Merging resnet34 into resnet152

Part 1 (2019) Advanced (Part 1 v3)
1

Hi,

I have a trained resnet34. Since resent is built in blocks is it possible to utilize the resnet34 and merge the learned weights into resnet152? Would this be useful? I feel like it could be a good approach instead of retraining entirely from scratch. Have there been any papers written on this topic and/or code published?

Is this possible with the fastai library?

2

resnet34 and resnet152 use different blocks in their architectures, so they are essentially completely different. The details can be found in the original paper or the torchvision.models source code.