为什么在 'with torch.no_grad():' 中包含 'loss.backward()' 后反向传播过程仍然有效？

Question

我正在 PyTorch 中处理线性回归示例。我知道我在 'with torch.no_grad():' 中包含 'loss.backward()' 时做错了，但为什么它与我的代码配合得很好？

根据pytorch docs，torch.autograd.no_grad是一个禁用梯度计算的上下文管理器。所以我真的很困惑。

代码在这里：

import torch
import torch.nn as nn
import numpy as np
import matplotlib.pyplot as plt
# Toy dataset
x_train = np.array([[3.3], [4.4], [5.5], [6.71], [6.93], [4.168], 
                    [9.779], [6.182], [7.59], [2.167], [7.042], 
                    [10.791], [5.313], [7.997], [3.1]], dtype=np.float32)

y_train = np.array([[1.7], [2.76], [2.09], [3.19], [1.694], [1.573], 
                    [3.366], [2.596], [2.53], [1.221], [2.827], 
                    [3.465], [1.65], [2.904], [1.3]], dtype=np.float32)

input_size = 1
output_size = 1
epochs = 100
learning_rate = 0.05

model = nn.Linear(input_size, output_size)


criterion = nn.MSELoss(reduction='sum')
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

# training
for epoch in range(epochs):
    # convert numpy to tensor
    inputs = torch.from_numpy(x_train)
    targets = torch.from_numpy(y_train)

    # forward
    out = model(inputs)
    loss = criterion(out, targets)

    # backward
    with torch.no_grad():
        model.zero_grad()
        loss.backward()
        optimizer.step()
        print('inputs grad : ', inputs.requires_grad)
    if epoch % 5 == 0:
        print ('Epoch [{}/{}], Loss: {:.4f}'.format(epoch+1, epochs, loss.item()))

predicted = model(torch.from_numpy(x_train)).detach().numpy()
plt.plot(x_train, y_train, 'ro', label='Original data')
plt.plot(x_train, predicted, label='Fitted line')
plt.legend()
plt.show()

# Save the model checkpoint
torch.save(model.state_dict(), 'model\linear_model.ckpt')

预先感谢您回答我的问题。

Answer 1

这是有效的，因为损失计算发生在 no_grad 之前，并且您根据该损失计算（启用了梯度的计算）继续计算梯度。

基本上，您使用 no_grad 之外计算的梯度继续更新层的权重。

当你实际使用no_grad时：

for epoch in range(epochs):
    # convert numpy to tensor
    inputs = torch.from_numpy(x_train)
    targets = torch.from_numpy(y_train)
    with torch.no_grad():  # no_grad used here
    # forward
        out = model(inputs)
        loss = criterion(out, targets)
        model.zero_grad()
        loss.backward()
        optimizer.step()
        print('inputs grad : ', inputs.requires_grad)
    if epoch % 5 == 0:
        print ('Epoch [{}/{}], Loss: {:.4f}'.format(epoch+1, epochs, loss.item()))

然后你会得到正确的错误，说：

element 0 of tensors does not require grad and does not have a grad_fn.

也就是你用no_grad不合适的地方。

如果你打印loss的.requires_grad，那么你会看到loss有requires_grad。

也就是说，当你这样做时：

for epoch in range(epochs):
    # convert numpy to tensor
    inputs = torch.from_numpy(x_train)
    targets = torch.from_numpy(y_train)

    # forward
    out = model(inputs)
    loss = criterion(out, targets)

    # backward
    with torch.no_grad():
        model.zero_grad()
        loss.backward()

        optimizer.step()
        print('inputs grad : ', inputs.requires_grad)
        print('loss grad : ', loss.requires_grad)  # Prints loss.require_rgad
    if epoch % 5 == 0:
        print ('Epoch [{}/{}], Loss: {:.4f}'.format(epoch+1, epochs, loss.item()))

您将看到：

inputs grad :  False
loss grad :  True

此外，

print('inputs grad : ', inputs.requires_grad)

将始终打印 False。也就是说，如果你这样做

for epoch in range(epochs):
    # convert numpy to tensor
    inputs = torch.from_numpy(x_train)
    targets = torch.from_numpy(y_train)

    print('inputs grad : ', inputs.requires_grad). # Print the inputs.requires_grad

    # forward
    out = model(inputs)
    loss = criterion(out, targets)

    # backward
    with torch.no_grad():
        model.zero_grad()
        loss.backward()

        optimizer.step()
        print('inputs grad : ', inputs.requires_grad)
        print('loss grad : ', loss.requires_grad)
    if epoch % 5 == 0:
        print ('Epoch [{}/{}], Loss: {:.4f}'.format(epoch+1, epochs, loss.item()))

您将获得：

inputs grad :  False
inputs grad :  False
loss grad :  True

就是你在用错误的东西来检查你做错了什么。你能做的最好的事情就是再次阅读 PyTorch 的梯度力学文档。

为什么在 'with torch.no_grad():' 中包含 'loss.backward()' 后反向传播过程仍然有效？

Why the backpropagation process can still work when I included 'loss.backward()' in 'with torch.no_grad():'?

backpropagation

pytorch