train acc about high resolution on cifar

[wangqiim]

Thanks for you work. I am reproducing your work on cifar dataset, but I find the high/low resolution acc always lower than independent train each resolution. Have you ever had this problem?

[yikaiw]

Hi, thanks for your interest. We did not try to evaluate on cifar. But I guess If independent BNs, and the multi-resolution setting in dataloader (the same random crop is resized multiple times, not multiple random crops) are properly implemented, the method could potentially still work.

[wangqiim]

I switch different resolution use

input # [batchsize, 3, 32, 32]
inputs = [
F.interpolate(input, size=[32, 32]).cuda(), 
F.interpolate(input, size=[16, 16]).cuda(), 
F.interpolate(input, size=[8, 8]).cuda()
]

instead of modify dataloader. May be the resolution is too small or this function is not fit I guess.........