Can I fine fune a modified network?

1

I want to do fine turning.

For example, there is a pre-trained weighyt traind from the layer1-> layer2-> layer3 structure.

But I want to fine-tune on layer1-> my_layer1-> layer2-> my_layer2-> layer3 structure.

Is there any way?

please. help me

2

Hi,

This is technically possible by setting the weights to the pretrained layers yourself with something like model.layer1.load_state_dict(pretrained_layer1_state_dict).
That being said, the insertion of your custom layers might change the behavior of the network completely and make the initialization useless.

1 Like