Results on WebVision

[MrChenFeng]

Hi,
thanks for your great work.
I'm so sorry to bother you because this repo is a few years old, but I'd still like to know about your setup for experiments on webvision. As you may know, the results of this article are widely used in many articles, mostly the baseline numbers on webvsion, but after doing direct supervised learning without any modifications[no small loss things], I found that the results of the webvsion dataset can reach top1 acc ~75. I use 64batchsize, inceptionresnetv2, 100epochs, lr=0.01, changing to 0.001 and 0.0001 in 50th and 80th epoch. I'm worried whether most of the current works are really improving or not?

Best,
Chen