Converting pre-trained resnet34 to resnet50

I was just wondering if there was an in-built way to convert a pre-trained resnet34 to a deeper resnet. That is, I have a trained resnet34 that classifies well, and would be interested in now training a resnet50, and was hoping there’d be a way to reclaim all that initial training.
In my mind I was thinking it might be possible to insert different types of identity matrices that would mean that you have new trainable parameters without changing the current performance of the network, and was wondering if anyone knew if someone had looked into this before.
Thanks in advance :slight_smile:

It is not possible. If you are thinking about adding some identity matrices, than the problem with that approach is a convnet learns representations from sequentially from (left to right) so if you insert a new layer than it essentially means is has to learn all the layers after that. So in essence you are learning the complete model from scratch.

1 Like

Oh well, thank you for the answer, that actually makes perfect sense why that would not work