About Transfer Learning - Frozen weights?

[mitchellwiebe]

Hi there,

I have a quick question about the application of transfer learning used for LUAD/LUSC classification that achieved the reported AUC=0.847. I understand that kernels within convolutional layers are initialized with ImageNet weights. However, were these weights frozen, or did they remain trainable? When I utilize the fine-tuning method and investigate weights, it seems that the kernel weights remain trainable.

Thanks, Mitch