Let’s say I have two cov neural networks (network A and network B) where the architectures (network A has 19 layers and network B has 20 layers) are almost identical.
How could I transfer the first five layers of weights/biases/optimizer from network A to network B?
I need some help with coding. Thanks.
I think the cleanest approach would be to create a mapping between the layer names of
modelB and load each
state_dict of these layers separately as described here.
This approach would need an explicit definition of which layers should be loaded and would not rely on e.g. using the
strict=False argument, which might yield unexpected results if the output is ignored.