记录如何用Pytorch搭建LeNet-5,大致步骤包括:网络的搭建->前向传播->定义Loss和Optimizer->训练html
# -*- 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,须要继承nn.Module,并实现它的forward方法,把网络中具备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) # 全链接层,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函数只能因为contiguous的张量上,就是在内存中连续存储的张量,当tensor以前调用了transpose, # permute函数就会是tensor内存中变得再也不连续,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() # 网络的可学习参数经过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__': """ 计算图以下: 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,观察调用以前和调用以后的grad net.zero_grad() # 把net中全部可学习参数的梯度清零 print('反向传播以前 conv1.bias的梯度') print(net.conv1.bias.grad) loss.backward() print('反向传播以后 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 # 在反向传播计算完全部参数的梯度后,还须要使用优化方法来更新网络的权重和参数,例如随机梯度降低法(SGD)的更新策略以下: # weight = weight - learning_rate * gradient optimizer = optim.SGD(net.parameters(), lr=0.01) # 在训练过程当中 # 先梯度清零(与net.zero_grad()效果同样) 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()
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
python
Size of Feature Map = (W - F + 2P)/S + 1git
W : 输入图像尺寸宽度github
F : 卷积核宽度网络
P:边界填充0数量框架
S:滑动步长less
例如:ide
输入(227,227,3)模块化
卷积层 kernel_size = 11函数
stride = 4
padding = 0
n(卷积核数量) = 96
输出 (55,55,96)
(227 - 11 + 0) /4 +1 = 55
**nn.Conv2d()详解:**https://www.aiuai.cn/aifarm618.html