Thank you for your great works.
I wonder what is the benefit of loading resnet50 pretrained? Why don't you training from scratch?
If I load resnet50 pretrained, fix the parameters of the Resnet50 and Only train the GCNN part? How do you think about the result of this?
Regards,