How to use linear layer after AdaptiveAvgPool2d?

Maybe it might be better for some use cases. However, the current implementation just sticks to the original ResNet paper.