卷积神经网络基础.ipynb

{"cells":[{"cell_type":"markdown","metadata":{"graffitiCellId":"id_3dafypt","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"8C5B8FD488F74207A409DC379EAAFD63","mdEditEnable":false},"source":"# 卷积神经网络基础\n\n本节我们介绍卷积神经网络的基础概念，主要是卷积层和池化层，并解释填充、步幅、输入通道和输出通道的含义。"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_hm819qn","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"254FA5A2E556460089490994E31604D0","mdEditEnable":false},"source":"## 二维卷积层\n\n本节介绍的是最常见的二维卷积层，常用于处理图像数据。"},{"attachments":{"%E5%9B%BE%E7%89%87.png":{"image/png":"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"}},"cell_type":"markdown","metadata":{"graffitiCellId":"id_1w29iz7","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"839259A7CAD349A7B2CE930E14F9E7B5","mdEditEnable":false},"source":"### 二维互相关运算\n\n二维互相关（cross-correlation）运算的输入是一个二维输入数组和一个二维核（kernel）数组，输出也是一个二维数组，其中核数组通常称为卷积核或过滤器（filter）。卷积核的尺寸通常小于输入数组，卷积核在输入数组上滑动，在每个位置上，卷积核与该位置处的输入子数组按元素相乘并求和，得到输出数组中相应位置的元素。图1展示了一个互相关运算的例子，阴影部分分别是输入的第一个计算区域、核数组以及对应的输出。\n\n![Image Name](https://cdn.kesci.com/upload/image/q5nfdbhcw5.png?imageView2/0/w/640/h/640)\n图1 二维互相关运算\n\n下面我们用`corr2d`函数实现二维互相关运算，它接受输入数组`X`与核数组`K`，并输出数组`Y`。"},{"cell_type":"code","execution_count":2,"metadata":{"graffitiCellId":"id_nculyfq","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"27AD46C637CE4BF88964BE3C22E6D4DB","collapsed":false,"scrolled":false},"outputs":[],"source":"import torch \nimport torch.nn as nn\n\ndef corr2d(X, K):\n    H, W = X.shape\n    h, w = K.shape\n    Y = torch.zeros(H - h + 1, W - w + 1)\n    for i in range(Y.shape[0]):\n        for j in range(Y.shape[1]):\n            Y[i, j] = (X[i: i + h, j: j + w] * K).sum()\n    return Y"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_nochl8t","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"12E4EDD4780B4ED3BF46FC98E0671987","mdEditEnable":false},"source":"构造上图中的输入数组`X`、核数组`K`来验证二维互相关运算的输出。"},{"cell_type":"code","execution_count":2,"metadata":{"graffitiCellId":"id_ou9gykf","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"D6B5388CE30B4280A8EAC9C1E6160CED","collapsed":false,"scrolled":false},"outputs":[{"output_type":"stream","text":"tensor([[19., 25.],\n        [37., 43.]])\n","name":"stdout"}],"source":"X = torch.tensor([[0, 1, 2], [3, 4, 5], [6, 7, 8]])\nK = torch.tensor([[0, 1], [2, 3]])\nY = corr2d(X, K)\nprint(Y)"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_kyrga1i","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"A667995033B54B998459E0CC127B6D96","mdEditEnable":false},"source":"### 二维卷积层\n\n二维卷积层将输入和卷积核做互相关运算，并加上一个标量偏置来得到输出。卷积层的模型参数包括卷积核和标量偏置。"},{"cell_type":"code","execution_count":3,"metadata":{"graffitiCellId":"id_wqfawao","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"02665E066AC949A2985642A92FC6352A","collapsed":false,"scrolled":false},"outputs":[],"source":"class Conv2D(nn.Module):\n    def __init__(self, kernel_size):\n        super(Conv2D, self).__init__()\n        self.weight = nn.Parameter(torch.randn(kernel_size))\n        self.bias = nn.Parameter(torch.randn(1))\n\n    def forward(self, x):\n        return corr2d(x, self.weight) + self.bias"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_ywvx25d","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"C8662D63813F4E5DA12ABFE199E36F24","mdEditEnable":false},"source":"下面我们看一个例子，我们构造一张$6 \\times 8$的图像，中间4列为黑（0），其余为白（1），希望检测到颜色边缘。我们的标签是一个$6 \\times 7$的二维数组，第2列是1（从1到0的边缘），第6列是-1（从0到1的边缘）。\n"},{"cell_type":"code","execution_count":4,"metadata":{"graffitiCellId":"id_w4j38om","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"F7CC03E43D714175ADD7D490EE3BC0CB","collapsed":false,"scrolled":false},"outputs":[{"output_type":"stream","text":"tensor([[1., 1., 0., 0., 0., 0., 1., 1.],\n        [1., 1., 0., 0., 0., 0., 1., 1.],\n        [1., 1., 0., 0., 0., 0., 1., 1.],\n        [1., 1., 0., 0., 0., 0., 1., 1.],\n        [1., 1., 0., 0., 0., 0., 1., 1.],\n        [1., 1., 0., 0., 0., 0., 1., 1.]])\ntensor([[ 0.,  1.,  0.,  0.,  0., -1.,  0.],\n        [ 0.,  1.,  0.,  0.,  0., -1.,  0.],\n        [ 0.,  1.,  0.,  0.,  0., -1.,  0.],\n        [ 0.,  1.,  0.,  0.,  0., -1.,  0.],\n        [ 0.,  1.,  0.,  0.,  0., -1.,  0.],\n        [ 0.,  1.,  0.,  0.,  0., -1.,  0.]])\n","name":"stdout"}],"source":"X = torch.ones(6, 8)\nY = torch.zeros(6, 7)\nX[:, 2: 6] = 0\nY[:, 1] = 1\nY[:, 5] = -1\nprint(X)\nprint(Y)"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_re55xx5","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"03689A0772F94ABBBE066101211F6299","mdEditEnable":false},"source":"我们希望学习一个$1 \\times 2$卷积层，通过卷积层来检测颜色边缘。\n"},{"cell_type":"code","execution_count":5,"metadata":{"graffitiCellId":"id_d05ejpd","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"1D790E2D3B7E4835994C7198A830DA9D","collapsed":false,"scrolled":false},"outputs":[{"output_type":"stream","text":"Step 5, loss 4.569\nStep 10, loss 0.949\nStep 15, loss 0.228\nStep 20, loss 0.060\nStep 25, loss 0.016\nStep 30, loss 0.004\ntensor([[ 1.0161, -1.0177]])\ntensor([0.0009])\n","name":"stdout"}],"source":"conv2d = Conv2D(kernel_size=(1, 2))\nstep = 30\nlr = 0.01\nfor i in range(step):\n    Y_hat = conv2d(X)\n    l = ((Y_hat - Y) ** 2).sum()\n    l.backward()\n    # 梯度下降\n    conv2d.weight.data -= lr * conv2d.weight.grad\n    conv2d.bias.data -= lr * conv2d.bias.grad\n    \n    # 梯度清零\n    conv2d.weight.grad.zero_()\n    conv2d.bias.grad.zero_()\n    if (i + 1) % 5 == 0:\n        print('Step %d, loss %.3f' % (i + 1, l.item()))\n        \nprint(conv2d.weight.data)\nprint(conv2d.bias.data)"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_oytmkp4","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"6D2014DD040C4AE8928E8D2E8B0E2B27","mdEditEnable":false},"source":"### 互相关运算与卷积运算\n\n卷积层得名于卷积运算，但卷积层中用到的并非卷积运算而是互相关运算。我们将核数组上下翻转、左右翻转，再与输入数组做互相关运算，这一过程就是卷积运算。由于卷积层的核数组是可学习的，所以使用互相关运算与使用卷积运算并无本质区别。"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_iv5rdyk","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"232D9ADECF97429A802C60C0000EF53B","mdEditEnable":false},"source":"### 特征图与感受野\n\n二维卷积层输出的二维数组可以看作是输入在空间维度（宽和高）上某一级的表征，也叫特征图（feature map）。影响元素$x$的前向计算的所有可能输入区域（可能大于输入的实际尺寸）叫做$x$的感受野（receptive field）。\n\n以图1为例，输入中阴影部分的四个元素是输出中阴影部分元素的感受野。我们将图中形状为$2 \\times 2$的输出记为$Y$，将$Y$与另一个形状为$2 \\times 2$的核数组做互相关运算，输出单个元素$z$。那么，$z$在$Y$上的感受野包括$Y$的全部四个元素，在输入上的感受野包括其中全部9个元素。可见，我们可以通过更深的卷积神经网络使特征图中单个元素的感受野变得更加广阔，从而捕捉输入上更大尺寸的特征。"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_adfhv52","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"A4CE7D97E9744A138790A9700063083C","mdEditEnable":false},"source":"## 填充和步幅\n\n我们介绍卷积层的两个超参数，即填充和步幅，它们可以对给定形状的输入和卷积核改变输出形状。"},{"attachments":{"%E5%9B%BE%E7%89%87.png":{"image/png":"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"}},"cell_type":"markdown","metadata":{"graffitiCellId":"id_oquq26z","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"6199F060EBF74BB9894EEDE4F3F845B5","mdEditEnable":false},"source":"### 填充\n\n填充（padding）是指在输入高和宽的两侧填充元素（通常是0元素），图2里我们在原输入高和宽的两侧分别添加了值为0的元素。\n\n\n![Image Name](https://cdn.kesci.com/upload/image/q5nfl6ejy4.png?imageView2/0/w/640/h/640)\n\n图2 在输入的高和宽两侧分别填充了0元素的二维互相关计算\n\n如果原输入的高和宽是$n_h$和$n_w$，卷积核的高和宽是$k_h$和$k_w$，在高的两侧一共填充$p_h$行，在宽的两侧一共填充$p_w$列，则输出形状为：\n\n\n$$\n\n(n_h+p_h-k_h+1)\\times(n_w+p_w-k_w+1)\n\n$$\n\n\n我们在卷积神经网络中使用奇数高宽的核，比如$3 \\times 3$，$5 \\times 5$的卷积核，对于高度（或宽度）为大小为$2 k + 1$的核，令步幅为1，在高（或宽）两侧选择大小为$k$的填充，便可保持输入与输出尺寸相同。"},{"attachments":{"%E5%9B%BE%E7%89%87.png":{"image/png":"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"}},"cell_type":"markdown","metadata":{"graffitiCellId":"id_co3p9ym","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"C31F0DD02A144FC29F738F6A1389E372","mdEditEnable":false},"source":"### 步幅\n\n在互相关运算中，卷积核在输入数组上滑动，每次滑动的行数与列数即是步幅（stride）。此前我们使用的步幅都是1，图3展示了在高上步幅为3、在宽上步幅为2的二维互相关运算。\n\n\n![Image Name](https://cdn.kesci.com/upload/image/q5nflohnqg.png?imageView2/0/w/640/h/640)\n\n图3 高和宽上步幅分别为3和2的二维互相关运算\n\n一般来说，当高上步幅为$s_h$，宽上步幅为$s_w$时，输出形状为：\n\n\n$$\n\n\\lfloor(n_h+p_h-k_h+s_h)/s_h\\rfloor \\times \\lfloor(n_w+p_w-k_w+s_w)/s_w\\rfloor\n\n$$\n\n\n如果$p_h=k_h-1$，$p_w=k_w-1$，那么输出形状将简化为$\\lfloor(n_h+s_h-1)/s_h\\rfloor \\times \\lfloor(n_w+s_w-1)/s_w\\rfloor$。更进一步，如果输入的高和宽能分别被高和宽上的步幅整除，那么输出形状将是$(n_h / s_h) \\times (n_w/s_w)$。\n\n当$p_h = p_w = p$时，我们称填充为$p$；当$s_h = s_w = s$时，我们称步幅为$s$。"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_1idmra0","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"D6F0F17D97B54D4E843A8792A8B2D568","mdEditEnable":false},"source":"## 多输入通道和多输出通道\n\n之前的输入和输出都是二维数组，但真实数据的维度经常更高。例如，彩色图像在高和宽2个维度外还有RGB（红、绿、蓝）3个颜色通道。假设彩色图像的高和宽分别是$h$和$w$（像素），那么它可以表示为一个$3 \\times h \\times w$的多维数组，我们将大小为3的这一维称为通道（channel）维。"},{"attachments":{"%E5%9B%BE%E7%89%87.png":{"image/png":"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"}},"cell_type":"markdown","metadata":{"graffitiCellId":"id_n6snjjh","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"9D38698675574834BCF819E18FAA6F21","mdEditEnable":false},"source":"### 多输入通道\n\n卷积层的输入可以包含多个通道，图4展示了一个含2个输入通道的二维互相关计算的例子。\n\n\n![Image Name](https://cdn.kesci.com/upload/image/q5nfmdnwbq.png?imageView2/0/w/640/h/640)\n\n图4 含2个输入通道的互相关计算\n\n假设输入数据的通道数为$c_i$，卷积核形状为$k_h\\times k_w$，我们为每个输入通道各分配一个形状为$k_h\\times k_w$的核数组，将$c_i$个互相关运算的二维输出按通道相加，得到一个二维数组作为输出。我们把$c_i$个核数组在通道维上连结，即得到一个形状为$c_i\\times k_h\\times k_w$的卷积核。"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_xpcj8zq","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"B43CBA5D4D854AB68291F5D7CB1B5FA5","mdEditEnable":false},"source":"### 多输出通道\n\n卷积层的输出也可以包含多个通道，设卷积核输入通道数和输出通道数分别为$c_i$和$c_o$，高和宽分别为$k_h$和$k_w$。如果希望得到含多个通道的输出，我们可以为每个输出通道分别创建形状为$c_i\\times k_h\\times k_w$的核数组，将它们在输出通道维上连结，卷积核的形状即$c_o\\times c_i\\times k_h\\times k_w$。\n\n对于输出通道的卷积核，我们提供这样一种理解，一个$c_i \\times k_h \\times k_w$的核数组可以提取某种局部特征，但是输入可能具有相当丰富的特征，我们需要有多个这样的$c_i \\times k_h \\times k_w$的核数组，不同的核数组提取的是不同的特征。"},{"attachments":{"%E5%9B%BE%E7%89%87.png":{"image/png":"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"}},"cell_type":"markdown","metadata":{"graffitiCellId":"id_t861gfe","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"2EDA6611AFD148D291D2B9E7AB8C1B50","mdEditEnable":false},"source":"### 1x1卷积层\n\n最后讨论形状为$1 \\times 1$的卷积核，我们通常称这样的卷积运算为$1 \\times 1$卷积，称包含这种卷积核的卷积层为$1 \\times 1$卷积层。图5展示了使用输入通道数为3、输出通道数为2的$1\\times 1$卷积核的互相关计算。\n\n\n![Image Name](https://cdn.kesci.com/upload/image/q5nfmq980r.png?imageView2/0/w/640/h/640)\n\n图5 1x1卷积核的互相关计算。输入和输出具有相同的高和宽\n\n$1 \\times 1$卷积核可在不改变高宽的情况下，调整通道数。$1 \\times 1$卷积核不识别高和宽维度上相邻元素构成的模式，其主要计算发生在通道维上。假设我们将通道维当作特征维，将高和宽维度上的元素当成数据样本，那么$1\\times 1$卷积层的作用与全连接层等价。"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_6hb1wnk","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"DFB80D1A07A5496194BA81030D54E45E","mdEditEnable":false},"source":"## 卷积层与全连接层的对比\n\n二维卷积层经常用于处理图像，与此前的全连接层相比，它主要有两个优势：\n\n一是全连接层把图像展平成一个向量，在输入图像上相邻的元素可能因为展平操作不再相邻，网络难以捕捉局部信息。而卷积层的设计，天然地具有提取局部信息的能力。\n\n二是卷积层的参数量更少。不考虑偏置的情况下，一个形状为$(c_i, c_o, h, w)$的卷积核的参数量是$c_i \\times c_o \\times h \\times w$，与输入图像的宽高无关。假如一个卷积层的输入和输出形状分别是$(c_1, h_1, w_1)$和$(c_2, h_2, w_2)$，如果要用全连接层进行连接，参数数量就是$c_1 \\times c_2 \\times h_1 \\times w_1 \\times h_2 \\times w_2$。使用卷积层可以以较少的参数数量来处理更大的图像。"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_qhv3vji","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"3B6B64EC47384FD78A0CF2BD157ABDC5","mdEditEnable":false},"source":"## 卷积层的简洁实现\n\n我们使用Pytorch中的`nn.Conv2d`类来实现二维卷积层，主要关注以下几个构造函数参数：\n\n* `in_channels` (python:int) – Number of channels in the input imag\n* `out_channels` (python:int) – Number of channels produced by the convolution\n* `kernel_size` (python:int or tuple) – Size of the convolving kernel\n* `stride` (python:int or tuple, optional) – Stride of the convolution. Default: 1\n* `padding` (python:int or tuple, optional) – Zero-padding added to both sides of the input. Default: 0\n* `bias` (bool, optional) – If True, adds a learnable bias to the output. Default: True\n\n`forward`函数的参数为一个四维张量，形状为$(N, C_{in}, H_{in}, W_{in})$，返回值也是一个四维张量，形状为$(N, C_{out}, H_{out}, W_{out})$，其中$N$是批量大小，$C, H, W$分别表示通道数、高度、宽度。\n\n代码讲解"},{"cell_type":"code","execution_count":6,"metadata":{"graffitiCellId":"id_50vlm3f","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"8C48648624FC4CB98EB0FC69A020F59C","collapsed":false,"scrolled":false},"outputs":[{"output_type":"stream","text":"torch.Size([4, 2, 3, 5])\nY.shape:  torch.Size([4, 3, 3, 5])\nweight.shape:  torch.Size([3, 2, 3, 5])\nbias.shape:  torch.Size([3])\n","name":"stdout"}],"source":"X = torch.rand(4, 2, 3, 5)\nprint(X.shape)\n\nconv2d = nn.Conv2d(in_channels=2, out_channels=3, kernel_size=(3, 5), stride=1, padding=(1, 2))\nY = conv2d(X)\nprint('Y.shape: ', Y.shape)\nprint('weight.shape: ', conv2d.weight.shape)\nprint('bias.shape: ', conv2d.bias.shape)"},{"attachments":{"%E5%9B%BE%E7%89%87.png":{"image/png":"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"}},"cell_type":"markdown","metadata":{"graffitiCellId":"id_h4crb33","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"C18215EE5FBC43F5867D1F384A2FC328","mdEditEnable":false},"source":"## 池化\n\n### 二维池化层\n\n池化层主要用于缓解卷积层对位置的过度敏感性。同卷积层一样，池化层每次对输入数据的一个固定形状窗口（又称池化窗口）中的元素计算输出，池化层直接计算池化窗口内元素的最大值或者平均值，该运算也分别叫做最大池化或平均池化。图6展示了池化窗口形状为$2\\times 2$的最大池化。\n\n\n![Image Name](https://cdn.kesci.com/upload/image/q5nfob3odo.png?imageView2/0/w/640/h/640)\n\n图6 池化窗口形状为 2 x 2 的最大池化\n\n二维平均池化的工作原理与二维最大池化类似，但将最大运算符替换成平均运算符。池化窗口形状为$p \\times q$的池化层称为$p \\times q$池化层，其中的池化运算叫作$p \\times q$池化。\n\n池化层也可以在输入的高和宽两侧填充并调整窗口的移动步幅来改变输出形状。池化层填充和步幅与卷积层填充和步幅的工作机制一样。\n\n在处理多通道输入数据时，池化层对每个输入通道分别池化，但不会像卷积层那样将各通道的结果按通道相加。这意味着池化层的输出通道数与输入通道数相等。"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_xeq3hyv","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"9AB630787D9B4A93866C9A98A23A3C95","mdEditEnable":false},"source":"### 池化层的简洁实现\n\n我们使用Pytorch中的`nn.MaxPool2d`实现最大池化层，关注以下构造函数参数：\n\n* `kernel_size` – the size of the window to take a max over\n* `stride` – the stride of the window. Default value is kernel_size\n* `padding` – implicit zero padding to be added on both sides\n\n`forward`函数的参数为一个四维张量，形状为$(N, C, H_{in}, W_{in})$，返回值也是一个四维张量，形状为$(N, C, H_{out}, W_{out})$，其中$N$是批量大小，$C, H, W$分别表示通道数、高度、宽度。\n\n代码讲解"},{"cell_type":"code","execution_count":3,"metadata":{"graffitiCellId":"id_euby6h0","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"055AED0960484D6F8B219367AC8542F5","collapsed":false,"scrolled":false},"outputs":[{"output_type":"stream","text":"tensor([[[[ 0.,  1.,  2.,  3.],\n          [ 4.,  5.,  6.,  7.],\n          [ 8.,  9., 10., 11.],\n          [12., 13., 14., 15.]],\n\n         [[16., 17., 18., 19.],\n          [20., 21., 22., 23.],\n          [24., 25., 26., 27.],\n          [28., 29., 30., 31.]]]])\ntensor([[[[ 5.,  6.,  7.,  7.],\n          [13., 14., 15., 15.]],\n\n         [[21., 22., 23., 23.],\n          [29., 30., 31., 31.]]]])\ntorch.Size([1, 2, 2, 4])\n","name":"stdout"}],"source":"X = torch.arange(32, dtype=torch.float32).view(1, 2, 4, 4)\npool2d = nn.MaxPool2d(kernel_size=3, padding=1, stride=(2, 1))\nY = pool2d(X)\nprint(X)\nprint(Y)\nprint(Y.shape)"},{"cell_type":"markdown","metadata":{"graffitiCellId":"id_br4s548","jupyter":{},"tags":[],"slideshow":{"slide_type":"slide"},"id":"E06C633237BF4157872396D0C262310B","mdEditEnable":false},"source":"平均池化层使用的是`nn.AvgPool2d`，使用方法与`nn.MaxPool2d`相同。"}],"metadata":{"kernelspec":{"name":"python3","display_name":"Python 3","language":"python"},"language_info":{"name":"python","version":"3.7.3","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"}},"nbformat":4,"nbformat_minor":4}