Transfer Learning train two last layer (Bloc)

1

Hello,
it is possible to set params in layer before fully-connected layer to train params like case 2 in the table ?
thanks :slight_smile:

Capture d’écran 2021-08-09 à 12.48.56
Capture d’écran 2021-08-09 à 12.48.56893×318 118 KB

2

Yes, you can freeze specific layers and train others by manipulating the .requires_grad attribute of each parameter (setting it to False would freeze the parameter). The Finetuning tutorial would give you an example.