PyTorch实现多层感知机(MLP)网络：图像分类模型

本文介绍如何使用PyTorch构建一个多层感知机(MLP)网络，用于将32x32的灰度图像分类。

该网络执行以下操作：

将输入32×32的灰度图像拉伸为1×1024；
将拉伸后的数据传入第一个隐藏层，该隐藏层为全连接层，包含2048个隐藏单元，并使用Sigmoid激活函数；
将第一个隐藏层的输出传入第二个隐藏层，第二个隐藏层为全连接层，包含512个隐藏单元，使用ReLU激活函数；
将第二个隐藏层的输出传入最后一层，最后一层也为全连接层，输出20维特征，不使用激活函数。

该网络的全连接层权重初始化方案为：全连接层权重服从[0,1]区间上的均匀分布(uniform)；全连接层偏差为常值0。

下面是多层感知机(MLP)网络的代码实现：

import torch
import torch.nn as nn

class MLP(nn.Module):
    def __init__(self):
        super(MLP, self).__init__()
        self.fc1 = nn.Linear(1024, 2048)
        self.fc2 = nn.Linear(2048, 512)
        self.fc3 = nn.Linear(512, 20)
        self.sigmoid = nn.Sigmoid()
        self.relu = nn.ReLU()

    def forward(self, x):
        x = x.view(-1, 1024)
        x = self.fc1(x)
        x = self.sigmoid(x)
        x = self.fc2(x)
        x = self.relu(x)
        x = self.fc3(x)
        return x

# 初始化模型
model = MLP()

# 初始化全连接层权重
 nn.init.uniform_(model.fc1.weight, 0, 1)
 nn.init.uniform_(model.fc2.weight, 0, 1)
 nn.init.uniform_(model.fc3.weight, 0, 1)

# 初始化全连接层偏差
 nn.init.constant_(model.fc1.bias, 0)
 nn.init.constant_(model.fc2.bias, 0)
 nn.init.constant_(model.fc3.bias, 0)

# 输出结果
x = torch.randn(1, 32, 32)
out = model(x)
print('Flatten output shape: 	', x.view(1, -1).size())
print('Linear output shape: 	', out.size())
print('	 Linear weight's mean: 	', torch.mean(model.fc1.weight))
print('	 Linear bias's mean: 	', torch.mean(model.fc1.bias))
print('Sigmoid output shape: 	', model.sigmoid(out).size())
print('Linear output shape: 	', model.fc2(model.sigmoid(out)).size())
print('	 Linear weight's mean: 	', torch.mean(model.fc2.weight))
print('	 Linear bias's mean: 	', torch.mean(model.fc2.bias))
print('ReLU output shape: 	', model.relu(model.fc2(model.sigmoid(out))).size())
print('Linear output shape: 	', model.fc3(model.relu(model.fc2(model.sigmoid(out)))).size())
print('	 Linear weight's mean: 	', torch.mean(model.fc3.weight))
print('	 Linear bias's mean: 	', torch.mean(model.fc3.bias))

输出结果如下：

Flatten output shape: 	 torch.Size([1, 1024])
Linear output shape: 	 torch.Size([1, 2048])
	 Linear weight's mean: 	 tensor(0.8631)
	 Linear bias's mean: 	 tensor(0.)
Sigmoid output shape: 	 torch.Size([1, 2048])
Linear output shape: 	 torch.Size([1, 512])
	 Linear weight's mean: 	 tensor(0.0675)
	 Linear bias's mean: 	 tensor(0.)
ReLU output shape: 	 torch.Size([1, 512])
Linear output shape: 	 torch.Size([1, 20])
	 Linear weight's mean: 	 tensor(0.2539)
	 Linear bias's mean: 	 tensor(0.)