Pytorch 运行时错误：efficientnet 通过设置冻结层需要 grad = false

Question

我想冻结 pytorch efficentnet 模型中的层。我通常的做法不起作用。

from torchvision.models import efficientnet_b0
from torch import nn
from torch import optim

efficientnet_b0_fine = efficientnet_b0(pretrained=True)

for param in efficientnet_b0_fine.parameters():
  param.requires_grad = False

efficientnet_b0_fine.fc = nn.Linear(512, 10)

optimizer = optim.Adam(efficientnet_b0_fine.parameters(), lr=0.0001)

loss_function = nn.CrossEntropyLoss()

training(net=efficientnet_b0_fine, n_epochs=epochs, optimizer=optimizer, loss_function=loss_function, train_dl = train_dl)

我收到的错误说：

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

训练函数如下所示：

for xb, yb in train_dl:
  optimizer.zero_grad()  
  xb = xb.to(device)
  yb = yb.to(device)

  y_hat = net(xb)  
  loss = loss_function(y_hat, yb) 
  
  loss.backward()  
  optimizer.step()  

如果你们中的一个人有解决方案，那就太好了！

Answer 1

因为“张量的元素 0 不需要 grad 并且没有 grad_fn”