{"id":4792,"date":"2022-07-13T20:49:52","date_gmt":"2022-07-13T12:49:52","guid":{"rendered":"http:\/\/139.9.1.231\/?p=4792"},"modified":"2022-11-10T10:47:48","modified_gmt":"2022-11-10T02:47:48","slug":"repvgg-making-vgg-style-convnets-great-again","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/07\/13\/repvgg-making-vgg-style-convnets-great-again\/","title":{"rendered":"RepVGG: Making VGG-style ConvNets Great Again"},"content":{"rendered":"\n

\u8bba\u6587\u4e0b\u8f7d\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2101.03697<\/a>
\u5b98\u65b9\u6e90\u7801\uff08Pytorch\u5b9e\u73b0\uff09\uff1ahttps:\/\/github.com\/DingXiaoH\/RepVGG<\/a><\/p>\n\n\n\n

\u8fd9\u7bc7\u8bba\u6587\u5bf9\u4e8e\u6211\u6765\u8bf4\u6700\u5927\u7684\u7528\u5904\u662f\u63d0\u51fa\u4e86\u7ed3\u6784\u7684\u91cd\u91cd\u53c2\u6570\u5316\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u5728\u63a8\u7406\u65f6\u5c06\u4e09\u4e2a\u5e76\u884c\u5206\u652f\u5408\u5e76\u6210\u5355\u4e2a\u5206\u652f\uff0c\u5e76\u4fdd\u8bc1\u8f93\u51fa\u8f93\u51fa\u4e0d\u53d8\u3002<\/p>\n\n\n\n

\u7ed3\u6784\u91cd\u53c2\u6570\u5316\u4e3b\u8981\u5206\u4e3a\u4e24\u6b65\uff0c\u7b2c\u4e00\u6b65\u4e3b\u8981\u662f\u5c06Conv2d\u7b97\u5b50\u548cBN\u7b97\u5b50\u878d\u5408\u4ee5\u53ca\u5c06\u53ea\u6709BN\u7684\u5206\u652f\u8f6c\u6362\u6210\u4e00\u4e2aConv2d\u7b97\u5b50\uff0c\u7b2c\u4e8c\u6b65\u5c06\u6bcf\u4e2a\u5206\u652f\u4e0a\u76843x3<\/code>\u5377\u79ef\u5c42\u878d\u5408\u6210\u4e00\u4e2a\u5377\u79ef\u5c42\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

1\u3001Conv2d\u548cBN \u8fd9\u4e2a\u5df2\u7ecf\u662f\u975e\u5e38\u5e38\u89c1\u7684\uff0c\u56e0\u4e3a\u5377\u79ef\u6838bn\u90fd\u662f\u7ebf\u6027\u8fd0\u7b97\uff0c\u6240\u4ee5\u53ef\u4ee5\u8fdb\u884c\u5408\u5e76\u3002<\/p>\n\n\n\n

\u8fd9\u91cc\u5047\u8bbe\u8f93\u5165\u7684\u7279\u5f81\u56fe\uff08Input feature map\uff09\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u8f93\u5165\u901a\u9053\u6570\u4e3a2\uff0c\u7136\u540e\u91c7\u7528\u4e24\u4e2a\u5377\u79ef\u6838\uff08\u56fe\u4e2d\u53ea\u753b\u4e86\u7b2c\u4e00\u4e2a\u5377\u79ef\u6838\u5bf9\u5e94\u53c2\u6570\uff09\u3002<\/p>\n\n\n\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"\/<\/figure>\n\n\n\n

\u63a5\u7740\u8ba1\u7b97\u4e00\u4e0b\u8f93\u51fa\u7279\u5f81\u56fe\uff08Output feature map\uff09\u901a\u90531\u4e0a\u7684\u7b2c\u4e00\u4e2a\u5143\u7d20\uff0c\u5373\u5f53\u5377\u79ef\u68381\u5728\u8f93\u5165\u7279\u5f81\u56fe\u7ea2\u8272\u6846\u533a\u57df\u5377\u79ef\u65f6\u5f97\u5230\u7684\u503c\uff08\u4e3a\u4e86\u4fdd\u8bc1\u8f93\u5165\u8f93\u51fa\u7279\u5f81\u56fe\u9ad8\u5bbd\u4e0d\u53d8\uff0c\u6240\u4ee5\u5bf9Input feature map\u8fdb\u884c\u4e86Padding\uff09\u3002\u5176\u4ed6\u4f4d\u7f6e\u7684\u8ba1\u7b97\u8fc7\u7a0b\u7c7b\u4f3c\u8fd9\u91cc\u5c31\u4e0d\u53bb\u6f14\u793a\u4e86\u3002<\/p>\n\n\n\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"\/<\/figure>\n\n\n\n

\u7136\u540e\u518d\u5c06\u5377\u79ef\u5c42\u8f93\u51fa\u7684\u7279\u5f81\u56fe\u4f5c\u4e3aBN\u5c42\u7684\u8f93\u5165\uff0c\u8fd9\u91cc\u540c\u6837\u8ba1\u7b97\u4e00\u4e0b\u8f93\u51fa\u7279\u5f81\u56fe\uff08Output feature map\uff09\u901a\u90531\u4e0a\u7684\u7b2c\u4e00\u4e2a\u5143\u7d20\uff0c\u6309\u7167\u4e0a\u8ff0BN\u5728\u63a8\u7406\u65f6\u7684\u8ba1\u7b97\u516c\u5f0f\u5373\u53ef\u5f97\u5230\u5982\u4e0b\u56fe\u6240\u793a\u7684\u8ba1\u7b97\u7ed3\u679c\u3002<\/p>\n\n\n\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"\/<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u4ee3\u7801<\/p>\n\n\n\n

Conv2d+BN\u878d\u5408\u5b9e\u9a8c(Pytorch)
\u4e0b\u9762\u662f\u53c2\u8003\u4f5c\u8005\u63d0\u4f9b\u7684\u6e90\u7801\u6539\u7684\u4e00\u4e2a\u5c0f\u5b9e\u9a8c\uff0c\u9996\u5148\u521b\u5efa\u4e86\u4e00\u4e2amodule\u5305\u542b\u4e86\u5377\u79ef\u548cBN\u6a21\u5757\uff0c\u7136\u540e\u6309\u7167\u4e0a\u8ff0\u8f6c\u6362\u516c\u5f0f\u5c06\u5377\u79ef\u5c42\u7684\u6743\u91cd\u548cBN\u7684\u6743\u91cd\u8fdb\u884c\u878d\u5408\u8f6c\u6362\uff0c\u63a5\u7740\u8f7d\u5165\u5230\u65b0\u5efa\u7684\u5377\u79ef\u6a21\u5757fused_conv\u4e2d\uff0c\u6700\u540e\u968f\u673a\u521b\u5efa\u4e00\u4e2aTensor\uff08f1\uff09\u5c06\u5b83\u5206\u522b\u8f93\u5165\u5230module\u4ee5\u53cafused_conv\u4e2d\uff0c\u901a\u8fc7\u5bf9\u6bd4\u4e24\u8005\u7684\u8f93\u51fa\u53ef\u4ee5\u53d1\u73b0\u5b83\u4eec\u7684\u7ed3\u679c\u662f\u4e00\u81f4\u7684\u3002
<\/p>\n\n\n\n

from collections import OrderedDict\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\n\n\ndef main():\n    torch.random.manual_seed(0)\n\n    f1 = torch.randn(1, 2, 3, 3)\n\n    module = nn.Sequential(OrderedDict(\n        conv=nn.Conv2d(in_channels=2, out_channels=2, kernel_size=3, stride=1, padding=1, bias=False),\n        bn=nn.BatchNorm2d(num_features=2)\n    ))\n\n    module.eval()\n\n    with torch.no_grad():\n        output1 = module(f1)\n        print(output1)\n\n    # fuse conv + bn\n    kernel = module.conv.weight \n    running_mean = module.bn.running_mean\n    running_var = module.bn.running_var\n    gamma = module.bn.weight\n    beta = module.bn.bias\n    eps = module.bn.eps\n    std = (running_var + eps).sqrt()\n    t = (gamma \/ std).reshape(-1, 1, 1, 1)  # [ch] -> [ch, 1, 1, 1]\n    kernel = kernel * t\n    bias = beta - running_mean * gamma \/ std\n    fused_conv = nn.Conv2d(in_channels=2, out_channels=2, kernel_size=3, stride=1, padding=1, bias=True)\n    fused_conv.load_state_dict(OrderedDict(weight=kernel, bias=bias))\n\n    with torch.no_grad():\n        output2 = fused_conv(f1)\n        print(output2)\n\n    np.testing.assert_allclose(output1.numpy(), output2.numpy(), rtol=1e-03, atol=1e-05)\n    print(\"convert module has been tested, and the result looks good!\")\n\n\nif __name__ == '__main__':\n    main()\n<\/code><\/pre>\n\n\n\n
\u200b repVGG\u4e2d\u5927\u91cf\u8fd0\u7528conv+BN\u5c42\uff0c\u6211\u4eec\u77e5\u9053\u5c06\u5c42\u5408\u5e76\uff0c\u51cf\u5c11\u5c42\u6570\u80fd\u63d0\u5347\u7f51\u7edc\u6027\u80fd\uff0c\u4e0b\u9762\u7684\u63a8\u7406\u662fconv\u5e26\u6709bias\u7684\u8fc7\u7a0b\uff1a<\/strong><\/p>\n\n\n\n
\"\"\/<\/figure><\/div>\n\n\n\n
\u8fd9\u5176\u5b9e\u5c31\u662f\u4e00\u4e2a\u5377\u79ef\u5c42\uff0c\u53ea\u4e0d\u8fc7\u6743\u91cd\u8003\u8651\u4e86BN\u7684\u53c2\u6570 \u6211\u4eec\u4ee4\uff1a<\/p>\n\n\n\n
\"\"\/<\/figure><\/div>\n\n\n\n
\u6700\u7ec8\u7684\u878d\u5408\u7ed3\u679c\u5373\u4e3a\uff1a<\/p>\n\n\n\n
\"\"<\/figure><\/div>\n\n\n\n
\u76f8\u5173\u878d\u5408\u4ee3\u7801\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/strong><\/p>\n\n\n\n
def _fuse_bn_tensor(self, branch):\r\n        if branch is None:\r\n            return 0, 0\r\n        if isinstance(branch, nn.Sequential):\r\n            kernel = branch.conv.weight\r\n            running_mean = branch.bn.running_mean\r\n            running_var = branch.bn.running_var\r\n            gamma = branch.bn.weight\r\n            beta = branch.bn.bias\r\n            eps = branch.bn.eps\r\n        else:\r\n            ...\r\n        std = (running_var + eps).sqrt()\r\n        t = (gamma \/ std).reshape(-1, 1, 1, 1)\r\n        return kernel * t, beta - running_mean * gamma \/ std<\/code><\/pre>\n\n\n\n
2\u3001\u5982\u4f55\u5c06\u4e0d\u540c\u5206\u652f\u5408\u5e76\uff1a<\/p>\n\n\n\n
\u4f5c\u8005\u8fd9\u91cc\u9996\u5148\u5c06\u4e0d\u540c\u5206\u652f\u7684\u5377\u79ef\u6838\u90fd\u53d8\u62103*3\uff1a<\/p>\n\n\n\n
2.1 \u5c061×1\u5377\u79ef\u8f6c\u6362\u62103×3\u5377\u79ef
\u8fd9\u4e2a\u8fc7\u7a0b\u6bd4\u8f83\u7b80\u5355\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u4ee51×1\u5377\u79ef\u5c42\u4e2d\u67d0\u4e00\u4e2a\u5377\u79ef\u6838\u4e3a\u4f8b\uff0c\u53ea\u9700\u5728\u539f\u6765\u6743\u91cd\u5468\u56f4\u8865\u4e00\u5708\u96f6\u5c31\u884c\u4e86\uff0c\u8fd9\u6837\u5c31\u53d8\u6210\u4e863×3\u7684\u5377\u79ef\u5c42\uff0c\u6ce8\u610f\u4e3a\u4e86\u4fdd\u8bc1\u8f93\u5165\u8f93\u51fa\u7279\u5f81\u56fe\u9ad8\u5bbd\u4e0d\u53d8\uff0c\u6b64\u65f6\u9700\u8981\u5c06padding\u8bbe\u7f6e\u62101\uff08\u539f\u6765\u5377\u79ef\u6838\u5927\u5c0f\u4e3a1×1\u65f6padding\u4e3a0\uff09\u3002\u6700\u540e\u6309\u7167\u4e0a\u8ff02.1\u4e2d\u8bb2\u7684\u5185\u5bb9\u5c06\u5377\u79ef\u5c42\u548cBN\u5c42\u8fdb\u884c\u878d\u5408\u5373\u53ef\u3002<\/p>\n\n\n\n
\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"\/<\/figure>\n\n\n\n
2.2\u5c06BN\u8f6c\u6362\u62103×3\u5377\u79ef
\u5bf9\u4e8e\u53ea\u6709BN\u7684\u5206\u652f\u7531\u4e8e\u6ca1\u6709\u5377\u79ef\u5c42\uff0c\u6240\u4ee5\u6211\u4eec\u53ef\u4ee5\u5148\u81ea\u5df1\u6784\u5efa\u51fa\u4e00\u4e2a\u5377\u79ef\u5c42\u6765\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u6784\u5efa\u4e86\u4e00\u4e2a3×3\u7684\u5377\u79ef\u5c42\uff0c\u8be5\u5377\u79ef\u5c42\u53ea\u505a\u4e86\u6052\u7b49\u6620\u5c04\uff0c\u5373\u8f93\u5165\u8f93\u51fa\u7279\u5f81\u56fe\u4e0d\u53d8\u3002\u65e2\u7136\u6709\u4e86\u5377\u79ef\u5c42\uff0c\u90a3\u4e48\u53c8\u53ef\u4ee5\u6309\u7167\u4e0a\u8ff02.1\u4e2d\u8bb2\u7684\u5185\u5bb9\u5c06\u5377\u79ef\u5c42\u548cBN\u5c42\u8fdb\u884c\u878d\u5408\u3002<\/p>\n\n\n\n
\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"\/<\/figure>\n\n\n\n
2.3 \u591a\u5206\u652f\u878d\u5408
\u5728\u4e0a\u9762\u7684\u7ae0\u8282\u4e2d\uff0c\u6211\u4eec\u5df2\u7ecf\u8bb2\u4e86\u600e\u4e48\u628a\u6bcf\u4e2a\u5206\u652f\u878d\u5408\u8f6c\u6362\u6210\u4e00\u4e2a3×3\u7684\u5377\u79ef\u5c42\uff0c\u63a5\u4e0b\u6765\u9700\u8981\u8fdb\u4e00\u6b65\u5c06\u591a\u5206\u652f\u8f6c\u6362\u6210\u4e00\u4e2a\u5355\u8def3×3\u5377\u79ef\u5c42\u3002<\/p>\n\n\n\n
\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"\/<\/figure>\n\n\n\n
\u5408\u5e76\u7684\u8fc7\u7a0b\u5176\u5b9e\u4e5f\u5f88\u7b80\u5355\uff0c\u76f4\u63a5\u5c06\u8fd9\u4e09\u4e2a\u5377\u79ef\u5c42\u7684\u53c2\u6570\u76f8\u52a0\u5373\u53ef\uff0c\u5177\u4f53\u63a8\u7406\u8fc7\u7a0b\u5c31\u4e0d\u8bb2\u4e86\uff0c\u5982\u679c\u4e0d\u4e86\u89e3\u7684\u53ef\u4ee5\u81ea\u5df1\u52a8\u624b\u7b97\u7b97\u3002<\/p>\n\n\n\n
\u603b\u7684\u6765\u8bf4\uff0c\u8fd9\u7bc7\u8bba\u6587\u7684\u76ee\u6807\u662fSimple is Fast, Memory-economical, Flexible\uff0c\u63d0\u51fa\u4e86\u5f88\u591a\u60f3\u6cd5\u53bb\u5b9e\u73b0\u4e0a\u8ff0\u76ee\u6807\uff0c\u5bf9\u4e8e\u5f53\u524d\u6211\u7684\u5de5\u4f5c\u8fd8\u662f\u6bd4\u8f83\u6709\u542f\u53d1\u7684\uff0c\u5c24\u5176\u662f\u6700\u540e\u5bf9\u7f51\u7edc\u8fdb\u884c\u5408\u5e76\u4ee5\u53ca\u91cf\u5316\u90e8\u5206\u3002\u4e0b\u4e00\u6b65\u8981\u597d\u597d\u5b66\u4e60\u4e0btorch\u7684\u91cf\u5316QAT (torch.quantization.prepare_qat)<\/p>\n\n\n\n\n\n\n\n
RepVGG\u4e2d\u7684bn \u878d\u5408\u4ee5\u53ca\u6743\u91cd\u91cf\u5316\uff08\u4ee3\u7801\u5206\u6790\uff09<\/h2>\n\n\n\n
\u76ee\u524d\u6211\u5728\u505a\u76f8\u5173CNN\u7684\u91cf\u5316\u5de5\u4f5c\uff0c\u6240\u4ee5\u8fd9\u5757\u6211\u7279\u610f\u53bb\u4f5c\u8005\u4ee3\u7801\u4e2d\u53bb\u770b\u4e86\u8fd9\u90e8\u5206\uff0c\u505a\u4e0b\u8bb0\u5f55\u3002<\/p>\n\n\n\n
torch \uff1a https:\/\/pytorch.org\/docs\/stable\/quantization.html<\/a><\/p>\n\n\n\n
github \uff1a  https:\/\/github.com\/DingXiaoH\/RepVGG<\/a> <\/p>\n","protected":false},"excerpt":{"rendered":"
\u8bba\u6587\u4e0b\u8f7d\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2101.03697\u5b98\u65b9\u6e90\u7801\uff08Pytorch\u5b9e\u73b0\uff09\uff1a … \u7ee7\u7eed\u9605\u8bfbRepVGG: Making VGG-style ConvNets Great Again<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[13,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/4792"}],"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=4792"}],"version-history":[{"count":21,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/4792\/revisions"}],"predecessor-version":[{"id":10238,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/4792\/revisions\/10238"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=4792"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=4792"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=4792"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}