使用pytorch修改模型权重时没有梯度问题

Question

当我尝试向模型添加扰动并优化扰动本身而不是模型参数时。我确实喜欢以下内容，非常简单：

import torch
import torch.nn as nn
import torch.optim as optim

# Define a simple model
class SimpleModel(nn.Module):
    def __init__(self):
        super(SimpleModel, self).__init__()
        self.linear = nn.Linear(1, 1)

    def forward(self, x, perturbation, clean):
        self.linear.weight.data = perturbation + clean
        return self.linear(x) 

# Generate clean dataset
X = torch.tensor([[1.0], [2.0], [3.0], [4.0]], dtype=torch.float32)
Y = 2 * X

# Create the model
model = SimpleModel()

# Define perturbation as a separate parameter
perturbation = torch.tensor(0.1, requires_grad=True)
clean = model.linear.weight.data
# Train the model and optimize the perturbation
criterion = nn.MSELoss()
optimizer_perturbation = optim.SGD([perturbation], lr=0.01)  # Optimize the perturbation

for epoch in range(100):
    optimizer_perturbation.zero_grad()

    outputs = model(X,clean,perturbation)
    loss = criterion(outputs, Y)
    loss.backward()
    # Update the perturbation
    optimizer_perturbation.step()

但是，执行loss.backward()后，扰动的梯度仍然是None。我不明白为什么。是什么导致了这个问题？我应该怎样做才能实现我需要的结果？

我认为执行loss.backward()后perturbation.grad不应该是None，但它是。

#before executing loss.backward()
model.linear.weight.grad
None
perturbation.grad
None
#after executing loss.backward()
perturbation.grad
None
model.linear.weight.grad
tensor([[-26.8524]])

Answer 1

您追踪

loss.backward()

，但您从不打电话给

perturbation.backward()

。

如果将

perturbation.backward()

添加到循环中，则

perturbation.grad

应该返回 smth。