{
"cells": [
{
"cell_type": "markdown",
"id": "d5217b24",
"metadata": {
"origin_pos": 0
},
"source": [
"# 多层感知机的简洁实现\n",
":label:`sec_mlp_concise`\n",
"\n",
"本节将介绍(**通过高级API更简洁地实现多层感知机**)。\n"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "f4b9d183",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:04:20.711610Z",
"iopub.status.busy": "2023-08-18T07:04:20.711337Z",
"iopub.status.idle": "2023-08-18T07:04:22.715766Z",
"shell.execute_reply": "2023-08-18T07:04:22.714884Z"
},
"origin_pos": 2,
"tab": [
"pytorch"
]
},
"outputs": [],
"source": [
"import torch\n",
"from torch import nn\n",
"from d2l import torch as d2l"
]
},
{
"cell_type": "markdown",
"id": "d1b8af0c",
"metadata": {
"origin_pos": 5
},
"source": [
"## 模型\n",
"\n",
"与softmax回归的简洁实现（ :numref:`sec_softmax_concise`）相比，\n",
"唯一的区别是我们添加了2个全连接层（之前我们只添加了1个全连接层）。\n",
"第一层是[**隐藏层**]，它(**包含256个隐藏单元，并使用了ReLU激活函数**)。\n",
"第二层是输出层。\n"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "a11cfbe9",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:04:22.719981Z",
"iopub.status.busy": "2023-08-18T07:04:22.719298Z",
"iopub.status.idle": "2023-08-18T07:04:22.748628Z",
"shell.execute_reply": "2023-08-18T07:04:22.747813Z"
},
"origin_pos": 7,
"tab": [
"pytorch"
]
},
"outputs": [],
"source": [
"net = nn.Sequential(nn.Flatten(),\n",
" nn.Linear(784, 256),\n",
" nn.ReLU(),\n",
" nn.Linear(256, 10))\n",
"\n",
"def init_weights(m):\n",
" if type(m) == nn.Linear:\n",
" nn.init.normal_(m.weight, std=0.01)\n",
"\n",
"net.apply(init_weights);"
]
},
{
"cell_type": "markdown",
"id": "f5aceed6",
"metadata": {
"origin_pos": 10
},
"source": [
"[**训练过程**]的实现与我们实现softmax回归时完全相同，\n",
"这种模块化设计使我们能够将与模型架构有关的内容独立出来。\n"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "b23e8ab9",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:04:22.753701Z",
"iopub.status.busy": "2023-08-18T07:04:22.753406Z",
"iopub.status.idle": "2023-08-18T07:04:22.758051Z",
"shell.execute_reply": "2023-08-18T07:04:22.757284Z"
},
"origin_pos": 12,
"tab": [
"pytorch"
]
},
"outputs": [],
"source": [
"batch_size, lr, num_epochs = 256, 0.1, 10\n",
"loss = nn.CrossEntropyLoss(reduction='none')\n",
"trainer = torch.optim.SGD(net.parameters(), lr=lr)"
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "78ac9bf1",
"metadata": {
"execution": {
"iopub.execute_input": "2023-08-18T07:04:22.761842Z",
"iopub.status.busy": "2023-08-18T07:04:22.761295Z",
"iopub.status.idle": "2023-08-18T07:05:05.308680Z",
"shell.execute_reply": "2023-08-18T07:05:05.307786Z"
},
"origin_pos": 15,
"tab": [
"pytorch"
]
},
"outputs": [
{
"data": {
"image/svg+xml": [
"\n",
"\n",
"\n"
],
"text/plain": [
""
]
},
"metadata": {
"needs_background": "light"
},
"output_type": "display_data"
}
],
"source": [
"train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size)\n",
"d2l.train_ch3(net, train_iter, test_iter, loss, num_epochs, trainer)"
]
},
{
"cell_type": "markdown",
"id": "9b636c57",
"metadata": {
"origin_pos": 16
},
"source": [
"## 小结\n",
"\n",
"* 我们可以使用高级API更简洁地实现多层感知机。\n",
"* 对于相同的分类问题，多层感知机的实现与softmax回归的实现相同，只是多层感知机的实现里增加了带有激活函数的隐藏层。\n",
"\n",
"## 练习\n",
"\n",
"1. 尝试添加不同数量的隐藏层（也可以修改学习率），怎么样设置效果最好？\n",
"1. 尝试不同的激活函数，哪个效果最好？\n",
"1. 尝试不同的方案来初始化权重，什么方法效果最好？\n"
]
},
{
"cell_type": "markdown",
"id": "36201fb3",
"metadata": {
"origin_pos": 18,
"tab": [
"pytorch"
]
},
"source": [
"[Discussions](https://discuss.d2l.ai/t/1802)\n"
]
}
],
"metadata": {
"language_info": {
"name": "python"
},
"required_libs": []
},
"nbformat": 4,
"nbformat_minor": 5
}