记录如何用Pytorch搭建LeNet-5,大体步骤包括:网络的搭建->前向传播->定义Loss和Optimizer->训练
# -*- coding: utf-8 -*-
# All codes and comments from <<深度学习框架Pytorch入门与实践>>
# Code url : https://github.com/zhouzhoujack/pytorch-book
# lesson_2 : Neural network of PT(Pytorch)
# torch.nn是专门为神经网络设计的模块化接口,nn构建于 Autograd之上,可用来定义和运行神经网络
# 定义网络时,需要继承nn.Module,并实现它的forward方法,把网络中具有可学习参数的层放在构造函数__init__中
# 下面是LeNet-5网络结构
import torch as t
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
class Net(nn.Module):
def __init__(self):
# nn.Module子类的函数必须在构造函数中执行父类的构造函数
# 下式等价于nn.Module.__init__(self)
super(Net, self).__init__()
self.conv1 = nn.Conv2d(1, 6, 5) # 卷积层'1'表示输入图片为单通道, '6'表示输出通道数,'5'表示卷积核为5*5
self.conv2 = nn.Conv2d(6, 16, 5)
self.fc1 = nn.Linear(in_features=16 * 5 * 5, out_features=120, bias=True) # 全连接层,y = x*transposition(A) + b
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10)
def forward(self, x):
x = F.max_pool2d(input=F.relu(self.conv1(x)), kernel_size=(2, 2)) # 卷积 -> 激活 -> 池化
x = F.max_pool2d(F.relu(self.conv2(x)), 2)
# view函数只能由于contiguous的张量上,就是在内存中连续存储的张量,当tensor之前调用了transpose,
# permute函数就会是tensor内存中变得不再连续,就不能调用view函数。
# tensor.view() = np.reshape()
x = x.view(x.size()[0], -1)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
"""
Net(
(conv1): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1))
(conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
(fc1): Linear(in_features=400, out_features=120, bias=True)
(fc2): Linear(in_features=120, out_features=84, bias=True)
(fc3): Linear(in_features=84, out_features=10, bias=True)
)
"""
net = Net()
# 网络的可学习参数通过net.parameters()返回,net.named_parameters可同时返回可学习的参数及名称
"""
conv1.weight : torch.Size([6, 1, 5, 5])
conv1.bias : torch.Size([6])
conv2.weight : torch.Size([16, 6, 5, 5])
conv2.bias : torch.Size([16])
fc1.weight : torch.Size([120, 400])
fc1.bias : torch.Size([120])
fc2.weight : torch.Size([84, 120])
fc2.bias : torch.Size([84])
fc3.weight : torch.Size([10, 84])
fc3.bias : torch.Size([10])
"""
# parameters infomation of network
# params = list(net.parameters())
# for name,parameters in net.named_parameters():
# print(name,':',parameters.size())
if __name__ == '__main__':
"""
计算图如下:
input -> conv2d -> relu -> maxpool2d -> conv2d -> relu -> maxpool2d
-> view -> linear -> relu -> linear -> relu -> linear
-> MSELoss
-> loss
"""
input = t.randn(1, 1, 32, 32)
output = net(input)
# >>torch.arange(1., 4.)
# >>1 2 3 [torch.FloatTensor of size 3]
# if missing . , the type of torch will change to int
target = t.arange(0., 10.).view(1, 10)
criterion = nn.MSELoss()
loss = criterion(output, target)
print(loss)
# 运行.backward,观察调用之前和调用之后的grad
net.zero_grad() # 把net中所有可学习参数的梯度清零
print('反向传播之前 conv1.bias的梯度')
print(net.conv1.bias.grad)
loss.backward()
print('反向传播之后 conv1.bias的梯度')
print(net.conv1.bias.grad)
# Optimizer
# torch.optim中实现了深度学习中绝大多数的优化方法,例如RMSProp、Adam、SGD等
# 在反向传播计算完所有参数的梯度后,还需要使用优化方法来更新网络的权重和参数,例如随机梯度下降法(SGD)的更新策略如下:
# weight = weight - learning_rate * gradient
optimizer = optim.SGD(net.parameters(), lr=0.01)
# 在训练过程中
# 先梯度清零(与net.zero_grad()效果一样)
optimizer.zero_grad()
# 计算损失
output = net(input)
loss = criterion(output, target)
# 反向传播
loss.backward()
# 更新参数
optimizer.step()
nn.Conv2d()详解
torch.nn.Conv2d(in_channels, # input channels
out_channels, # output channels
kernel_size, # conv kernel size
stride=1,
padding=0, # add the number of zeros per dimension
dilation=1,
groups=1,
bias=True # default=True
)
其中Conv2d 的输入 input 尺寸为
,输出 output 尺寸为
Feature Map 大小计算
Size of Feature Map = (W - F + 2P)/S + 1
W : 输入图像尺寸宽度
F : 卷积核宽度
P:边界填充0数量
S:滑动步长
例如:
输入(227,227,3)
卷积层 kernel_size = 11
stride = 4
padding = 0
n(卷积核数量) = 96
输出 (55,55,96)
(227 - 11 + 0) /4 +1 = 55
参考资料
nn.Conv2d()详解:https://www.aiuai.cn/aifarm618.html