{
"cells": [
{
"cell_type": "markdown",
"id": "324028bf",
"metadata": {
"origin_pos": 0
},
"source": [
"# 网络中的网络（NiN）\n",
":label:`sec_nin`\n",
"\n",
"LeNet、AlexNet和VGG都有一个共同的设计模式：通过一系列的卷积层与汇聚层来提取空间结构特征；然后通过全连接层对特征的表征进行处理。\n",
"AlexNet和VGG对LeNet的改进主要在于如何扩大和加深这两个模块。\n",
"或者，可以想象在这个过程的早期使用全连接层。然而，如果使用了全连接层，可能会完全放弃表征的空间结构。\n",
"*网络中的网络*（*NiN*）提供了一个非常简单的解决方案：在每个像素的通道上分别使用多层感知机 :cite:`Lin.Chen.Yan.2013`\n",
"\n",
"## (**NiN块**)\n",
"\n",
"回想一下，卷积层的输入和输出由四维张量组成，张量的每个轴分别对应样本、通道、高度和宽度。\n",
"另外，全连接层的输入和输出通常是分别对应于样本和特征的二维张量。\n",
"NiN的想法是在每个像素位置（针对每个高度和宽度）应用一个全连接层。\n",
"如果我们将权重连接到每个空间位置，我们可以将其视为$1\\times 1$卷积层（如 :numref:`sec_channels`中所述），或作为在每个像素位置上独立作用的全连接层。\n",
"从另一个角度看，即将空间维度中的每个像素视为单个样本，将通道维度视为不同特征（feature）。\n",
"\n",
" :numref:`fig_nin`说明了VGG和NiN及它们的块之间主要架构差异。\n",
"NiN块以一个普通卷积层开始，后面是两个$1 \\times 1$的卷积层。这两个$1 \\times 1$卷积层充当带有ReLU激活函数的逐像素全连接层。\n",
"第一层的卷积窗口形状通常由用户设置。\n",
"随后的卷积窗口形状固定为$1 \\times 1$。\n",
"\n",
"\n",
":width:`600px`\n",
":label:`fig_nin`\n"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "9b116832",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:19:49.388461Z",
"iopub.status.busy": "2023-08-18T07:19:49.387694Z",
"iopub.status.idle": "2023-08-18T07:19:52.241405Z",
"shell.execute_reply": "2023-08-18T07:19:52.240502Z"
},
"origin_pos": 2,
"tab": [
"pytorch"
]
},
"outputs": [],
"source": [
"import torch\n",
"from torch import nn\n",
"from d2l import torch as d2l\n",
"\n",
"\n",
"def nin_block(in_channels, out_channels, kernel_size, strides, padding):\n",
" return nn.Sequential(\n",
" nn.Conv2d(in_channels, out_channels, kernel_size, strides, padding),\n",
" nn.ReLU(),\n",
" nn.Conv2d(out_channels, out_channels, kernel_size=1), nn.ReLU(),\n",
" nn.Conv2d(out_channels, out_channels, kernel_size=1), nn.ReLU())"
]
},
{
"cell_type": "markdown",
"id": "fc11a427",
"metadata": {
"origin_pos": 5
},
"source": [
"## [**NiN模型**]\n",
"\n",
"最初的NiN网络是在AlexNet后不久提出的，显然从中得到了一些启示。\n",
"NiN使用窗口形状为$11\\times 11$、$5\\times 5$和$3\\times 3$的卷积层，输出通道数量与AlexNet中的相同。\n",
"每个NiN块后有一个最大汇聚层，汇聚窗口形状为$3\\times 3$，步幅为2。\n",
"\n",
"NiN和AlexNet之间的一个显著区别是NiN完全取消了全连接层。\n",
"相反，NiN使用一个NiN块，其输出通道数等于标签类别的数量。最后放一个*全局平均汇聚层*（global average pooling layer），生成一个对数几率\t（logits）。NiN设计的一个优点是，它显著减少了模型所需参数的数量。然而，在实践中，这种设计有时会增加训练模型的时间。\n"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "8ba4ca30",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:19:52.246912Z",
"iopub.status.busy": "2023-08-18T07:19:52.246237Z",
"iopub.status.idle": "2023-08-18T07:19:52.294688Z",
"shell.execute_reply": "2023-08-18T07:19:52.293840Z"
},
"origin_pos": 7,
"tab": [
"pytorch"
]
},
"outputs": [],
"source": [
"net = nn.Sequential(\n",
" nin_block(1, 96, kernel_size=11, strides=4, padding=0),\n",
" nn.MaxPool2d(3, stride=2),\n",
" nin_block(96, 256, kernel_size=5, strides=1, padding=2),\n",
" nn.MaxPool2d(3, stride=2),\n",
" nin_block(256, 384, kernel_size=3, strides=1, padding=1),\n",
" nn.MaxPool2d(3, stride=2),\n",
" nn.Dropout(0.5),\n",
" # 标签类别数是10\n",
" nin_block(384, 10, kernel_size=3, strides=1, padding=1),\n",
" nn.AdaptiveAvgPool2d((1, 1)),\n",
" # 将四维的输出转成二维的输出，其形状为(批量大小,10)\n",
" nn.Flatten())"
]
},
{
"cell_type": "markdown",
"id": "76c9e9ac",
"metadata": {
"origin_pos": 10
},
"source": [
"我们创建一个数据样本来[**查看每个块的输出形状**]。\n"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "2917eb46",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:19:52.298946Z",
"iopub.status.busy": "2023-08-18T07:19:52.298371Z",
"iopub.status.idle": "2023-08-18T07:19:52.325792Z",
"shell.execute_reply": "2023-08-18T07:19:52.324591Z"
},
"origin_pos": 12,
"tab": [
"pytorch"
]
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Sequential output shape:\t torch.Size([1, 96, 54, 54])\n",
"MaxPool2d output shape:\t torch.Size([1, 96, 26, 26])\n",
"Sequential output shape:\t torch.Size([1, 256, 26, 26])\n",
"MaxPool2d output shape:\t torch.Size([1, 256, 12, 12])\n",
"Sequential output shape:\t torch.Size([1, 384, 12, 12])\n",
"MaxPool2d output shape:\t torch.Size([1, 384, 5, 5])\n",
"Dropout output shape:\t torch.Size([1, 384, 5, 5])\n",
"Sequential output shape:\t torch.Size([1, 10, 5, 5])\n",
"AdaptiveAvgPool2d output shape:\t torch.Size([1, 10, 1, 1])\n",
"Flatten output shape:\t torch.Size([1, 10])\n"
]
}
],
"source": [
"X = torch.rand(size=(1, 1, 224, 224))\n",
"for layer in net:\n",
" X = layer(X)\n",
" print(layer.__class__.__name__,'output shape:\\t', X.shape)"
]
},
{
"cell_type": "markdown",
"id": "95a4c245",
"metadata": {
"origin_pos": 15
},
"source": [
"## [**训练模型**]\n",
"\n",
"和以前一样，我们使用Fashion-MNIST来训练模型。训练NiN与训练AlexNet、VGG时相似。\n"
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "3c362e21",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:19:52.329626Z",
"iopub.status.busy": "2023-08-18T07:19:52.329083Z",
"iopub.status.idle": "2023-08-18T07:24:13.568006Z",
"shell.execute_reply": "2023-08-18T07:24:13.566948Z"
},
"origin_pos": 16,
"tab": [
"pytorch"
]
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"loss 0.563, train acc 0.786, test acc 0.790\n",
"3087.6 examples/sec on cuda:0\n"
]
},
{
"data": {
"image/svg+xml": [
"\n",
"\n",
"\n"
],
"text/plain": [
""
]
},
"metadata": {
"needs_background": "light"
},
"output_type": "display_data"
}
],
"source": [
"lr, num_epochs, batch_size = 0.1, 10, 128\n",
"train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size, resize=224)\n",
"d2l.train_ch6(net, train_iter, test_iter, num_epochs, lr, d2l.try_gpu())"
]
},
{
"cell_type": "markdown",
"id": "847261ce",
"metadata": {
"origin_pos": 17
},
"source": [
"## 小结\n",
"\n",
"* NiN使用由一个卷积层和多个$1\\times 1$卷积层组成的块。该块可以在卷积神经网络中使用，以允许更多的每像素非线性。\n",
"* NiN去除了容易造成过拟合的全连接层，将它们替换为全局平均汇聚层（即在所有位置上进行求和）。该汇聚层通道数量为所需的输出数量（例如，Fashion-MNIST的输出为10）。\n",
"* 移除全连接层可减少过拟合，同时显著减少NiN的参数。\n",
"* NiN的设计影响了许多后续卷积神经网络的设计。\n",
"\n",
"## 练习\n",
"\n",
"1. 调整NiN的超参数，以提高分类准确性。\n",
"1. 为什么NiN块中有两个$1\\times 1$卷积层？删除其中一个，然后观察和分析实验现象。\n",
"1. 计算NiN的资源使用情况。\n",
" 1. 参数的数量是多少？\n",
" 1. 计算量是多少？\n",
" 1. 训练期间需要多少显存？\n",
" 1. 预测期间需要多少显存？\n",
"1. 一次性直接将$384 \\times 5 \\times 5$的表示缩减为$10 \\times 5 \\times 5$的表示，会存在哪些问题？\n"
]
},
{
"cell_type": "markdown",
"id": "288616e8",
"metadata": {
"origin_pos": 19,
"tab": [
"pytorch"
]
},
"source": [
"[Discussions](https://discuss.d2l.ai/t/1869)\n"
]
}
],
"metadata": {
"language_info": {
"name": "python"
},
"required_libs": []
},
"nbformat": 4,
"nbformat_minor": 5
}