Why is the performance of resnet50 higher than tv_resnet50 so much?

[lartpang]

From https://github.com/rwightman/pytorch-image-models/blob/master/results/results-imagenet.csv, we can see that:

model	top1	top1_err	top5	top5_err	param_count	img_size	cropt_pct	interpolation
tv_resnet50	76.138	23.862	92.864	7.136	25.56	224	0.875	bilinear
resnet50	79.038	20.962	94.390	5.610	25.56	224	0.875	bicubic

Why is the performance of resnet50 higher than tv_resnet50 so much?
The structure of the two should be the same. Is the training strategy different?

[rwightman]

@lartpang yes, completely different training strategy. I'm at 79.98 (Acc@1 79.982 (20.018) Acc@5 94.838 (5.162)) now with a base resnet50 but haven't updated the weights as I'm trying to crack 80 and release a technical report w/ details.

[lartpang]

Thanks for your extremely good job!
I was surprised when I realized that the resnet50 provided by timm is so good. 😀