Performance for RegNetY

[zhanghang1989]

First, thanks for the awesome work of re-implementing RegNet.

I am having difficulty of reproducing the results for RegNetY-0.4GF. The configurations are taken from the original repo:

group_width = 8
initial_width = 48
slope = 27.89
quantized_param = 2.09
network_depth = 16

I only get 72.92 top-1 accuracy, but the original paper reported 74.2. Any thoughts on that?

[uvipen]

Hi, Thanks for your interest. Do you change anything else on our code (e.g data augmentation, schedule, ...)?

[zhanghang1989]

Thanks for the quick reply!

I used different training HPs learning rate: 0.2, batch size: 512, weight decay: 1e-4

When switching back to paper default: learning rate: 0.8, batch size: 1024, weight decay: 5e-5, I got 73.66 now.

[uvipen]

Great. Pls notice that data-preprocessing stage also influences performance. Since the paper didnt explicitly mention this step, we only applied very basic techniques. That's the reason why performance could be slightly different from paper

[zhanghang1989]

Thanks @uvipen Could you the results in your experiments? I saw the README is still pending.