dim=1\u7ef4\u5ea6\uff0c\u5206\u6210 self.shift_size(5)\u4e2a\u5757\n        xs = torch.chunk(xn, self.shift_size, 1)\n        #torch.roll(x,y,d)\u5c06x\uff0c\u6cbf\u7740d\u7ef4\u5ea6\uff0c\u5411\u4e0a\/\u4e0broll y\u4e2a\u503c\n        x_shift = [torch.roll(x_c, shift, 2) for x_c, shift in zip(xs, range(-self.pad, self.pad+1))]\n        x_cat = torch.cat(x_shift, 1)\n        #x.narrow(*dimension*, *start*, *length*) \u2192 Tensor \u8868\u793a\u53d6\u53d8\u91cfx\u7684\u7b2cdimension\u7ef4,\u4ece\u7d22\u5f15start\u5f00\u59cb\u5230(start+length-1)\u8303\u56f4\u7684\u503c\u3002\n        x_cat = torch.narrow(x_cat, 2, self.pad, H)\n        x_s = torch.narrow(x_cat, 3, self.pad, W)\n\n        x_s = x_s.reshape(B,C,H*W).contiguous()\n        x_shift_r = x_s.transpose(1,2)\n\n        x = self.fc1(x_shift_r)\n\n        x = self.dwconv(x, H, W)\n        x = self.act(x) \n        x = self.drop(x)\n\n        xn = x.transpose(1, 2).view(B, C, H, W).contiguous()\n        xn = F.pad(xn, (self.pad, self.pad, self.pad, self.pad) , \"constant\", 0)\n        xs = torch.chunk(xn, self.shift_size, 1)\n        x_shift = [torch.roll(x_c, shift, 3) for x_c, shift in zip(xs, range(-self.pad, self.pad+1))]\n        x_cat = torch.cat(x_shift, 1)\n        x_cat = torch.narrow(x_cat, 2, self.pad, H)\n        x_s = torch.narrow(x_cat, 3, self.pad, W)\n        x_s = x_s.reshape(B,C,H*W).contiguous()\n        x_shift_c = x_s.transpose(1,2)\n\n        x = self.fc2(x_shift_c)\n        x = self.drop(x)\n        return x<\/code><\/pre>\n\n\n\n
    <\/p>\n\n\n\n
    class shiftedBlock(nn.Module):\n    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,\n                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):\n        super().__init__()\n\n\n        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = shiftmlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n        self.apply(self._init_weights)\n\n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            trunc_normal_(m.weight, std=.02)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n        elif isinstance(m, nn.Conv2d):\n            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels\n            fan_out \/\/= m.groups\n            m.weight.data.normal_(0, math.sqrt(2.0 \/ fan_out))\n            if m.bias is not None:\n                m.bias.data.zero_()\n\n    def forward(self, x, H, W):\n\n        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))\n        return x<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"
    \u8bba\u6587\u5730\u5740: https:\/\/arxiv.org\/abs\/2203.04967 github:https:\/\/g … \u7ee7\u7eed\u9605\u8bfbMICCAI 2022\uff1a\u57fa\u4e8e MLP \u7684\u5feb\u901f\u533b\u5b66\u56fe\u50cf\u5206\u5272\u7f51\u7edc UNeXt<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[24,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/8319"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=8319"}],"version-history":[{"count":44,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/8319\/revisions"}],"predecessor-version":[{"id":8905,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/8319\/revisions\/8905"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=8319"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=8319"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=8319"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}