Fine tuning resnet18 for cifar10

zz z 0
Viewed 20

I just want fine tuning ResNet18 on cifar10 datasets. so I just want to change the last linear layer from 1000 to 10. I tried use children function to get the previous layers

ResModel = resnet18(weights=ResNet18_Weights)
model = nn.Sequential(
    *list(ResModel.children())[:-1],
    nn.Linear(512,10)
)

so it raised error RuntimeError: mat1 and mat2 shapes cannot be multiplied (32768x1 and 512x10) and then I tried this way ResModel.fc=nn.Linear(512,10) it works fine. so why?

resnet deep-learning pytorch python
1 Answers

The difference between stacking all layers into a single nn.Sequential and overriding only the last layer is the forward function:
Your ResModel is of type torchvision.models.ResNet, while your model is a simple nn.Sequential. The forward pass of ResNet has an additional flatten operation before the last linear layer -- you do not have this operation in your nn.Sequential model.

usero1c70
Related