pytorch学习（2）构造简单神经网络Lenet-5

pytorch学习（2）构造简单神经网络Lenet-5

用pytorch构造一个神经网络一般有以下几个步骤：

定义一个有训练参数的神经网络

import torch.nn as nn
import torch.nn.functional as F
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        # 定义卷积核权重参数
        self.conv1 = nn.Conv2d(1, 6, 3) # 图片输入通道为1，输出通道为6，卷积核大小为3x3
        self.conv2 = nn.Conv2d(6, 16, 3)
        # 定义全连接层参数
        self.fc1 = nn.Linear(16 * 6 * 6, 120)  # 输入为16个通道的6x6特征图
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        # 池化的窗口大小为2x2
        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
        # 若池化窗口为方形，也可以用单个数字
        x = F.max_pool2d(F.relu(self.conv2(x)), 2)
        # 把3维张量flat为1维
        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)

output：

Net(
  (conv1): Conv2d(1, 6, kernel_size=(3, 3), stride=(1, 1))
  (conv2): Conv2d(6, 16, kernel_size=(3, 3), stride=(1, 1))
  (fc1): Linear(in_features=576, 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)
)

迭代的输入数据集

先尝试性的输入一个张量

input = torch.ones(1, 1, 32, 32)
out = net(input)
print(out)

output：

tensor([[ 0.0736,  0.0561,  0.0124, -0.0494, -0.0058, -0.1239, -0.1036,  0.0572,
         -0.0019, -0.0115]], grad_fn=<AddmmBackward>)

计算LOSS

先设计一个简单随机标签，然后计算其损失，其中 nn.MSELoss() 是一个均方差损失函数

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)

output：

tensor(0.7285, grad_fn=<MseLossBackward>)

反向传播梯度

有了Loss后，我们就可以得到所有参数在loss上的梯度啦

print(loss.grad_fn)  # MSELoss
print(loss.grad_fn.next_functions[0][0])  # Linear
print(loss.grad_fn.next_functions[0][0].next_functions[0][0])  # ReLU

output：

<MseLossBackward object at 0x000001AAA626C208>
<AddmmBackward object at 0x000001AAA7E88438>
<AccumulateGrad object at 0x000001AAA626C208>

要反向传播损失，我们只需要 loss.backward() ，但是需要先清楚现有的梯度，否则会积累在一起。

net.zero_grad()     # zeroes the gradient buffers of all parameters

print('conv1.bias.grad before backward')
print(net.conv1.bias.grad)

loss.backward()

print('conv1.bias.grad after backward')
print(net.conv1.bias.grad)

output：

conv1.bias.grad before backward
None
conv1.bias.grad after backward
tensor([-0.0010,  0.0000,  0.0000,  0.0000, -0.0117, -0.0164])

更新训练参数的值

神经网络中有许多更新参数的方法： SGD,Nesterov-SGD,Adam,RMSProp 等，在pytorch中的 torch.optim 封装了这些更新(优化)方法
这里以SGD为例： weight = weight - learning_rate * gradient

import torch.optim as optim
# 创建优化器
optimizer = optim.SGD(net.parameters(), lr=1) # 为了让结果更明显，我们把学习率设为1
optimizer.zero_grad()   # 把梯度缓存置0
output = net(input)
loss = criterion(output, target)
loss.backward()
print(net.conv1.bias)
optimizer.step()    # 更新
print(net.conv1.bias.grad)
print(net.conv1.bias)

output：

Parameter containing:
tensor([ 0.2609,  0.3270,  0.2890,  0.1068, -0.3012,  0.1862],
       requires_grad=True)
tensor([ 0.0037, -0.0128,  0.0154,  0.0000,  0.0000, -0.0043])
Parameter containing:
tensor([ 0.2573,  0.3398,  0.2735,  0.1068, -0.3012,  0.1905],
       requires_grad=True)