{"id":412,"date":"2022-08-21T16:18:36","date_gmt":"2022-08-21T08:18:36","guid":{"rendered":"http:\/\/www.gislxz.top\/?p=412"},"modified":"2022-08-21T16:38:44","modified_gmt":"2022-08-21T08:38:44","slug":"%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e7%ac%94%e8%ae%b0%ef%bc%8812%ef%bc%89","status":"publish","type":"post","link":"https:\/\/www.gislxz.com\/index.php\/2022\/08\/21\/%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e7%ac%94%e8%ae%b0%ef%bc%8812%ef%bc%89\/","title":{"rendered":"\u6df1\u5ea6\u5b66\u4e60\u7b14\u8bb0\uff0812\uff09"},"content":{"rendered":"\n<p>\u8fd9\u4e00\u7bc7\u662f<a href=\"https:\/\/aistudio.baidu.com\/aistudio\/projectdetail\/1613144\" target=\"_blank\"  rel=\"nofollow\" >\u5b98\u65b9\u6559\u7a0b\u56fe\u50cf\u5206\u7c7b<\/a>\u7684\u6700\u540e\u4e00\u90e8\u5206\uff0c\u4ecb\u7ecd\u4e86VGG\uff0cGoogLeNet\u4ee5\u53caResNet<\/p>\n\n\n\n<p>\u8fd9\u4e00\u7ae0\u8981\u8bb0\u5f97\u7b14\u8bb0\u5012\u662f\u4e0d\u591a\uff0c\u56e0\u4e3apaddle\u7684\u6559\u7a0b\u5df2\u7ecf\u8bb2\u7684\u5f88\u660e\u767d\u4e86<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">VGG<\/h2>\n\n\n\n<p>VGG\u7ed3\u6784\u76f8\u5bf9\u7b80\u5355\u6613\u61c2\uff0c\u6bcf\u4e00\u5c42\u7684\u5377\u79ef\u5e76\u4e0d\u4f1a\u6539\u53d8\u8f93\u51fa\u7684size\uff0c\u53ea\u662f\u589e\u52a0\u4e86channels\uff0c\u6c60\u5316\u5c42\u6bcf\u6b21\u5c06size\u51cf\u534a\uff0c\u6700\u540e\u5f97\u5230512\u00d77\u00d77\u7684output\uff0c\u4e4b\u540e\u7ecf\u8fc7\u4e09\u5c42\u7ebf\u6027\u5f97\u5230\u6700\u7ec8\u8f93\u51fa\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/ai-studio-static-online.cdn.bcebos.com\/657026651e084b639e011fe1fd4ba0bed502807ffa764fceb4796b9ee2a8736b\" alt=\"\"\/><\/figure>\n\n\n\n<pre class=\"wp-block-code\"><code># -*- coding:utf-8 -*-\n\n# VGG\u6a21\u578b\u4ee3\u7801\nimport numpy as np\nimport paddle\n# from paddle.nn import Conv2D, MaxPool2D, BatchNorm, Linear\nfrom paddle.nn import Conv2D, MaxPool2D, BatchNorm2D, Linear\n\n# \u5b9a\u4e49vgg\u7f51\u7edc\nclass VGG(paddle.nn.Layer):\n    def __init__(self):\n        super(VGG, self).__init__()\n\n        in_channels = &#91;3, 64, 128, 256, 512, 512]\n        # \u5b9a\u4e49\u7b2c\u4e00\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e24\u4e2a\u5377\u79ef\n        self.conv1_1 = Conv2D(in_channels=in_channels&#91;0], out_channels=in_channels&#91;1], kernel_size=3, padding=1, stride=1)\n        self.conv1_2 = Conv2D(in_channels=in_channels&#91;1], out_channels=in_channels&#91;1], kernel_size=3, padding=1, stride=1)\n        # \u5b9a\u4e49\u7b2c\u4e8c\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e24\u4e2a\u5377\u79ef\n        self.conv2_1 = Conv2D(in_channels=in_channels&#91;1], out_channels=in_channels&#91;2], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv2_2 = Conv2D(in_channels=in_channels&#91;2], out_channels=in_channels&#91;2], kernel_size=3, padding=1,\n                              stride=1)\n        # \u5b9a\u4e49\u7b2c\u4e09\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e09\u4e2a\u5377\u79ef\n        self.conv3_1 = Conv2D(in_channels=in_channels&#91;2], out_channels=in_channels&#91;3], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv3_2 = Conv2D(in_channels=in_channels&#91;3], out_channels=in_channels&#91;3], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv3_3 = Conv2D(in_channels=in_channels&#91;3], out_channels=in_channels&#91;3], kernel_size=3, padding=1,\n                              stride=1)\n        # \u5b9a\u4e49\u7b2c\u56db\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e09\u4e2a\u5377\u79ef\n        self.conv4_1 = Conv2D(in_channels=in_channels&#91;3], out_channels=in_channels&#91;4], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv4_2 = Conv2D(in_channels=in_channels&#91;4], out_channels=in_channels&#91;4], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv4_3 = Conv2D(in_channels=in_channels&#91;4], out_channels=in_channels&#91;4], kernel_size=3, padding=1,\n                              stride=1)\n        # \u5b9a\u4e49\u7b2c\u4e94\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e09\u4e2a\u5377\u79ef\n        self.conv5_1 = Conv2D(in_channels=in_channels&#91;4], out_channels=in_channels&#91;5], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv5_2 = Conv2D(in_channels=in_channels&#91;5], out_channels=in_channels&#91;5], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv5_3 = Conv2D(in_channels=in_channels&#91;5], out_channels=in_channels&#91;5], kernel_size=3, padding=1,\n                              stride=1)\n\n        # \u4f7f\u7528Sequential \u5c06\u5377\u79ef\u548crelu\u7ec4\u6210\u4e00\u4e2a\u7ebf\u6027\u7ed3\u6784\uff08fc + relu\uff09\n        # \u5f53\u8f93\u5165\u4e3a224x224\u65f6\uff0c\u7ecf\u8fc7\u4e94\u4e2a\u5377\u79ef\u5757\u548c\u6c60\u5316\u5c42\u540e\uff0c\u7279\u5f81\u7ef4\u5ea6\u53d8\u4e3a&#91;512x7x7]\n        self.fc1 = paddle.nn.Sequential(paddle.nn.Linear(512 * 7 * 7, 4096), paddle.nn.ReLU())\n        self.drop1_ratio = 0.5\n        self.dropout1 = paddle.nn.Dropout(self.drop1_ratio, mode='upscale_in_train')\n        # \u4f7f\u7528Sequential \u5c06\u5377\u79ef\u548crelu\u7ec4\u6210\u4e00\u4e2a\u7ebf\u6027\u7ed3\u6784\uff08fc + relu\uff09\n        self.fc2 = paddle.nn.Sequential(paddle.nn.Linear(4096, 4096), paddle.nn.ReLU())\n\n        self.drop2_ratio = 0.5\n        self.dropout2 = paddle.nn.Dropout(self.drop2_ratio, mode='upscale_in_train')\n        self.fc3 = paddle.nn.Linear(4096, 1)\n\n        self.relu = paddle.nn.ReLU()\n        self.pool = MaxPool2D(stride=2, kernel_size=2)\n\n    def forward(self, x):\n        x = self.relu(self.conv1_1(x))\n        x = self.relu(self.conv1_2(x))\n        x = self.pool(x)\n\n        x = self.relu(self.conv2_1(x))\n        x = self.relu(self.conv2_2(x))\n        x = self.pool(x)\n\n        x = self.relu(self.conv3_1(x))\n        x = self.relu(self.conv3_2(x))\n        x = self.relu(self.conv3_3(x))\n        x = self.pool(x)\n\n        x = self.relu(self.conv4_1(x))\n        x = self.relu(self.conv4_2(x))\n        x = self.relu(self.conv4_3(x))\n        x = self.pool(x)\n\n        x = self.relu(self.conv5_1(x))\n        x = self.relu(self.conv5_2(x))\n        x = self.relu(self.conv5_3(x))\n        x = self.pool(x)\n\n        x = paddle.flatten(x, 1, -1)\n        x = self.dropout1(self.relu(self.fc1(x)))\n        x = self.dropout2(self.relu(self.fc2(x)))\n        x = self.fc3(x)\n        return x<\/code><\/pre>\n\n\n\n<p>\u9ad8\u5185\u805a\u4f4e\u8026\u5408\uff0c\u5176\u4ed6\u90fd\u4e0d\u7528\u6539<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\n# \u521b\u5efa\u6a21\u578b\nmodel = VGG()\n# opt = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())\nopt = paddle.optimizer.Momentum(learning_rate=0.001, momentum=0.9, parameters=model.parameters())\n\n# \u542f\u52a8\u8bad\u7ec3\u8fc7\u7a0b\ntrain_pm(model, opt)<\/code><\/pre>\n\n\n\n<p>pytorch\u6784\u5efaVGG<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u5b9a\u4e49vgg\u7f51\u7edc\nclass VGG(torch.nn.Module):\n    def __init__(self):\n        super(VGG, self).__init__()\n\n        in_channels = &#91;3, 64, 128, 256, 512, 512]\n        # \u5b9a\u4e49\u7b2c\u4e00\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e24\u4e2a\u5377\u79ef\n        self.conv1_1 = Conv2d(in_channels=in_channels&#91;0], out_channels=in_channels&#91;1], kernel_size=3, padding=1, stride=1)\n        self.conv1_2 = Conv2d(in_channels=in_channels&#91;1], out_channels=in_channels&#91;1], kernel_size=3, padding=1, stride=1)\n        # \u5b9a\u4e49\u7b2c\u4e8c\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e24\u4e2a\u5377\u79ef\n        self.conv2_1 = Conv2d(in_channels=in_channels&#91;1], out_channels=in_channels&#91;2], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv2_2 = Conv2d(in_channels=in_channels&#91;2], out_channels=in_channels&#91;2], kernel_size=3, padding=1,\n                              stride=1)\n        # \u5b9a\u4e49\u7b2c\u4e09\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e09\u4e2a\u5377\u79ef\n        self.conv3_1 = Conv2d(in_channels=in_channels&#91;2], out_channels=in_channels&#91;3], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv3_2 = Conv2d(in_channels=in_channels&#91;3], out_channels=in_channels&#91;3], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv3_3 = Conv2d(in_channels=in_channels&#91;3], out_channels=in_channels&#91;3], kernel_size=3, padding=1,\n                              stride=1)\n        # \u5b9a\u4e49\u7b2c\u56db\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e09\u4e2a\u5377\u79ef\n        self.conv4_1 = Conv2d(in_channels=in_channels&#91;3], out_channels=in_channels&#91;4], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv4_2 = Conv2d(in_channels=in_channels&#91;4], out_channels=in_channels&#91;4], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv4_3 = Conv2d(in_channels=in_channels&#91;4], out_channels=in_channels&#91;4], kernel_size=3, padding=1,\n                              stride=1)\n        # \u5b9a\u4e49\u7b2c\u4e94\u4e2a\u5377\u79ef\u5757\uff0c\u5305\u542b\u4e09\u4e2a\u5377\u79ef\n        self.conv5_1 = Conv2d(in_channels=in_channels&#91;4], out_channels=in_channels&#91;5], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv5_2 = Conv2d(in_channels=in_channels&#91;5], out_channels=in_channels&#91;5], kernel_size=3, padding=1,\n                              stride=1)\n        self.conv5_3 = Conv2d(in_channels=in_channels&#91;5], out_channels=in_channels&#91;5], kernel_size=3, padding=1,\n                              stride=1)\n\n        # \u4f7f\u7528Sequential \u5c06\u5377\u79ef\u548crelu\u7ec4\u6210\u4e00\u4e2a\u7ebf\u6027\u7ed3\u6784\uff08fc + relu\uff09\n        # \u5f53\u8f93\u5165\u4e3a224x224\u65f6\uff0c\u7ecf\u8fc7\u4e94\u4e2a\u5377\u79ef\u5757\u548c\u6c60\u5316\u5c42\u540e\uff0c\u7279\u5f81\u7ef4\u5ea6\u53d8\u4e3a&#91;512x7x7]\n        self.fc1 = torch.nn.Sequential(torch.nn.Linear(512 * 7 * 7, 4096), torch.nn.ReLU())\n        self.drop1_ratio = 0.5\n        self.dropout1 = torch.nn.Dropout(self.drop1_ratio)\n        # \u4f7f\u7528Sequential \u5c06\u5377\u79ef\u548crelu\u7ec4\u6210\u4e00\u4e2a\u7ebf\u6027\u7ed3\u6784\uff08fc + relu\uff09\n        self.fc2 = torch.nn.Sequential(torch.nn.Linear(4096, 4096), torch.nn.ReLU())\n\n        self.drop2_ratio = 0.5\n        self.dropout2 = torch.nn.Dropout(self.drop2_ratio)\n        self.fc3 = torch.nn.Linear(4096, 1)\n\n        self.relu = torch.nn.ReLU()\n        self.pool = MaxPool2d(stride=2, kernel_size=2)\n\n    def forward(self, x):\n        x = self.relu(self.conv1_1(x))\n        x = self.relu(self.conv1_2(x))\n        x = self.pool(x)\n\n        x = self.relu(self.conv2_1(x))\n        x = self.relu(self.conv2_2(x))\n        x = self.pool(x)\n\n        x = self.relu(self.conv3_1(x))\n        x = self.relu(self.conv3_2(x))\n        x = self.relu(self.conv3_3(x))\n        x = self.pool(x)\n\n        x = self.relu(self.conv4_1(x))\n        x = self.relu(self.conv4_2(x))\n        x = self.relu(self.conv4_3(x))\n        x = self.pool(x)\n\n        x = self.relu(self.conv5_1(x))\n        x = self.relu(self.conv5_2(x))\n        x = self.relu(self.conv5_3(x))\n        x = self.pool(x)\n\n        x = torch.flatten(x, 1, -1)\n        x = self.dropout1(self.relu(self.fc1(x)))\n        x = self.dropout2(self.relu(self.fc2(x)))\n        x = self.fc3(x)\n        return x\n\nmodel = VGG()\n# opt = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())\nopt = optim.SGD(lr=0.001, params=model.parameters(), momentum=0.9)\n# \u542f\u52a8\u8bad\u7ec3\u8fc7\u7a0b\ntrain_pm(model, opt)<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">GoogLeNet<\/h2>\n\n\n\n<p>GoogLeNet\u662f2014\u5e74ImageNet\u6bd4\u8d5b\u7684\u51a0\u519b\uff0c\u5b83\u7684\u4e3b\u8981\u7279\u70b9\u662f\u7f51\u7edc\u4e0d\u4ec5\u6709\u6df1\u5ea6\uff0c\u8fd8\u5728\u6a2a\u5411\u4e0a\u5177\u6709\u201c\u5bbd\u5ea6\u201d\u3002\u7531\u4e8e\u56fe\u50cf\u4fe1\u606f\u5728\u7a7a\u95f4\u5c3a\u5bf8\u4e0a\u7684\u5de8\u5927\u5dee\u5f02\uff0c\u5982\u4f55\u9009\u62e9\u5408\u9002\u7684\u5377\u79ef\u6838\u6765\u63d0\u53d6\u7279\u5f81\u5c31\u663e\u5f97\u6bd4\u8f83\u56f0\u96be\u4e86\u3002\u7a7a\u95f4\u5206\u5e03\u8303\u56f4\u66f4\u5e7f\u7684\u56fe\u50cf\u4fe1\u606f\u9002\u5408\u7528\u8f83\u5927\u7684\u5377\u79ef\u6838\u6765\u63d0\u53d6\u5176\u7279\u5f81\uff1b\u800c\u7a7a\u95f4\u5206\u5e03\u8303\u56f4\u8f83\u5c0f\u7684\u56fe\u50cf\u4fe1\u606f\u5219\u9002\u5408\u7528\u8f83\u5c0f\u7684\u5377\u79ef\u6838\u6765\u63d0\u53d6\u5176\u7279\u5f81\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0cGoogLeNet\u63d0\u51fa\u4e86\u4e00\u79cd\u88ab\u79f0\u4e3aInception\u6a21\u5757\u7684\u65b9\u6848\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/ai-studio-static-online.cdn.bcebos.com\/ebc171e0281549a9b6aace1113f92fb72df08b947059446ca62a07b9af22e4b4\" alt=\"\"\/><\/figure>\n\n\n\n<p>Inception\u6a21\u5757\u5b9e\u73b0<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># GoogLeNet\u6a21\u578b\u4ee3\u7801\nimport numpy as np\nimport paddle\nfrom paddle.nn import Conv2D, MaxPool2D, AdaptiveAvgPool2D, Linear\n## \u7ec4\u7f51\nimport paddle.nn.functional as F\n\n# \u5b9a\u4e49Inception\u5757\nclass Inception(paddle.nn.Layer):\n    def __init__(self, c0, c1, c2, c3, c4, **kwargs):\n        '''\n        Inception\u6a21\u5757\u7684\u5b9e\u73b0\u4ee3\u7801\uff0c\n        \n        c1,\u56fe(b)\u4e2d\u7b2c\u4e00\u6761\u652f\u8def1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662f\u6574\u6570\n        c2,\u56fe(b)\u4e2d\u7b2c\u4e8c\u6761\u652f\u8def\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662ftuple\u6216list, \n               \u5176\u4e2dc2&#91;0]\u662f1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0cc2&#91;1]\u662f3x3\n        c3,\u56fe(b)\u4e2d\u7b2c\u4e09\u6761\u652f\u8def\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662ftuple\u6216list, \n               \u5176\u4e2dc3&#91;0]\u662f1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0cc3&#91;1]\u662f3x3\n        c4,\u56fe(b)\u4e2d\u7b2c\u4e00\u6761\u652f\u8def1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662f\u6574\u6570\n        '''\n        super(Inception, self).__init__()\n        # \u4f9d\u6b21\u521b\u5efaInception\u5757\u6bcf\u6761\u652f\u8def\u4e0a\u4f7f\u7528\u5230\u7684\u64cd\u4f5c\n        self.p1_1 = Conv2D(in_channels=c0,out_channels=c1, kernel_size=1)\n        self.p2_1 = Conv2D(in_channels=c0,out_channels=c2&#91;0], kernel_size=1)\n        self.p2_2 = Conv2D(in_channels=c2&#91;0],out_channels=c2&#91;1], kernel_size=3, padding=1)\n        self.p3_1 = Conv2D(in_channels=c0,out_channels=c3&#91;0], kernel_size=1)\n        self.p3_2 = Conv2D(in_channels=c3&#91;0],out_channels=c3&#91;1], kernel_size=5, padding=2)\n        self.p4_1 = MaxPool2D(kernel_size=3, stride=1, padding=1)\n        self.p4_2 = Conv2D(in_channels=c0,out_channels=c4, kernel_size=1)\n\n    def forward(self, x):\n        # \u652f\u8def1\u53ea\u5305\u542b\u4e00\u4e2a1x1\u5377\u79ef\n        p1 = F.relu(self.p1_1(x))\n        # \u652f\u8def2\u5305\u542b 1x1\u5377\u79ef + 3x3\u5377\u79ef\n        p2 = F.relu(self.p2_2(F.relu(self.p2_1(x))))\n        # \u652f\u8def3\u5305\u542b 1x1\u5377\u79ef + 5x5\u5377\u79ef\n        p3 = F.relu(self.p3_2(F.relu(self.p3_1(x))))\n        # \u652f\u8def4\u5305\u542b \u6700\u5927\u6c60\u5316\u548c1x1\u5377\u79ef\n        p4 = F.relu(self.p4_2(self.p4_1(x)))\n        # \u5c06\u6bcf\u4e2a\u652f\u8def\u7684\u8f93\u51fa\u7279\u5f81\u56fe\u62fc\u63a5\u5728\u4e00\u8d77\u4f5c\u4e3a\u6700\u7ec8\u7684\u8f93\u51fa\u7ed3\u679c\n        return paddle.concat(&#91;p1, p2, p3, p4], axis=1)<\/code><\/pre>\n\n\n\n<p>\u7b80\u5355\u6765\u8bf4Inception\u6a21\u5757\u5c31\u662f\u4e09\u79cd\u4e0d\u540c\u5c3a\u5bf8\u7684\u5377\u79ef\u6838\u5377\u79ef\u5b8c\u540e\u518dconcat\u8d77\u6765\uff0c\u56e0\u4e3astride\u90fd\u662f1\uff0c\u5e76\u4e14\u6709\u5404\u81ea\u7684padding\u6240\u4ee5\u8f93\u51fa\u7684\u56fe\u50cf\u5c3a\u5bf8\u662f\u4e0d\u53d8\u7684<\/p>\n\n\n\n<p>GoogLeNet\u7684\u67b6\u6784\u5982\u56fe\u6240\u793a\uff0c\u5728\u4e3b\u4f53\u5377\u79ef\u90e8\u5206\u4e2d\u4f7f\u75285\u4e2a\u6a21\u5757\uff08block\uff09\uff0c\u6bcf\u4e2a\u6a21\u5757\u4e4b\u95f4\u4f7f\u7528\u6b65\u5e45\u4e3a2\u76843 \u00d73\u6700\u5927\u6c60\u5316\u5c42\u6765\u51cf\u5c0f\u8f93\u51fa\u9ad8\u5bbd\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\"><li>\u7b2c\u4e00\u6a21\u5757\u4f7f\u7528\u4e00\u4e2a64\u901a\u9053\u76847 \u00d7 7\u5377\u79ef\u5c42\u3002<\/li><li>\u7b2c\u4e8c\u6a21\u5757\u4f7f\u75282\u4e2a\u5377\u79ef\u5c42:\u9996\u5148\u662f64\u901a\u9053\u76841 \u00d7 1\u5377\u79ef\u5c42\uff0c\u7136\u540e\u662f\u5c06\u901a\u9053\u589e\u59273\u500d\u76843 \u00d7 3\u5377\u79ef\u5c42\u3002<\/li><li>\u7b2c\u4e09\u6a21\u5757\u4e32\u80542\u4e2a\u5b8c\u6574\u7684Inception\u5757\u3002<\/li><li>\u7b2c\u56db\u6a21\u5757\u4e32\u8054\u4e865\u4e2aInception\u5757\u3002<\/li><li>\u7b2c\u4e94\u6a21\u5757\u4e32\u8054\u4e862 \u4e2aInception\u5757\u3002<\/li><li>\u7b2c\u4e94\u6a21\u5757\u7684\u540e\u9762\u7d27\u8ddf\u8f93\u51fa\u5c42\uff0c\u4f7f\u7528\u5168\u5c40\u5e73\u5747\u6c60\u5316\u5c42\u6765\u5c06\u6bcf\u4e2a\u901a\u9053\u7684\u9ad8\u548c\u5bbd\u53d8\u62101\uff0c\u6700\u540e\u63a5\u4e0a\u4e00\u4e2a\u8f93\u51fa\u4e2a\u6570\u4e3a\u6807\u7b7e\u7c7b\u522b\u6570\u7684\u5168\u8fde\u63a5\u5c42\u3002<\/li><\/ul>\n\n\n\n<hr class=\"wp-block-separator is-style-wide\"\/>\n\n\n\n<p>\u8bf4\u660e\uff1a \u5728\u539f\u4f5c\u8005\u7684\u8bba\u6587\u4e2d\u6dfb\u52a0\u4e86\u56fe\u4e2d\u6240\u793a\u7684softmax1\u548csoftmax2\u4e24\u4e2a\u8f85\u52a9\u5206\u7c7b\u5668\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u8bad\u7ec3\u65f6\u5c06\u4e09\u4e2a\u5206\u7c7b\u5668\u7684\u635f\u5931\u51fd\u6570\u8fdb\u884c\u52a0\u6743\u6c42\u548c\uff0c\u4ee5\u7f13\u89e3\u68af\u5ea6\u6d88\u5931\u73b0\u8c61\u3002\u8fd9\u91cc\u7684\u7a0b\u5e8f\u4f5c\u4e86\u7b80\u5316\uff0c\u6ca1\u6709\u52a0\u5165\u8f85\u52a9\u5206\u7c7b\u5668\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator is-style-wide\"\/>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/ai-studio-static-online.cdn.bcebos.com\/9d0794b330934bc9be72cba9f056d62eb77d3ba6c2ac450fae64cf86d86f2e04\" alt=\"\"\/><\/figure>\n\n\n\n<p>GoogLeNet\u5177\u4f53\u5b9e\u73b0<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># GoogLeNet\u6a21\u578b\u4ee3\u7801\nimport numpy as np\nimport paddle\nfrom paddle.nn import Conv2D, MaxPool2D, AdaptiveAvgPool2D, Linear\n## \u7ec4\u7f51\nimport paddle.nn.functional as F\n\n# \u5b9a\u4e49Inception\u5757\nclass Inception(paddle.nn.Layer):\n    def __init__(self, c0, c1, c2, c3, c4, **kwargs):\n        '''\n        Inception\u6a21\u5757\u7684\u5b9e\u73b0\u4ee3\u7801\uff0c\n        \n        c1,\u56fe(b)\u4e2d\u7b2c\u4e00\u6761\u652f\u8def1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662f\u6574\u6570\n        c2,\u56fe(b)\u4e2d\u7b2c\u4e8c\u6761\u652f\u8def\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662ftuple\u6216list, \n               \u5176\u4e2dc2&#91;0]\u662f1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0cc2&#91;1]\u662f3x3\n        c3,\u56fe(b)\u4e2d\u7b2c\u4e09\u6761\u652f\u8def\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662ftuple\u6216list, \n               \u5176\u4e2dc3&#91;0]\u662f1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0cc3&#91;1]\u662f3x3\n        c4,\u56fe(b)\u4e2d\u7b2c\u4e00\u6761\u652f\u8def1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662f\u6574\u6570\n        '''\n        super(Inception, self).__init__()\n        # \u4f9d\u6b21\u521b\u5efaInception\u5757\u6bcf\u6761\u652f\u8def\u4e0a\u4f7f\u7528\u5230\u7684\u64cd\u4f5c\n        self.p1_1 = Conv2D(in_channels=c0,out_channels=c1, kernel_size=1, stride=1)\n        self.p2_1 = Conv2D(in_channels=c0,out_channels=c2&#91;0], kernel_size=1, stride=1)\n        self.p2_2 = Conv2D(in_channels=c2&#91;0],out_channels=c2&#91;1], kernel_size=3, padding=1, stride=1)\n        self.p3_1 = Conv2D(in_channels=c0,out_channels=c3&#91;0], kernel_size=1, stride=1)\n        self.p3_2 = Conv2D(in_channels=c3&#91;0],out_channels=c3&#91;1], kernel_size=5, padding=2, stride=1)\n        self.p4_1 = MaxPool2D(kernel_size=3, stride=1, padding=1)\n        self.p4_2 = Conv2D(in_channels=c0,out_channels=c4, kernel_size=1, stride=1)\n        \n        # # \u65b0\u52a0\u4e00\u5c42batchnorm\u7a33\u5b9a\u6536\u655b\n        # self.batchnorm = paddle.nn.BatchNorm2D(c1+c2&#91;1]+c3&#91;1]+c4)\n\n    def forward(self, x):\n        # \u652f\u8def1\u53ea\u5305\u542b\u4e00\u4e2a1x1\u5377\u79ef\n        p1 = F.relu(self.p1_1(x))\n        # \u652f\u8def2\u5305\u542b 1x1\u5377\u79ef + 3x3\u5377\u79ef\n        p2 = F.relu(self.p2_2(F.relu(self.p2_1(x))))\n        # \u652f\u8def3\u5305\u542b 1x1\u5377\u79ef + 5x5\u5377\u79ef\n        p3 = F.relu(self.p3_2(F.relu(self.p3_1(x))))\n        # \u652f\u8def4\u5305\u542b \u6700\u5927\u6c60\u5316\u548c1x1\u5377\u79ef\n        p4 = F.relu(self.p4_2(self.p4_1(x)))\n        # \u5c06\u6bcf\u4e2a\u652f\u8def\u7684\u8f93\u51fa\u7279\u5f81\u56fe\u62fc\u63a5\u5728\u4e00\u8d77\u4f5c\u4e3a\u6700\u7ec8\u7684\u8f93\u51fa\u7ed3\u679c\n        return paddle.concat(&#91;p1, p2, p3, p4], axis=1)\n        # return self.batchnorm()\n    \nclass GoogLeNet(paddle.nn.Layer):\n    def __init__(self):\n        super(GoogLeNet, self).__init__()\n        # GoogLeNet\u5305\u542b\u4e94\u4e2a\u6a21\u5757\uff0c\u6bcf\u4e2a\u6a21\u5757\u540e\u9762\u7d27\u8ddf\u4e00\u4e2a\u6c60\u5316\u5c42\n        # \u7b2c\u4e00\u4e2a\u6a21\u5757\u5305\u542b1\u4e2a\u5377\u79ef\u5c42\n        self.conv1 = Conv2D(in_channels=3,out_channels=64, kernel_size=7, padding=3, stride=1)\n        # 3x3\u6700\u5927\u6c60\u5316\n        self.pool1 = MaxPool2D(kernel_size=3, stride=2, padding=1)\n        # \u7b2c\u4e8c\u4e2a\u6a21\u5757\u5305\u542b2\u4e2a\u5377\u79ef\u5c42\n        self.conv2_1 = Conv2D(in_channels=64,out_channels=64, kernel_size=1, stride=1)\n        self.conv2_2 = Conv2D(in_channels=64,out_channels=192, kernel_size=3, padding=1, stride=1)\n        # 3x3\u6700\u5927\u6c60\u5316\n        self.pool2 = MaxPool2D(kernel_size=3, stride=2, padding=1)\n        # \u7b2c\u4e09\u4e2a\u6a21\u5757\u5305\u542b2\u4e2aInception\u5757\n        self.block3_1 = Inception(192, 64, (96, 128), (16, 32), 32)\n        self.block3_2 = Inception(256, 128, (128, 192), (32, 96), 64)\n        # 3x3\u6700\u5927\u6c60\u5316\n        self.pool3 = MaxPool2D(kernel_size=3, stride=2, padding=1)\n        # \u7b2c\u56db\u4e2a\u6a21\u5757\u5305\u542b5\u4e2aInception\u5757\n        self.block4_1 = Inception(480, 192, (96, 208), (16, 48), 64)\n        self.block4_2 = Inception(512, 160, (112, 224), (24, 64), 64)\n        self.block4_3 = Inception(512, 128, (128, 256), (24, 64), 64)\n        self.block4_4 = Inception(512, 112, (144, 288), (32, 64), 64)\n        self.block4_5 = Inception(528, 256, (160, 320), (32, 128), 128)\n        # 3x3\u6700\u5927\u6c60\u5316\n        self.pool4 = MaxPool2D(kernel_size=3, stride=2, padding=1)\n        # \u7b2c\u4e94\u4e2a\u6a21\u5757\u5305\u542b2\u4e2aInception\u5757\n        self.block5_1 = Inception(832, 256, (160, 320), (32, 128), 128)\n        self.block5_2 = Inception(832, 384, (192, 384), (48, 128), 128)\n        # \u5168\u5c40\u6c60\u5316\uff0c\u5c3a\u5bf8\u7528\u7684\u662fglobal_pooling\uff0cpool_stride\u4e0d\u8d77\u4f5c\u7528\n        self.pool5 = AdaptiveAvgPool2D(output_size=1)\n        self.fc = Linear(in_features=1024, out_features=1)\n\n    def forward(self, x):\n        x = self.pool1(F.relu(self.conv1(x)))\n        x = self.pool2(F.relu(self.conv2_2(F.relu(self.conv2_1(x)))))\n        x = self.pool3(self.block3_2(self.block3_1(x)))\n        x = self.block4_3(self.block4_2(self.block4_1(x)))\n        x = self.pool4(self.block4_5(self.block4_4(x)))\n        x = self.pool5(self.block5_2(self.block5_1(x)))\n        x = paddle.reshape(x, &#91;x.shape&#91;0], -1])\n        x = self.fc(x)\n        return x\n\n# \u521b\u5efa\u6a21\u578b\nmodel = GoogLeNet()\nprint(len(model.parameters()))\nopt = paddle.optimizer.Momentum(learning_rate=0.001, momentum=0.9, parameters=model.parameters(), weight_decay=0.001)\n# \u542f\u52a8\u8bad\u7ec3\u8fc7\u7a0b\ntrain_pm(model, opt)<\/code><\/pre>\n\n\n\n<p>\u5176\u5b9eGoogLeNet\u7684\u7ed3\u6784\u8fd8\u662f\u633a\u660e\u767d\u7684\uff0c\u4e0d\u7ba1\u662f\u5377\u79ef\u5c42\u8fd8\u662f\u7ec4\u5408\u7684Inception\u5c42\u90fd\u4e0d\u4f1a\u6539\u53d8\u56fe\u50cf\u5c3a\u5bf8\u5927\u5c0f\u800c\u53ea\u662f\u589e\u52a0\u901a\u9053\u6570\uff0c\u6bcf\u6b21\u6c60\u5316\u5c06\u56fe\u50cf\u5927\u5c0f\u51cf\u534a\uff0c\u6700\u540e\u901a\u8fc7\u81ea\u9002\u5e94\u5e73\u5747\u6c60\u5316\u5c42\u4f7f\u6bcf\u4e2a\u901a\u9053\u53d8\u62101\u00d71\u7684size\uff0c\u518d\u901a\u8fc7\u7ebf\u6027\u5c42\u5f97\u5230\u6700\u7ec8\u8f93\u51fa\u3002<\/p>\n\n\n\n<p>\u540c\u6837\u4f7f\u7528pytorch\u6765\u5b9e\u73b0GoogLeNet\u7f51\u7edc\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># GoogLeNet\u6a21\u578b\u4ee3\u7801\nimport numpy as np\nimport torch\nfrom torch.nn import Conv2d, MaxPool2d, AdaptiveAvgPool2d, Linear\n## \u7ec4\u7f51\nimport torch.nn.functional as F\n\n# \u5b9a\u4e49Inception\u5757\nclass Inception(torch.nn.Module):\n    def __init__(self, c0, c1, c2, c3, c4, **kwargs):\n        '''\n        Inception\u6a21\u5757\u7684\u5b9e\u73b0\u4ee3\u7801\uff0c\n        \n        c1,\u56fe(b)\u4e2d\u7b2c\u4e00\u6761\u652f\u8def1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662f\u6574\u6570\n        c2,\u56fe(b)\u4e2d\u7b2c\u4e8c\u6761\u652f\u8def\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662ftuple\u6216list, \n               \u5176\u4e2dc2&#91;0]\u662f1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0cc2&#91;1]\u662f3x3\n        c3,\u56fe(b)\u4e2d\u7b2c\u4e09\u6761\u652f\u8def\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662ftuple\u6216list, \n               \u5176\u4e2dc3&#91;0]\u662f1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0cc3&#91;1]\u662f3x3\n        c4,\u56fe(b)\u4e2d\u7b2c\u4e00\u6761\u652f\u8def1x1\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\uff0c\u6570\u636e\u7c7b\u578b\u662f\u6574\u6570\n        '''\n        super(Inception, self).__init__()\n        # \u4f9d\u6b21\u521b\u5efaInception\u5757\u6bcf\u6761\u652f\u8def\u4e0a\u4f7f\u7528\u5230\u7684\u64cd\u4f5c\n        self.p1_1 = Conv2d(in_channels=c0,out_channels=c1, kernel_size=1, stride=1)\n        self.p2_1 = Conv2d(in_channels=c0,out_channels=c2&#91;0], kernel_size=1, stride=1)\n        self.p2_2 = Conv2d(in_channels=c2&#91;0],out_channels=c2&#91;1], kernel_size=3, padding=1, stride=1)\n        self.p3_1 = Conv2d(in_channels=c0,out_channels=c3&#91;0], kernel_size=1, stride=1)\n        self.p3_2 = Conv2d(in_channels=c3&#91;0],out_channels=c3&#91;1], kernel_size=5, padding=2, stride=1)\n        self.p4_1 = MaxPool2d(kernel_size=3, stride=1, padding=1)\n        self.p4_2 = Conv2d(in_channels=c0,out_channels=c4, kernel_size=1, stride=1)\n        \n        # # \u65b0\u52a0\u4e00\u5c42batchnorm\u7a33\u5b9a\u6536\u655b\n        # self.batchnorm = torch.nn.BatchNorm2d(c1+c2&#91;1]+c3&#91;1]+c4)\n\n    def forward(self, x):\n        # \u652f\u8def1\u53ea\u5305\u542b\u4e00\u4e2a1x1\u5377\u79ef\n        p1 = F.relu(self.p1_1(x))\n        # \u652f\u8def2\u5305\u542b 1x1\u5377\u79ef + 3x3\u5377\u79ef\n        p2 = F.relu(self.p2_2(F.relu(self.p2_1(x))))\n        # \u652f\u8def3\u5305\u542b 1x1\u5377\u79ef + 5x5\u5377\u79ef\n        p3 = F.relu(self.p3_2(F.relu(self.p3_1(x))))\n        # \u652f\u8def4\u5305\u542b \u6700\u5927\u6c60\u5316\u548c1x1\u5377\u79ef\n        p4 = F.relu(self.p4_2(self.p4_1(x)))\n        # \u5c06\u6bcf\u4e2a\u652f\u8def\u7684\u8f93\u51fa\u7279\u5f81\u56fe\u62fc\u63a5\u5728\u4e00\u8d77\u4f5c\u4e3a\u6700\u7ec8\u7684\u8f93\u51fa\u7ed3\u679c\n        return torch.concat(&#91;p1, p2, p3, p4], axis=1)\n        # return self.batchnorm()\n    \nclass GoogLeNet(torch.nn.Module):\n    def __init__(self):\n        super(GoogLeNet, self).__init__()\n        # GoogLeNet\u5305\u542b\u4e94\u4e2a\u6a21\u5757\uff0c\u6bcf\u4e2a\u6a21\u5757\u540e\u9762\u7d27\u8ddf\u4e00\u4e2a\u6c60\u5316\u5c42\n        # \u7b2c\u4e00\u4e2a\u6a21\u5757\u5305\u542b1\u4e2a\u5377\u79ef\u5c42\n        self.conv1 = Conv2d(in_channels=3,out_channels=64, kernel_size=7, padding=3, stride=1)\n        # 3x3\u6700\u5927\u6c60\u5316\n        self.pool1 = MaxPool2d(kernel_size=3, stride=2, padding=1)\n        # \u7b2c\u4e8c\u4e2a\u6a21\u5757\u5305\u542b2\u4e2a\u5377\u79ef\u5c42\n        self.conv2_1 = Conv2d(in_channels=64,out_channels=64, kernel_size=1, stride=1)\n        self.conv2_2 = Conv2d(in_channels=64,out_channels=192, kernel_size=3, padding=1, stride=1)\n        # 3x3\u6700\u5927\u6c60\u5316\n        self.pool2 = MaxPool2d(kernel_size=3, stride=2, padding=1)\n        # \u7b2c\u4e09\u4e2a\u6a21\u5757\u5305\u542b2\u4e2aInception\u5757\n        self.block3_1 = Inception(192, 64, (96, 128), (16, 32), 32)\n        self.block3_2 = Inception(256, 128, (128, 192), (32, 96), 64)\n        # 3x3\u6700\u5927\u6c60\u5316\n        self.pool3 = MaxPool2d(kernel_size=3, stride=2, padding=1)\n        # \u7b2c\u56db\u4e2a\u6a21\u5757\u5305\u542b5\u4e2aInception\u5757\n        self.block4_1 = Inception(480, 192, (96, 208), (16, 48), 64)\n        self.block4_2 = Inception(512, 160, (112, 224), (24, 64), 64)\n        self.block4_3 = Inception(512, 128, (128, 256), (24, 64), 64)\n        self.block4_4 = Inception(512, 112, (144, 288), (32, 64), 64)\n        self.block4_5 = Inception(528, 256, (160, 320), (32, 128), 128)\n        # 3x3\u6700\u5927\u6c60\u5316\n        self.pool4 = MaxPool2d(kernel_size=3, stride=2, padding=1)\n        # \u7b2c\u4e94\u4e2a\u6a21\u5757\u5305\u542b2\u4e2aInception\u5757\n        self.block5_1 = Inception(832, 256, (160, 320), (32, 128), 128)\n        self.block5_2 = Inception(832, 384, (192, 384), (48, 128), 128)\n        # \u5168\u5c40\u6c60\u5316\uff0c\u5c3a\u5bf8\u7528\u7684\u662fglobal_pooling\uff0cpool_stride\u4e0d\u8d77\u4f5c\u7528\n        self.pool5 = AdaptiveAvgPool2d(output_size=1)\n        self.fc = Linear(in_features=1024, out_features=1)\n\n    def forward(self, x):\n        x = self.pool1(F.relu(self.conv1(x)))\n        x = self.pool2(F.relu(self.conv2_2(F.relu(self.conv2_1(x)))))\n        x = self.pool3(self.block3_2(self.block3_1(x)))\n        x = self.block4_3(self.block4_2(self.block4_1(x)))\n        x = self.pool4(self.block4_5(self.block4_4(x)))\n        x = self.pool5(self.block5_2(self.block5_1(x)))\n        x = torch.reshape(x, &#91;x.shape&#91;0], -1])\n        x = self.fc(x)\n        return x\n\n# \u521b\u5efa\u6a21\u578b\nmodel = GoogLeNet()\n#print(model.parameters)\nopt = optim.SGD(lr=0.001, params=model.parameters(), momentum=0.9)\n# \u542f\u52a8\u8bad\u7ec3\u8fc7\u7a0b\ntrain_pm(model, opt)<\/code><\/pre>\n\n\n\n<p>\u603b\u7ed3\u4e00\u4e0b\u628apaddle\u6a21\u578b\u8f6c\u6362\u6210pytorch\u7684\u6b65\u9aa4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\"><li>paddle\u6362\u6210torch<\/li><li>\u7f51\u7edc\u7ee7\u627f\u7c7b\u7531paddle.nn.Layer\u6539\u6210torch.nn.Module<\/li><li>Conv2D\uff0cMaxPool2D\u7b49\u6539\u6210Conv2d\uff0cMaxPool2d<\/li><li>\u4f18\u5316\u5668\u6539\u6210optim.SGD\u7b49\u7b49\uff0c\u53c2\u6570\u90fd\u662f\u7f29\u5199<\/li><li>torch\u7684GPU\u6a21\u5f0f\u590d\u6742\u4e00\u70b9\uff0c\u8981\u52a0\u597d\u51e0\u4e2acuda()<\/li><\/ul>\n\n\n\n<p class=\"has-text-color\" style=\"color:#d55bac\">jupyter notebook\u7684\u66ff\u6362\u5feb\u6377\u952e\u662fESC+F<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">ResNet<\/h2>\n\n\n\n<p>ResNet\u662f2015\u5e74ImageNet\u6bd4\u8d5b\u7684\u51a0\u519b\uff0c\u5c06\u8bc6\u522b\u9519\u8bef\u7387\u964d\u4f4e\u5230\u4e863.6%\uff0c\u8fd9\u4e2a\u7ed3\u679c\u751a\u81f3\u8d85\u51fa\u4e86\u6b63\u5e38\u4eba\u773c\u8bc6\u522b\u7684\u7cbe\u5ea6\u3002<\/p>\n\n\n\n<p>\u901a\u8fc7\u524d\u9762\u51e0\u4e2a\u7ecf\u5178\u6a21\u578b\u5b66\u4e60\uff0c\u6211\u4eec\u53ef\u4ee5\u53d1\u73b0\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u4e0d\u65ad\u53d1\u5c55\uff0c\u6a21\u578b\u7684\u5c42\u6570\u8d8a\u6765\u8d8a\u591a\uff0c\u7f51\u7edc\u7ed3\u6784\u4e5f\u8d8a\u6765\u8d8a\u590d\u6742\u3002\u90a3\u4e48\u662f\u5426\u52a0\u6df1\u7f51\u7edc\u7ed3\u6784\uff0c\u5c31\u4e00\u5b9a\u4f1a\u5f97\u5230\u66f4\u597d\u7684\u6548\u679c\u5462\uff1f\u4ece\u7406\u8bba\u4e0a\u6765\u8bf4\uff0c\u5047\u8bbe\u65b0\u589e\u52a0\u7684\u5c42\u90fd\u662f\u6052\u7b49\u6620\u5c04\uff0c\u53ea\u8981\u539f\u6709\u7684\u5c42\u5b66\u51fa\u8ddf\u539f\u6a21\u578b\u4e00\u6837\u7684\u53c2\u6570\uff0c\u90a3\u4e48\u6df1\u6a21\u578b\u7ed3\u6784\u5c31\u80fd\u8fbe\u5230\u539f\u6a21\u578b\u7ed3\u6784\u7684\u6548\u679c\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u539f\u6a21\u578b\u7684\u89e3\u53ea\u662f\u65b0\u6a21\u578b\u7684\u89e3\u7684\u5b50\u7a7a\u95f4\uff0c\u5728\u65b0\u6a21\u578b\u89e3\u7684\u7a7a\u95f4\u91cc\u5e94\u8be5\u80fd\u627e\u5230\u6bd4\u539f\u6a21\u578b\u89e3\u5bf9\u5e94\u7684\u5b50\u7a7a\u95f4\u66f4\u597d\u7684\u7ed3\u679c\u3002\u4f46\u662f\u5b9e\u8df5\u8868\u660e\uff0c\u589e\u52a0\u7f51\u7edc\u7684\u5c42\u6570\u4e4b\u540e\uff0c\u8bad\u7ec3\u8bef\u5dee\u5f80\u5f80\u4e0d\u964d\u53cd\u5347\u3002<\/p>\n\n\n\n<p>Kaiming He\u7b49\u4eba\u63d0\u51fa\u4e86\u6b8b\u5dee\u7f51\u7edcResNet\u6765\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u5176\u57fa\u672c\u601d\u60f3\u5982\u56fe\u6240\u793a\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\"><li>\u56fe6(a)\uff1a\u8868\u793a\u589e\u52a0\u7f51\u7edc\u7684\u65f6\u5019\uff0c\u5c06<em>x<\/em>\u6620\u5c04\u6210y=F(x)\u8f93\u51fa\u3002<\/li><li>\u56fe6(b)\uff1a\u5bf9\u56fe6(a)\u4f5c\u4e86\u6539\u8fdb\uff0c\u8f93\u51fay=F(x)+x\u3002\u8fd9\u65f6\u4e0d\u662f\u76f4\u63a5\u5b66\u4e60\u8f93\u51fa\u7279\u5f81yy<em>y<\/em>\u7684\u8868\u793a\uff0c\u800c\u662f\u5b66\u4e60y\u2212x\u3002<ul><li>\u5982\u679c\u60f3\u5b66\u4e60\u51fa\u539f\u6a21\u578b\u7684\u8868\u793a\uff0c\u53ea\u9700\u5c06F(x)\u7684\u53c2\u6570\u5168\u90e8\u8bbe\u7f6e\u4e3a0\uff0c\u5219y=x\u662f\u6052\u7b49\u6620\u5c04\u3002<\/li><li>F(x)=y\u2212x\u4e5f\u53eb\u505a\u6b8b\u5dee\u9879\uff0c\u5982\u679cx\u2192y\u7684\u6620\u5c04\u63a5\u8fd1\u6052\u7b49\u6620\u5c04\uff0c\u56fe6(b)\u4e2d\u901a\u8fc7\u5b66\u4e60\u6b8b\u5dee\u9879\u4e5f\u6bd4\u56fe6(a)\u5b66\u4e60\u5b8c\u6574\u6620\u5c04\u5f62\u5f0f\u66f4\u52a0\u5bb9\u6613\u3002<\/li><\/ul><\/li><\/ul>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img decoding=\"async\" src=\"https:\/\/ai-studio-static-online.cdn.bcebos.com\/e10f22f054704daabf4261ab46719629a36749631db74eb0a368499de3e5d3d6\" alt=\"\"\/><\/figure><\/div>\n\n\n\n<p>\u56fe6(b)\u7684\u7ed3\u6784\u662f\u6b8b\u5dee\u7f51\u7edc\u7684\u57fa\u7840\uff0c\u8fd9\u79cd\u7ed3\u6784\u4e5f\u53eb\u505a\u6b8b\u5dee\u5757\uff08Residual block\uff09\u3002\u8f93\u5165xx<em>x<\/em>\u901a\u8fc7\u8de8\u5c42\u8fde\u63a5\uff0c\u80fd\u66f4\u5feb\u7684\u5411\u524d\u4f20\u64ad\u6570\u636e\uff0c\u6216\u8005\u5411\u540e\u4f20\u64ad\u68af\u5ea6\u3002\u901a\u4fd7\u7684\u6bd4\u55bb\uff0c\u5728\u706b\u70ed\u7684\u7535\u89c6\u8282\u76ee\u300a\u738b\u724c\u5bf9\u738b\u724c\u300b\u4e0a\u6709\u4e00\u4e2a\u201c\u4f20\u58f0\u7b52\u201d\u7684\u6e38\u620f\uff0c\u6392\u5728\u961f\u9996\u7684\u5609\u5bbe\u628a\u770b\u5230\u7684\u5f71\u89c6\u7247\u6bb5\u8868\u6f14\u7ed9\u540e\u9762\u4e00\u4e2a\u5609\u5bbe\u770b\uff0c\u7ecf\u8fc7\u56db\u4e94\u4e2a\u5609\u5bbe\u540e\uff0c\u6700\u540e\u4e00\u4e2a\u5609\u5bbe\u5982\u679c\u80fd\u8868\u6f14\u51fa\u66f4\u591a\u539f\u5267\u7684\u5185\u5bb9\uff0c\u5c31\u80fd\u53d6\u5f97\u9ad8\u5206\u3002\u6211\u4eec\u5e38\u5e38\u4f1a\u53d1\u73b0\u521a\u5f00\u59cb\u7684\u5609\u5bbe\u5f80\u5f80\u8868\u6f14\u51fa\u6700\u591a\u7684\u4fe1\u606f\uff08\u7c7b\u4f3c\u4e8eLoss\uff09\uff0c\u800c\u968f\u7740\u8868\u6f14\u7684\u4f20\u9012\uff0c\u6709\u6548\u7684\u8868\u6f14\u4fe1\u606f\u8d8a\u6765\u8d8a\u5c11\uff08\u7c7b\u4f3c\u4e8e\u68af\u5ea6\u5f25\u6563\uff09\u3002\u5982\u679c\u6bcf\u4e2a\u5609\u5bbe\u90fd\u80fd\u770b\u5230\u539f\u59cb\u7684\u5f71\u89c6\u7247\u6bb5\uff0c\u90a3\u4e48\u76f8\u4fe1\u4f20\u58f0\u7b52\u7684\u6548\u679c\u4f1a\u597d\u5f88\u591a\u3002\u7c7b\u4f3c\u7684\uff0c\u7531\u4e8eResNet\u6bcf\u5c42\u90fd\u5b58\u5728\u76f4\u8fde\u7684\u65c1\u8def\uff0c\u76f8\u5f53\u4e8e\u6bcf\u4e00\u5c42\u90fd\u548c\u6700\u7ec8\u7684\u635f\u5931\u6709\u201c\u76f4\u63a5\u5bf9\u8bdd\u201d\u7684\u673a\u4f1a\uff0c\u81ea\u7136\u53ef\u4ee5\u66f4\u597d\u7684\u89e3\u51b3\u68af\u5ea6\u5f25\u6563\u7684\u95ee\u9898\u3002\u6b8b\u5dee\u5757\u7684\u5177\u4f53\u8bbe\u8ba1\u65b9\u6848\u5982&nbsp;<strong>\u56fe<\/strong>7 \u6240\u793a\uff0c\u8fd9\u79cd\u8bbe\u8ba1\u65b9\u6848\u4e5f\u5e38\u79f0\u4f5c\u74f6\u9888\u7ed3\u6784\uff08BottleNeck\uff09\u30021*1\u7684\u5377\u79ef\u6838\u53ef\u4ee5\u975e\u5e38\u65b9\u4fbf\u7684\u8c03\u6574\u4e2d\u95f4\u5c42\u7684\u901a\u9053\u6570\uff0c\u5728\u8fdb\u51653*3\u7684\u5377\u79ef\u5c42\u4e4b\u524d\u51cf\u5c11\u901a\u9053\u6570\uff08256-&gt;64\uff09\uff0c\u7ecf\u8fc7\u8be5\u5377\u79ef\u5c42\u540e\u518d\u6062\u590d\u901a\u9053\u6570(64-&gt;256)\uff0c\u53ef\u4ee5\u663e\u8457\u51cf\u5c11\u7f51\u7edc\u7684\u53c2\u6570\u91cf\u3002\u8fd9\u4e2a\u7ed3\u6784\uff08256-&gt;64-&gt;256\uff09\u50cf\u4e00\u4e2a\u4e2d\u95f4\u7ec6\uff0c\u4e24\u5934\u7c97\u7684\u74f6\u9888\uff0c\u6240\u4ee5\u88ab\u79f0\u4e3a\u201cBottleNeck\u201d\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img decoding=\"async\" src=\"https:\/\/ai-studio-static-online.cdn.bcebos.com\/322b26358d43401ba81546dd134a310cfb11ecafb3314aab88b5885ff642870b\" alt=\"\"\/><\/figure><\/div>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img decoding=\"async\" src=\"https:\/\/ai-studio-static-online.cdn.bcebos.com\/8f42b3b5b7b34e45847a9c61580f1f8239a80ca6fa67448e8baeeb0209a2d556\" alt=\"\"\/><\/figure><\/div>\n\n\n\n<p>ResNet\u5b9e\u73b0\u4ee3\u7801\u5982\u4e0b<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># -*- coding:utf-8 -*-\n\n# ResNet\u6a21\u578b\u4ee3\u7801\nimport numpy as np\nimport paddle\nimport paddle.nn as nn\nimport paddle.nn.functional as F\n\n# ResNet\u4e2d\u4f7f\u7528\u4e86BatchNorm\u5c42\uff0c\u5728\u5377\u79ef\u5c42\u7684\u540e\u9762\u52a0\u4e0aBatchNorm\u4ee5\u63d0\u5347\u6570\u503c\u7a33\u5b9a\u6027\n# \u5b9a\u4e49\u5377\u79ef\u6279\u5f52\u4e00\u5316\u5757\nclass ConvBNLayer(paddle.nn.Layer):\n    def __init__(self,\n                 num_channels,\n                 num_filters,\n                 filter_size,\n                 stride=1,\n                 groups=1,\n                 act=None):\n       \n        \"\"\"\n        num_channels, \u5377\u79ef\u5c42\u7684\u8f93\u5165\u901a\u9053\u6570\n        num_filters, \u5377\u79ef\u5c42\u7684\u8f93\u51fa\u901a\u9053\u6570\n        stride, \u5377\u79ef\u5c42\u7684\u6b65\u5e45\n        groups, \u5206\u7ec4\u5377\u79ef\u7684\u7ec4\u6570\uff0c\u9ed8\u8ba4groups=1\u4e0d\u4f7f\u7528\u5206\u7ec4\u5377\u79ef\n        \"\"\"\n        super(ConvBNLayer, self).__init__()\n\n        # \u521b\u5efa\u5377\u79ef\u5c42\n        self._conv = nn.Conv2D(\n            in_channels=num_channels,\n            out_channels=num_filters,\n            kernel_size=filter_size,\n            stride=stride,\n            padding=(filter_size - 1) \/\/ 2,\n            groups=groups,\n            bias_attr=False)\n\n        # \u521b\u5efaBatchNorm\u5c42\n        self._batch_norm = paddle.nn.BatchNorm2D(num_filters)\n        \n        self.act = act\n\n    def forward(self, inputs):\n        y = self._conv(inputs)\n        y = self._batch_norm(y)\n        if self.act == 'leaky':\n            y = F.leaky_relu(x=out, negative_slope=0.1)\n        elif self.act == 'relu':\n            y = F.relu(x=y)\n        return y\n\n# \u5b9a\u4e49\u6b8b\u5dee\u5757\n# \u6bcf\u4e2a\u6b8b\u5dee\u5757\u4f1a\u5bf9\u8f93\u5165\u56fe\u7247\u505a\u4e09\u6b21\u5377\u79ef\uff0c\u7136\u540e\u8ddf\u8f93\u5165\u56fe\u7247\u8fdb\u884c\u77ed\u63a5\n# \u5982\u679c\u6b8b\u5dee\u5757\u4e2d\u7b2c\u4e09\u6b21\u5377\u79ef\u8f93\u51fa\u7279\u5f81\u56fe\u7684\u5f62\u72b6\u4e0e\u8f93\u5165\u4e0d\u4e00\u81f4\uff0c\u5219\u5bf9\u8f93\u5165\u56fe\u7247\u505a1x1\u5377\u79ef\uff0c\u5c06\u5176\u8f93\u51fa\u5f62\u72b6\u8c03\u6574\u6210\u4e00\u81f4\nclass BottleneckBlock(paddle.nn.Layer):\n    def __init__(self,\n                 num_channels,\n                 num_filters,\n                 stride,\n                 shortcut=True):\n        super(BottleneckBlock, self).__init__()\n        # \u521b\u5efa\u7b2c\u4e00\u4e2a\u5377\u79ef\u5c42 1x1\n        self.conv0 = ConvBNLayer(\n            num_channels=num_channels,\n            num_filters=num_filters,\n            filter_size=1,\n            act='relu')\n        # \u521b\u5efa\u7b2c\u4e8c\u4e2a\u5377\u79ef\u5c42 3x3\n        self.conv1 = ConvBNLayer(\n            num_channels=num_filters,\n            num_filters=num_filters,\n            filter_size=3,\n            stride=stride,\n            act='relu')\n        # \u521b\u5efa\u7b2c\u4e09\u4e2a\u5377\u79ef 1x1\uff0c\u4f46\u8f93\u51fa\u901a\u9053\u6570\u4e58\u4ee54\n        self.conv2 = ConvBNLayer(\n            num_channels=num_filters,\n            num_filters=num_filters * 4,\n            filter_size=1,\n            act=None)\n\n        # \u5982\u679cconv2\u7684\u8f93\u51fa\u8ddf\u6b64\u6b8b\u5dee\u5757\u7684\u8f93\u5165\u6570\u636e\u5f62\u72b6\u4e00\u81f4\uff0c\u5219shortcut=True\n        # \u5426\u5219shortcut = False\uff0c\u6dfb\u52a01\u4e2a1x1\u7684\u5377\u79ef\u4f5c\u7528\u5728\u8f93\u5165\u6570\u636e\u4e0a\uff0c\u4f7f\u5176\u5f62\u72b6\u53d8\u6210\u8ddfconv2\u4e00\u81f4\n        if not shortcut:\n            self.short = ConvBNLayer(\n                num_channels=num_channels,\n                num_filters=num_filters * 4,\n                filter_size=1,\n                stride=stride)\n\n        self.shortcut = shortcut\n\n        self._num_channels_out = num_filters * 4\n\n    def forward(self, inputs):\n        y = self.conv0(inputs)\n        conv1 = self.conv1(y)\n        conv2 = self.conv2(conv1)\n\n        # \u5982\u679cshortcut=True\uff0c\u76f4\u63a5\u5c06inputs\u8ddfconv2\u7684\u8f93\u51fa\u76f8\u52a0\n        # \u5426\u5219\u9700\u8981\u5bf9inputs\u8fdb\u884c\u4e00\u6b21\u5377\u79ef\uff0c\u5c06\u5f62\u72b6\u8c03\u6574\u6210\u8ddfconv2\u8f93\u51fa\u4e00\u81f4\n        if self.shortcut:\n            short = inputs\n        else:\n            short = self.short(inputs)\n\n        y = paddle.add(x=short, y=conv2)\n        y = F.relu(y)\n        return y\n\n# \u5b9a\u4e49ResNet\u6a21\u578b\nclass ResNet(paddle.nn.Layer):\n    def __init__(self, layers=50, class_dim=1):\n        \"\"\"\n        \n        layers, \u7f51\u7edc\u5c42\u6570\uff0c\u53ef\u4ee5\u662f50, 101\u6216\u8005152\n        class_dim\uff0c\u5206\u7c7b\u6807\u7b7e\u7684\u7c7b\u522b\u6570\n        \"\"\"\n        super(ResNet, self).__init__()\n        self.layers = layers\n        supported_layers = &#91;50, 101, 152]\n        assert layers in supported_layers, \\\n            \"supported layers are {} but input layer is {}\".format(supported_layers, layers)\n\n        if layers == 50:\n            #ResNet50\u5305\u542b\u591a\u4e2a\u6a21\u5757\uff0c\u5176\u4e2d\u7b2c2\u5230\u7b2c5\u4e2a\u6a21\u5757\u5206\u522b\u5305\u542b3\u30014\u30016\u30013\u4e2a\u6b8b\u5dee\u5757\n            depth = &#91;3, 4, 6, 3]\n        elif layers == 101:\n            #ResNet101\u5305\u542b\u591a\u4e2a\u6a21\u5757\uff0c\u5176\u4e2d\u7b2c2\u5230\u7b2c5\u4e2a\u6a21\u5757\u5206\u522b\u5305\u542b3\u30014\u300123\u30013\u4e2a\u6b8b\u5dee\u5757\n            depth = &#91;3, 4, 23, 3]\n        elif layers == 152:\n            #ResNet152\u5305\u542b\u591a\u4e2a\u6a21\u5757\uff0c\u5176\u4e2d\u7b2c2\u5230\u7b2c5\u4e2a\u6a21\u5757\u5206\u522b\u5305\u542b3\u30018\u300136\u30013\u4e2a\u6b8b\u5dee\u5757\n            depth = &#91;3, 8, 36, 3]\n        \n        # \u6b8b\u5dee\u5757\u4e2d\u4f7f\u7528\u5230\u7684\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\n        num_filters = &#91;64, 128, 256, 512]\n\n        # ResNet\u7684\u7b2c\u4e00\u4e2a\u6a21\u5757\uff0c\u5305\u542b1\u4e2a7x7\u5377\u79ef\uff0c\u540e\u9762\u8ddf\u77401\u4e2a\u6700\u5927\u6c60\u5316\u5c42\n        self.conv = ConvBNLayer(\n            num_channels=3,\n            num_filters=64,\n            filter_size=7,\n            stride=2,\n            act='relu')\n        self.pool2d_max = nn.MaxPool2D(\n            kernel_size=3,\n            stride=2,\n            padding=1)\n\n        # ResNet\u7684\u7b2c\u4e8c\u5230\u7b2c\u4e94\u4e2a\u6a21\u5757c2\u3001c3\u3001c4\u3001c5\n        self.bottleneck_block_list = &#91;]\n        num_channels = 64\n        for block in range(len(depth)):\n            shortcut = False\n            for i in range(depth&#91;block]):\n                bottleneck_block = self.add_sublayer(\n                    'bb_%d_%d' % (block, i),\n                    BottleneckBlock(\n                        num_channels=num_channels,\n                        num_filters=num_filters&#91;block],\n                        stride=2 if i == 0 and block != 0 else 1, # c3\u3001c4\u3001c5\u5c06\u4f1a\u5728\u7b2c\u4e00\u4e2a\u6b8b\u5dee\u5757\u4f7f\u7528stride=2\uff1b\u5176\u4f59\u6240\u6709\u6b8b\u5dee\u5757stride=1\n                        shortcut=shortcut))\n                num_channels = bottleneck_block._num_channels_out\n                self.bottleneck_block_list.append(bottleneck_block)\n                shortcut = True\n\n        # \u5728c5\u7684\u8f93\u51fa\u7279\u5f81\u56fe\u4e0a\u4f7f\u7528\u5168\u5c40\u6c60\u5316\n        self.pool2d_avg = paddle.nn.AdaptiveAvgPool2D(output_size=1)\n\n        # stdv\u7528\u6765\u4f5c\u4e3a\u5168\u8fde\u63a5\u5c42\u968f\u673a\u521d\u59cb\u5316\u53c2\u6570\u7684\u65b9\u5dee\n        import math\n        stdv = 1.0 \/ math.sqrt(2048 * 1.0)\n        \n        # \u521b\u5efa\u5168\u8fde\u63a5\u5c42\uff0c\u8f93\u51fa\u5927\u5c0f\u4e3a\u7c7b\u522b\u6570\u76ee\uff0c\u7ecf\u8fc7\u6b8b\u5dee\u7f51\u7edc\u7684\u5377\u79ef\u548c\u5168\u5c40\u6c60\u5316\u540e\uff0c\n        # \u5377\u79ef\u7279\u5f81\u7684\u7ef4\u5ea6\u662f&#91;B,2048,1,1]\uff0c\u6545\u6700\u540e\u4e00\u5c42\u5168\u8fde\u63a5\u7684\u8f93\u5165\u7ef4\u5ea6\u662f2048\n        self.out = nn.Linear(in_features=2048, out_features=class_dim,\n                      weight_attr=paddle.ParamAttr(\n                          initializer=paddle.nn.initializer.Uniform(-stdv, stdv)))\n\n    def forward(self, inputs):\n        y = self.conv(inputs)\n        y = self.pool2d_max(y)\n        for bottleneck_block in self.bottleneck_block_list:\n            y = bottleneck_block(y)\n        y = self.pool2d_avg(y)\n        y = paddle.reshape(y, &#91;y.shape&#91;0], -1])\n        y = self.out(y)\n        return y\n\n# \u521b\u5efa\u6a21\u578b\nmodel = ResNet()\n# \u5b9a\u4e49\u4f18\u5316\u5668\nopt = paddle.optimizer.Momentum(learning_rate=0.001, momentum=0.9, parameters=model.parameters(), weight_decay=0.001)\n# \u542f\u52a8\u8bad\u7ec3\u8fc7\u7a0b\ntrain_pm(model, opt)<\/code><\/pre>\n\n\n\n<p>\u7814\u7a76\u4e00\u4e0b\u5b9e\u73b0\u7684\u4ee3\u7801\uff0c\u9996\u5148\u5b9a\u4e49\u4e86<strong>ConvBNLayer<\/strong>\uff0c\u5176\u5b9e\u5c31\u662f\u5377\u79ef\u6838\u52a0\u4e86\u4e2a\u5f52\u4e00\u5316<\/p>\n\n\n\n<p>\u4e4b\u540e\u5b9a\u4e49<strong>BottleneckBlock<\/strong>\uff0c\u770b\u4ee3\u7801\u5c31\u660e\u767d\u4ec0\u4e48\u53eb\u6b8b\u5dee\u4e86\uff0c\u8f93\u51fa\u662f\u5377\u79ef\u7684\u8f93\u51fa\u4e0e\u8f93\u5165\u76f8\u52a0\uff08\u6ce8\u610f\u5982\u679csize\u4e0d\u4e00\u6837\u8981\u52a0short\u5c421\u00d71\u5377\u79ef\u589e\u52a0\u56db\u500d\u901a\u9053\u6570\uff0c\u5c31\u662f\u6bcf\u4e00\u6a21\u5757\u7b2c\u4e00\u4e2a\u6b8b\u5dee\u5757\uff0c\u8f93\u51fa\u901a\u9053\u4f1a\u56db\u500d\uff0c\u4e4b\u540e\u6bcf\u4e2a\u6b8b\u5dee\u5757\u8f93\u5165\u8f93\u51fa\u901a\u9053\u6570\u4e0d\u53d8\uff09<\/p>\n\n\n\n<p>\u6bcf\u4e00<strong>BottleneckBlock<\/strong>\u6a21\u5757\u7684channels\u5c31\u662fnum_filter\u7684\u56db\u500d\uff0c\u7b2c\u4e00\u6a21\u5757stride\u90fd\u662f1\u6240\u4ee5size\u4e0d\u53d8\uff0c\u5176\u540e\u6bcf\u6a21\u5757\u7684\u7b2c\u4e00\u4e2a\u6b8b\u5dee\u5757\u7684stride=2\uff0csize\u4f1a\u51cf\u534a<\/p>\n\n\n\n<p>\u6253\u5370\u4e00\u4e0b\u6bcf\u5c42\u8f93\u51fa\u7684size\u4e0e\u6211\u4eec\u624b\u52a8\u8ba1\u7b97\u7684\u5bf9\u6bd4\u4e00\u4e0b\uff0c\u6ca1\u6709\u9519\u8bef\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>x = np.random.randn(*&#91;1,3,224,224])\nx = x.astype('float32')\nx = paddle.to_tensor(x)\nm = ResNet()\ny = m.conv(x)\nprint(y.shape)\ny = m.pool2d_max(y)\nx = np.random.randn(*&#91;1,3,224,224])\nx = x.astype('float32')\nx = paddle.to_tensor(x)\nm = ResNet()\ny = m.conv(x)\nprint(y.shape)\ny = m.pool2d_max(y)\nfor bottleneck_block in m.bottleneck_block_list:\n    y = bottleneck_block(y)\n    print(y.shape)\ny = m.pool2d_avg(y)\nprint(y.shape)\ny = paddle.reshape(y, &#91;y.shape&#91;0], -1])\nprint(y.shape)\ny = m.out(y)\nprint(y.shape)\n\"\"\"-----------------------\u8f93\u51fa------------------------\"\"\"\n&#91;1, 64, 112, 112]\n&#91;1, 256, 56, 56]\n&#91;1, 256, 56, 56]\n&#91;1, 256, 56, 56]\n&#91;1, 512, 28, 28]\n&#91;1, 512, 28, 28]\n&#91;1, 512, 28, 28]\n&#91;1, 512, 28, 28]\n&#91;1, 1024, 14, 14]\n&#91;1, 1024, 14, 14]\n&#91;1, 1024, 14, 14]\n&#91;1, 1024, 14, 14]\n&#91;1, 1024, 14, 14]\n&#91;1, 1024, 14, 14]\n&#91;1, 2048, 7, 7]\n&#91;1, 2048, 7, 7]\n&#91;1, 2048, 7, 7]\n&#91;1, 2048, 1, 1]\n&#91;1, 2048]\n&#91;1, 1]<\/code><\/pre>\n\n\n\n<p>\u63a5\u7740\u4f7f\u7528torch\u6765\u5b9e\u73b0<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># -*- coding:utf-8 -*-\n\n# ResNet\u6a21\u578b\u4ee3\u7801\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n# ResNet\u4e2d\u4f7f\u7528\u4e86BatchNorm\u5c42\uff0c\u5728\u5377\u79ef\u5c42\u7684\u540e\u9762\u52a0\u4e0aBatchNorm\u4ee5\u63d0\u5347\u6570\u503c\u7a33\u5b9a\u6027\n# \u5b9a\u4e49\u5377\u79ef\u6279\u5f52\u4e00\u5316\u5757\nclass ConvBNLayer(torch.nn.Module):\n    def __init__(self,\n                 num_channels,\n                 num_filters,\n                 filter_size,\n                 stride=1,\n                 groups=1,\n                 act=None):\n       \n        \"\"\"\n        num_channels, \u5377\u79ef\u5c42\u7684\u8f93\u5165\u901a\u9053\u6570\n        num_filters, \u5377\u79ef\u5c42\u7684\u8f93\u51fa\u901a\u9053\u6570\n        stride, \u5377\u79ef\u5c42\u7684\u6b65\u5e45\n        groups, \u5206\u7ec4\u5377\u79ef\u7684\u7ec4\u6570\uff0c\u9ed8\u8ba4groups=1\u4e0d\u4f7f\u7528\u5206\u7ec4\u5377\u79ef\n        \"\"\"\n        super(ConvBNLayer, self).__init__()\n\n        # \u521b\u5efa\u5377\u79ef\u5c42\n        self._conv = nn.Conv2d(\n            in_channels=num_channels,\n            out_channels=num_filters,\n            kernel_size=filter_size,\n            stride=stride,\n            padding=(filter_size - 1) \/\/ 2,\n            groups=groups,).cuda()\n\n        # \u521b\u5efaBatchNorm\u5c42\n        self._batch_norm = torch.nn.BatchNorm2d(num_filters).cuda()\n        \n        self.act = act\n\n    def forward(self, inputs):\n        y = self._conv(inputs)\n        y = self._batch_norm(y)\n        if self.act == 'leaky':\n            y = F.leaky_relu(y, negative_slope=0.1)\n        elif self.act == 'relu':\n            y = F.relu(y)\n        return y\n\n# \u5b9a\u4e49\u6b8b\u5dee\u5757\n# \u6bcf\u4e2a\u6b8b\u5dee\u5757\u4f1a\u5bf9\u8f93\u5165\u56fe\u7247\u505a\u4e09\u6b21\u5377\u79ef\uff0c\u7136\u540e\u8ddf\u8f93\u5165\u56fe\u7247\u8fdb\u884c\u77ed\u63a5\n# \u5982\u679c\u6b8b\u5dee\u5757\u4e2d\u7b2c\u4e09\u6b21\u5377\u79ef\u8f93\u51fa\u7279\u5f81\u56fe\u7684\u5f62\u72b6\u4e0e\u8f93\u5165\u4e0d\u4e00\u81f4\uff0c\u5219\u5bf9\u8f93\u5165\u56fe\u7247\u505a1x1\u5377\u79ef\uff0c\u5c06\u5176\u8f93\u51fa\u5f62\u72b6\u8c03\u6574\u6210\u4e00\u81f4\nclass BottleneckBlock(torch.nn.Module):\n    def __init__(self,\n                 num_channels,\n                 num_filters,\n                 stride,\n                 shortcut=True):\n        super(BottleneckBlock, self).__init__()\n        # \u521b\u5efa\u7b2c\u4e00\u4e2a\u5377\u79ef\u5c42 1x1\n        self.conv0 = ConvBNLayer(\n            num_channels=num_channels,\n            num_filters=num_filters,\n            filter_size=1,\n            act='relu')\n        # \u521b\u5efa\u7b2c\u4e8c\u4e2a\u5377\u79ef\u5c42 3x3\n        self.conv1 = ConvBNLayer(\n            num_channels=num_filters,\n            num_filters=num_filters,\n            filter_size=3,\n            stride=stride,\n            act='relu')\n        # \u521b\u5efa\u7b2c\u4e09\u4e2a\u5377\u79ef 1x1\uff0c\u4f46\u8f93\u51fa\u901a\u9053\u6570\u4e58\u4ee54\n        self.conv2 = ConvBNLayer(\n            num_channels=num_filters,\n            num_filters=num_filters * 4,\n            filter_size=1,\n            act=None)\n\n        # \u5982\u679cconv2\u7684\u8f93\u51fa\u8ddf\u6b64\u6b8b\u5dee\u5757\u7684\u8f93\u5165\u6570\u636e\u5f62\u72b6\u4e00\u81f4\uff0c\u5219shortcut=True\n        # \u5426\u5219shortcut = False\uff0c\u6dfb\u52a01\u4e2a1x1\u7684\u5377\u79ef\u4f5c\u7528\u5728\u8f93\u5165\u6570\u636e\u4e0a\uff0c\u4f7f\u5176\u5f62\u72b6\u53d8\u6210\u8ddfconv2\u4e00\u81f4\n        if not shortcut:\n            self.short = ConvBNLayer(\n                num_channels=num_channels,\n                num_filters=num_filters * 4,\n                filter_size=1,\n                stride=stride)\n\n        self.shortcut = shortcut\n\n        self._num_channels_out = num_filters * 4\n\n    def forward(self, inputs):\n        y = self.conv0(inputs)\n        conv1 = self.conv1(y)\n        conv2 = self.conv2(conv1)\n\n        # \u5982\u679cshortcut=True\uff0c\u76f4\u63a5\u5c06inputs\u8ddfconv2\u7684\u8f93\u51fa\u76f8\u52a0\n        # \u5426\u5219\u9700\u8981\u5bf9inputs\u8fdb\u884c\u4e00\u6b21\u5377\u79ef\uff0c\u5c06\u5f62\u72b6\u8c03\u6574\u6210\u8ddfconv2\u8f93\u51fa\u4e00\u81f4\n        if self.shortcut:\n            short = inputs\n        else:\n            short = self.short(inputs)\n\n        y = torch.add(short, conv2)\n        y = F.relu(y)\n        return y\n\n# \u5b9a\u4e49ResNet\u6a21\u578b\nclass ResNet(torch.nn.Module):\n    def __init__(self, layers=50, class_dim=1):\n        \"\"\"\n        \n        layers, \u7f51\u7edc\u5c42\u6570\uff0c\u53ef\u4ee5\u662f50, 101\u6216\u8005152\n        class_dim\uff0c\u5206\u7c7b\u6807\u7b7e\u7684\u7c7b\u522b\u6570\n        \"\"\"\n        super(ResNet, self).__init__()\n        self.layers = layers\n        supported_layers = &#91;50, 101, 152]\n        assert layers in supported_layers, \\\n            \"supported layers are {} but input layer is {}\".format(supported_layers, layers)\n\n        if layers == 50:\n            #ResNet50\u5305\u542b\u591a\u4e2a\u6a21\u5757\uff0c\u5176\u4e2d\u7b2c2\u5230\u7b2c5\u4e2a\u6a21\u5757\u5206\u522b\u5305\u542b3\u30014\u30016\u30013\u4e2a\u6b8b\u5dee\u5757\n            depth = &#91;3, 4, 6, 3]\n        elif layers == 101:\n            #ResNet101\u5305\u542b\u591a\u4e2a\u6a21\u5757\uff0c\u5176\u4e2d\u7b2c2\u5230\u7b2c5\u4e2a\u6a21\u5757\u5206\u522b\u5305\u542b3\u30014\u300123\u30013\u4e2a\u6b8b\u5dee\u5757\n            depth = &#91;3, 4, 23, 3]\n        elif layers == 152:\n            #ResNet152\u5305\u542b\u591a\u4e2a\u6a21\u5757\uff0c\u5176\u4e2d\u7b2c2\u5230\u7b2c5\u4e2a\u6a21\u5757\u5206\u522b\u5305\u542b3\u30018\u300136\u30013\u4e2a\u6b8b\u5dee\u5757\n            depth = &#91;3, 8, 36, 3]\n        \n        # \u6b8b\u5dee\u5757\u4e2d\u4f7f\u7528\u5230\u7684\u5377\u79ef\u7684\u8f93\u51fa\u901a\u9053\u6570\n        num_filters = &#91;64, 128, 256, 512]\n\n        # ResNet\u7684\u7b2c\u4e00\u4e2a\u6a21\u5757\uff0c\u5305\u542b1\u4e2a7x7\u5377\u79ef\uff0c\u540e\u9762\u8ddf\u77401\u4e2a\u6700\u5927\u6c60\u5316\u5c42\n        self.conv = ConvBNLayer(\n            num_channels=3,\n            num_filters=64,\n            filter_size=7,\n            stride=2,\n            act='relu')\n        self.pool2d_max = nn.MaxPool2d(\n            kernel_size=3,\n            stride=2,\n            padding=1)\n\n        # ResNet\u7684\u7b2c\u4e8c\u5230\u7b2c\u4e94\u4e2a\u6a21\u5757c2\u3001c3\u3001c4\u3001c5\n        self.bottleneck_block_list = &#91;]\n        num_channels = 64\n        for block in range(len(depth)):\n            shortcut = False\n            for i in range(depth&#91;block]):\n                self.bottleneck_block_list.append(\n                    BottleneckBlock(\n                        num_channels=num_channels,\n                        num_filters=num_filters&#91;block],\n                        stride=2 if i == 0 and block != 0 else 1, # c3\u3001c4\u3001c5\u5c06\u4f1a\u5728\u7b2c\u4e00\u4e2a\u6b8b\u5dee\u5757\u4f7f\u7528stride=2\uff1b\u5176\u4f59\u6240\u6709\u6b8b\u5dee\u5757stride=1\n                        shortcut=shortcut))\n                num_channels = self.bottleneck_block_list&#91;-1]._num_channels_out\n                shortcut = True\n\n        # \u5728c5\u7684\u8f93\u51fa\u7279\u5f81\u56fe\u4e0a\u4f7f\u7528\u5168\u5c40\u6c60\u5316\n        self.pool2d_avg = torch.nn.AdaptiveAvgPool2d(output_size=1)\n        self.bn1=nn.BatchNorm1d(2048)\n        \n        self.out = nn.Linear(in_features=2048, out_features=class_dim)\n\n    def forward(self, inputs):\n        y = self.conv(inputs)\n        y = self.pool2d_max(y)\n        for bottleneck_block in self.bottleneck_block_list:\n            y = bottleneck_block(y)\n        y = self.pool2d_avg(y)\n        y = torch.reshape(y, &#91;y.shape&#91;0], -1])\n        y = self.bn1(y)\n        y = self.out(y)\n        return y<\/code><\/pre>\n\n\n\n<p>\u5b9e\u73b0\u7684\u65f6\u5019\u8fd8\u662f\u9047\u5230\u4e0d\u5c11\u95ee\u9898\uff0capi\u4e0d\u4e00\u6837\uff0ctorch\u6ca1\u6709add_sublayer\u8fd9\u4e2aapi\uff0c\u6211\u5c31\u76f4\u63a5\u7528append\u4ee3\u66ff\u4e86\uff0c\u53e6\u5916torch\u4e0d\u5141\u8bb8y = F.relu(x=y)\u8fd9\u79cd\u62ec\u53f7\u91cc\u8d4b\u503c\u7684\u5199\u6cd5\uff0c\u9664\u6b64\u4e4b\u5916\u5b9a\u4e49\u7684conv\u548cbatchnorm\u5c42\u90fd\u8981\u52a0.cuda()\u3002\u6700\u540epaddle\u5e94\u8be5\u662f\u7ebf\u6027\u5c42\u7684\u53c2\u6570\u6b63\u5219\u5316\uff0ctorch\u7684\u5b9e\u73b0\u8fd8\u6ca1\u641e\u660e\u767d\uff0c\u5199\u7684\u65f6\u5019\u4ee5\u4e3a\u662fbatchnorm\uff0c\u4f46\u60f3\u4e86\u60f3\u4e0d\u4e00\u6837\uff0cbatchnorm\u8fd8\u662f\u628a\u8fd9\u4e2abatch\u5230\u8fd9\u5c42\u7684\u8f93\u5165\u7ed9\u5f52\u4e00\u5316\u7684\uff0c\u4e0d\u662f\u5bf9\u53c2\u6570\u8fdb\u884c\u9650\u5236\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"576\" src=\"http:\/\/www.gislxz.com\/wp-content\/uploads\/2022\/08\/IMG_20220821_112355-1024x576.png\" alt=\"\" class=\"wp-image-433\" srcset=\"https:\/\/www.gislxz.com\/wp-content\/uploads\/2022\/08\/IMG_20220821_112355-1024x576.png 1024w, https:\/\/www.gislxz.com\/wp-content\/uploads\/2022\/08\/IMG_20220821_112355-300x169.png 300w, https:\/\/www.gislxz.com\/wp-content\/uploads\/2022\/08\/IMG_20220821_112355-768x432.png 768w, https:\/\/www.gislxz.com\/wp-content\/uploads\/2022\/08\/IMG_20220821_112355-1536x864.png 1536w, https:\/\/www.gislxz.com\/wp-content\/uploads\/2022\/08\/IMG_20220821_112355-2048x1152.png 2048w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption>\u751c\u751c\u751c\uff0c\u592a\u751c\u5566<\/figcaption><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>VGG\uff0cGoogLeNet\u4ee5\u53caResNet<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[21],"tags":[],"class_list":["post-412","post","type-post","status-publish","format-standard","hentry","category-21"],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/posts\/412","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/comments?post=412"}],"version-history":[{"count":8,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/posts\/412\/revisions"}],"predecessor-version":[{"id":439,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/posts\/412\/revisions\/439"}],"wp:attachment":[{"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/media?parent=412"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/categories?post=412"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.gislxz.com\/index.php\/wp-json\/wp\/v2\/tags?post=412"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}