ResNet Strikes Back

[amaarora]

Hey @rwightman! Firstly, congratulations on the wonderful paper! Really enjoy seeing papers from you!

Have a quick question - for Figure-1 of the paper, with ImageNet-Val and ImageNet-V2 analysis on 100 seeds, which of the ImageNet-V2 datasets did you use?

I see three options:

1. imagenetv2-matched-frequency
2. imagenetv2-threshold0.7
3. imagenetv2-topimages

Also, will the validate.py script be enough to replicate the chart? I've added random_seed to validate.py similar to train.py.

Thank you!

[rwightman]

@amaarora it's matched-frequency for v2, I've rarely seen the others used.

Nothing in the validation process relies on seeds, it's deterministic with the exception of possible cudnn layer algorithm selection.

100 seeds means 100 different model weights were trained each with a different training seed (different initial weights and different sequence of augmentations / sample selections). Each of those resulting models was run through 1k-val and v2