What's the difference between resnet50_a1_0 and resnet50_a1h ?

[laisimiao]

What's the difference between resnet50_a1_0 and resnet50_a1h in releases?

[rwightman]

@laisimiao a1h was a training run after the initial ver of the paper, it had increased augmentation and regularization relative to the seed 0 a1 (more dropout, 3 rand augmetn layers instead of 2) that's in the main paper tables. It lowered the 1k val score a bit (but still not far off the mean result across seeds), but appeared a bit better on some of the OOD test sets...