Is it possible to upgrade a trained ResNet34 to ResNet50?

I may show my lack of understanding of how CNNs work by asking this question, but from how I understand it, I dont see why we would not be able to.

I image a scenario where I am training using ResNet34 until I am satisfied, save it, then upgrade it to 50 to try and get better results.

Is this too wild of an idea? I would love some bounce back on this :slight_smile:

Note: I should also add, I imagine random wights would populate the layers past 34, and my training would freeze layers 1-34 so that I would only be training 35+ and possibly unfreeze later.

Resnet-50 and Resnet-34 have very different architectures. I don’t think there’s going to be any obvious way to use the weights from one to initialize the other.

1 Like