#requires-grad

Question 1: Single choice

Following d2l, you build finetune_net from a pretrained ResNet-18, replace its output layer, and then try to freeze the pretrained layers:

finetune_net = torchvision.models.resnet18(weights="IMAGENET1K_V1")
finetune_net.fc = nn.Linear(finetune_net.fc.in_features, 2)
for param in finetune_net.parameters():
    param.requires_grad = False

You meant to train only the new output layer. What happens when training starts?

All layers still train, because requires_grad only matters for layers created after it is set1
The first loss.backward() raises an error, because the new fc layer is frozen too2
Only the new fc layer is trained, exactly as intended3
Training runs without errors, but no weight changes, so accuracy stays where it started4

Timed test