Fine-tuning resnet, learning rate


(Pigeon Jole) #1

Hello guys,

I am trying to fine-tune resnet18 for Image classification on German Traffic Sign Dataset to compare with my “from-scratch” module.
I am using resnet as a fixed feature extractor.
I have updated the FC layer for the number of classes for this problem.
I am using Adam optimizer with 1e-4 (I have seen for transfer learning, learning rate should be really small??)
However, training is really slow and accuracy is really bad.
Any pointers?


(Max) #2

There is a nice thing called learn rate finder, based on Leslie N. Smith’s idea.