Files
2025-12-16 09:23:53 +08:00

527 lines
16 KiB
Plaintext
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
{
"cells": [
{
"cell_type": "markdown",
"id": "3406a2db",
"metadata": {
"origin_pos": 0
},
"source": [
"# 汇聚层\n",
":label:`sec_pooling`\n",
"\n",
"通常当我们处理图像时,我们希望逐渐降低隐藏表示的空间分辨率、聚集信息,这样随着我们在神经网络中层叠的上升,每个神经元对其敏感的感受野(输入)就越大。\n",
"\n",
"而我们的机器学习任务通常会跟全局图像的问题有关(例如,“图像是否包含一只猫呢?”),所以我们最后一层的神经元应该对整个输入的全局敏感。通过逐渐聚合信息,生成越来越粗糙的映射,最终实现学习全局表示的目标,同时将卷积图层的所有优势保留在中间层。\n",
"\n",
"此外,当检测较底层的特征时(例如 :numref:`sec_conv_layer`中所讨论的边缘),我们通常希望这些特征保持某种程度上的平移不变性。例如,如果我们拍摄黑白之间轮廓清晰的图像`X`,并将整个图像向右移动一个像素,即`Z[i, j] = X[i, j + 1]`,则新图像`Z`的输出可能大不相同。而在现实中,随着拍摄角度的移动,任何物体几乎不可能发生在同一像素上。即使用三脚架拍摄一个静止的物体,由于快门的移动而引起的相机振动,可能会使所有物体左右移动一个像素(除了高端相机配备了特殊功能来解决这个问题)。\n",
"\n",
"本节将介绍*汇聚*(pooling)层,它具有双重目的:降低卷积层对位置的敏感性,同时降低对空间降采样表示的敏感性。\n",
"\n",
"## 最大汇聚层和平均汇聚层\n",
"\n",
"与卷积层类似,汇聚层运算符由一个固定形状的窗口组成,该窗口根据其步幅大小在输入的所有区域上滑动,为固定形状窗口(有时称为*汇聚窗口*)遍历的每个位置计算一个输出。\n",
"然而,不同于卷积层中的输入与卷积核之间的互相关计算,汇聚层不包含参数。\n",
"相反,池运算是确定性的,我们通常计算汇聚窗口中所有元素的最大值或平均值。这些操作分别称为*最大汇聚层*(maximum pooling)和*平均汇聚层*average pooling)。\n",
"\n",
"在这两种情况下,与互相关运算符一样,汇聚窗口从输入张量的左上角开始,从左往右、从上往下的在输入张量内滑动。在汇聚窗口到达的每个位置,它计算该窗口中输入子张量的最大值或平均值。计算最大值或平均值是取决于使用了最大汇聚层还是平均汇聚层。\n",
"\n",
"![汇聚窗口形状为 $2\\times 2$ 的最大汇聚层。着色部分是第一个输出元素,以及用于计算这个输出的输入元素: $\\max(0, 1, 3, 4)=4$.](../img/pooling.svg)\n",
":label:`fig_pooling`\n",
"\n",
" :numref:`fig_pooling`中的输出张量的高度为$2$,宽度为$2$。这四个元素为每个汇聚窗口中的最大值:\n",
"\n",
"$$\n",
"\\max(0, 1, 3, 4)=4,\\\\\n",
"\\max(1, 2, 4, 5)=5,\\\\\n",
"\\max(3, 4, 6, 7)=7,\\\\\n",
"\\max(4, 5, 7, 8)=8.\\\\\n",
"$$\n",
"\n",
"汇聚窗口形状为$p \\times q$的汇聚层称为$p \\times q$汇聚层,汇聚操作称为$p \\times q$汇聚。\n",
"\n",
"回到本节开头提到的对象边缘检测示例,现在我们将使用卷积层的输出作为$2\\times 2$最大汇聚的输入。\n",
"设置卷积层输入为`X`,汇聚层输出为`Y`。\n",
"无论`X[i, j]`和`X[i, j + 1]`的值相同与否,或`X[i, j + 1]`和`X[i, j + 2]`的值相同与否,汇聚层始终输出`Y[i, j] = 1`。\n",
"也就是说,使用$2\\times 2$最大汇聚层,即使在高度或宽度上移动一个元素,卷积层仍然可以识别到模式。\n",
"\n",
"在下面的代码中的`pool2d`函数,我们(**实现汇聚层的前向传播**)。\n",
"这类似于 :numref:`sec_conv_layer`中的`corr2d`函数。\n",
"然而,这里我们没有卷积核,输出为输入中每个区域的最大值或平均值。\n"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "292e979e",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:18.192662Z",
"iopub.status.busy": "2023-08-18T07:02:18.191844Z",
"iopub.status.idle": "2023-08-18T07:02:20.224371Z",
"shell.execute_reply": "2023-08-18T07:02:20.223413Z"
},
"origin_pos": 2,
"tab": [
"pytorch"
]
},
"outputs": [],
"source": [
"import torch\n",
"from torch import nn\n",
"from d2l import torch as d2l"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "fe35adac",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:20.228639Z",
"iopub.status.busy": "2023-08-18T07:02:20.227964Z",
"iopub.status.idle": "2023-08-18T07:02:20.234155Z",
"shell.execute_reply": "2023-08-18T07:02:20.233266Z"
},
"origin_pos": 4,
"tab": [
"pytorch"
]
},
"outputs": [],
"source": [
"def pool2d(X, pool_size, mode='max'):\n",
" p_h, p_w = pool_size\n",
" Y = torch.zeros((X.shape[0] - p_h + 1, X.shape[1] - p_w + 1))\n",
" for i in range(Y.shape[0]):\n",
" for j in range(Y.shape[1]):\n",
" if mode == 'max':\n",
" Y[i, j] = X[i: i + p_h, j: j + p_w].max()\n",
" elif mode == 'avg':\n",
" Y[i, j] = X[i: i + p_h, j: j + p_w].mean()\n",
" return Y"
]
},
{
"cell_type": "markdown",
"id": "27b51b5e",
"metadata": {
"origin_pos": 6
},
"source": [
"我们可以构建 :numref:`fig_pooling`中的输入张量`X`,[**验证二维最大汇聚层的输出**]。\n"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "3a781c85",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:20.237767Z",
"iopub.status.busy": "2023-08-18T07:02:20.237211Z",
"iopub.status.idle": "2023-08-18T07:02:20.268065Z",
"shell.execute_reply": "2023-08-18T07:02:20.267212Z"
},
"origin_pos": 7,
"tab": [
"pytorch"
]
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[4., 5.],\n",
" [7., 8.]])"
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"X = torch.tensor([[0.0, 1.0, 2.0], [3.0, 4.0, 5.0], [6.0, 7.0, 8.0]])\n",
"pool2d(X, (2, 2))"
]
},
{
"cell_type": "markdown",
"id": "8cc88d86",
"metadata": {
"origin_pos": 8
},
"source": [
"此外,我们还可以(**验证平均汇聚层**)。\n"
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "4f9a1ffd",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:20.272001Z",
"iopub.status.busy": "2023-08-18T07:02:20.271411Z",
"iopub.status.idle": "2023-08-18T07:02:20.277849Z",
"shell.execute_reply": "2023-08-18T07:02:20.276928Z"
},
"origin_pos": 9,
"tab": [
"pytorch"
]
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[2., 3.],\n",
" [5., 6.]])"
]
},
"execution_count": 4,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pool2d(X, (2, 2), 'avg')"
]
},
{
"cell_type": "markdown",
"id": "447c6999",
"metadata": {
"origin_pos": 10
},
"source": [
"## [**填充和步幅**]\n",
"\n",
"与卷积层一样,汇聚层也可以改变输出形状。和以前一样,我们可以通过填充和步幅以获得所需的输出形状。\n",
"下面,我们用深度学习框架中内置的二维最大汇聚层,来演示汇聚层中填充和步幅的使用。\n",
"我们首先构造了一个输入张量`X`,它有四个维度,其中样本数和通道数都是1。\n"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "140d08f5",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:20.281458Z",
"iopub.status.busy": "2023-08-18T07:02:20.280874Z",
"iopub.status.idle": "2023-08-18T07:02:20.287391Z",
"shell.execute_reply": "2023-08-18T07:02:20.286578Z"
},
"origin_pos": 12,
"tab": [
"pytorch"
]
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[[[ 0., 1., 2., 3.],\n",
" [ 4., 5., 6., 7.],\n",
" [ 8., 9., 10., 11.],\n",
" [12., 13., 14., 15.]]]])"
]
},
"execution_count": 5,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"X = torch.arange(16, dtype=torch.float32).reshape((1, 1, 4, 4))\n",
"X"
]
},
{
"cell_type": "markdown",
"id": "f95f2492",
"metadata": {
"origin_pos": 15
},
"source": [
"默认情况下,(**深度学习框架中的步幅与汇聚窗口的大小相同**)。\n",
"因此,如果我们使用形状为`(3, 3)`的汇聚窗口,那么默认情况下,我们得到的步幅形状为`(3, 3)`。\n"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "a3cc01e3",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:20.291052Z",
"iopub.status.busy": "2023-08-18T07:02:20.290402Z",
"iopub.status.idle": "2023-08-18T07:02:20.296276Z",
"shell.execute_reply": "2023-08-18T07:02:20.295476Z"
},
"origin_pos": 17,
"tab": [
"pytorch"
]
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[[[10.]]]])"
]
},
"execution_count": 6,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pool2d = nn.MaxPool2d(3)\n",
"pool2d(X)"
]
},
{
"cell_type": "markdown",
"id": "0b19d625",
"metadata": {
"origin_pos": 20
},
"source": [
"[**填充和步幅可以手动设定**]。\n"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "9c247428",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:20.299965Z",
"iopub.status.busy": "2023-08-18T07:02:20.299310Z",
"iopub.status.idle": "2023-08-18T07:02:20.307455Z",
"shell.execute_reply": "2023-08-18T07:02:20.306477Z"
},
"origin_pos": 22,
"tab": [
"pytorch"
]
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[[[ 5., 7.],\n",
" [13., 15.]]]])"
]
},
"execution_count": 7,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pool2d = nn.MaxPool2d(3, padding=1, stride=2)\n",
"pool2d(X)"
]
},
{
"cell_type": "markdown",
"id": "635b4034",
"metadata": {
"origin_pos": 26,
"tab": [
"pytorch"
]
},
"source": [
"当然,我们可以(**设定一个任意大小的矩形汇聚窗口,并分别设定填充和步幅的高度和宽度**)。\n"
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "7c169b2f",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:20.311794Z",
"iopub.status.busy": "2023-08-18T07:02:20.311492Z",
"iopub.status.idle": "2023-08-18T07:02:20.320399Z",
"shell.execute_reply": "2023-08-18T07:02:20.319108Z"
},
"origin_pos": 30,
"tab": [
"pytorch"
]
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[[[ 5., 7.],\n",
" [13., 15.]]]])"
]
},
"execution_count": 8,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pool2d = nn.MaxPool2d((2, 3), stride=(2, 3), padding=(0, 1))\n",
"pool2d(X)"
]
},
{
"cell_type": "markdown",
"id": "a893596a",
"metadata": {
"origin_pos": 33
},
"source": [
"## 多个通道\n",
"\n",
"在处理多通道输入数据时,[**汇聚层在每个输入通道上单独运算**],而不是像卷积层一样在通道上对输入进行汇总。\n",
"这意味着汇聚层的输出通道数与输入通道数相同。\n",
"下面,我们将在通道维度上连结张量`X`和`X + 1`,以构建具有2个通道的输入。\n"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "c0a30a7f",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:20.325617Z",
"iopub.status.busy": "2023-08-18T07:02:20.324879Z",
"iopub.status.idle": "2023-08-18T07:02:20.335303Z",
"shell.execute_reply": "2023-08-18T07:02:20.334055Z"
},
"origin_pos": 35,
"tab": [
"pytorch"
]
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[[[ 0., 1., 2., 3.],\n",
" [ 4., 5., 6., 7.],\n",
" [ 8., 9., 10., 11.],\n",
" [12., 13., 14., 15.]],\n",
"\n",
" [[ 1., 2., 3., 4.],\n",
" [ 5., 6., 7., 8.],\n",
" [ 9., 10., 11., 12.],\n",
" [13., 14., 15., 16.]]]])"
]
},
"execution_count": 9,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"X = torch.cat((X, X + 1), 1)\n",
"X"
]
},
{
"cell_type": "markdown",
"id": "45add004",
"metadata": {
"origin_pos": 37
},
"source": [
"如下所示,汇聚后输出通道的数量仍然是2。\n"
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "e534c8f3",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:02:20.340529Z",
"iopub.status.busy": "2023-08-18T07:02:20.339767Z",
"iopub.status.idle": "2023-08-18T07:02:20.349365Z",
"shell.execute_reply": "2023-08-18T07:02:20.348159Z"
},
"origin_pos": 39,
"tab": [
"pytorch"
]
},
"outputs": [
{
"data": {
"text/plain": [
"tensor([[[[ 5., 7.],\n",
" [13., 15.]],\n",
"\n",
" [[ 6., 8.],\n",
" [14., 16.]]]])"
]
},
"execution_count": 10,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pool2d = nn.MaxPool2d(3, padding=1, stride=2)\n",
"pool2d(X)"
]
},
{
"cell_type": "markdown",
"id": "0a91fd9f",
"metadata": {
"origin_pos": 43
},
"source": [
"## 小结\n",
"\n",
"* 对于给定输入元素,最大汇聚层会输出该窗口内的最大值,平均汇聚层会输出该窗口内的平均值。\n",
"* 汇聚层的主要优点之一是减轻卷积层对位置的过度敏感。\n",
"* 我们可以指定汇聚层的填充和步幅。\n",
"* 使用最大汇聚层以及大于1的步幅,可减少空间维度(如高度和宽度)。\n",
"* 汇聚层的输出通道数与输入通道数相同。\n",
"\n",
"## 练习\n",
"\n",
"1. 尝试将平均汇聚层作为卷积层的特殊情况实现。\n",
"1. 尝试将最大汇聚层作为卷积层的特殊情况实现。\n",
"1. 假设汇聚层的输入大小为$c\\times h\\times w$,则汇聚窗口的形状为$p_h\\times p_w$,填充为$(p_h, p_w)$,步幅为$(s_h, s_w)$。这个汇聚层的计算成本是多少?\n",
"1. 为什么最大汇聚层和平均汇聚层的工作方式不同?\n",
"1. 我们是否需要最小汇聚层?可以用已知函数替换它吗?\n",
"1. 除了平均汇聚层和最大汇聚层,是否有其它函数可以考虑(提示:回想一下`softmax`)?为什么它不流行?\n"
]
},
{
"cell_type": "markdown",
"id": "f53a8320",
"metadata": {
"origin_pos": 45,
"tab": [
"pytorch"
]
},
"source": [
"[Discussions](https://discuss.d2l.ai/t/1857)\n"
]
}
],
"metadata": {
"language_info": {
"name": "python"
},
"required_libs": []
},
"nbformat": 4,
"nbformat_minor": 5
}