ide without camstyle got higher mAP(my batch-size is 64) about camstyle

[zhunzhong07]

I fixed the bug of camstyle_loader, could you try to run it again? In my experiment, the rank-1 accuracy is around about 85% without CamStyle.

[jian-xu]

OK, I am running now. And I want to ask the meaning of the number given by arg.camstyle. I would like to try some different ratios between real and fake images.

[jian-xu]

Here is my understanding about the number given by args.camstyle: every time I choose 128 images from trainset, I also take 46 images from camstyle, then train the network (using different loss function).

[zhunzhong07]

Yes, the overall input images are 128+46. The ratio of real and fake is 128/46.