【674】PyTorch —— 神经网络

PyTorch 神经网络

　一个典型的神经网络训练过程包括以下几点：

1. 定义一个包含可训练参数的神经网络

2. 通过神经网络处理输入
   

3. 计算损失(loss)

4. 反向传播梯度到神经网络的参数

5. 更新网络的参数

---

 

1. 定义一个包含可训练参数的神经网络

• torch.nn 定义了相关神经网络层

• torch.nn.functional 定义了相关函数

• 自定义的函数就是为了获取除batch size外的总元素个数

• __init__(self) 函数：用来初始化层的结构

• forward(self, x) 函数：通过函数将数据流建立起来

• net.parameters()：一个模型可训练的参数

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39

import torch import torch.nn as nn import torch.nn.functional as F     class Net(nn.Module):       def __init__(self):         super(Net, self).__init__()         # 1 input image channel, 6 output channels, 5x5 square convolution         # kernel         self.conv1 = nn.Conv2d(1, 6, 5)         self.conv2 = nn.Conv2d(6, 16, 5)         # an affine operation: y = Wx + b         self.fc1 = nn.Linear(16 * 5 * 5, 120)         self.fc2 = nn.Linear(120, 84)         self.fc3 = nn.Linear(84, 10)       def forward(self, x):         # Max pooling over a (2, 2) window         x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))         # If the size is a square you can only specify a single number         x = F.max_pool2d(F.relu(self.conv2(x)), 2)         x = x.view(-1, self.num_flat_features(x))         x = F.relu(self.fc1(x))         x = F.relu(self.fc2(x))         x = self.fc3(x)         return x       def num_flat_features(self, x):         size = x.size()[1:]  # all dimensions except the batch dimension         num_features = 1         for s in size:             num_features *= s         return num_features     net = Net() print(net)

 

2. 通过神经网络处理输入

• 创建网络

?

1 2 3 4 5 6 7

input = torch.randn(1, 1, 32, 32) out = net(input) print(out)   #把所有参数梯度缓存器置零，用随机的梯度来反向传播 net.zero_grad() out.backward(torch.randn(1, 10))

 

3. 计算损失(loss)

• 定义损失函数

?

1 2 3 4 5 6 7

output = net(input) target = torch.randn(10)  # a dummy target, for example target = target.view(1, -1)  # make it the same shape as output criterion = nn.MSELoss()   loss = criterion(output, target) print(loss)

 

还有添加 optimizer 以及其他训练策略