The problem of the top 1 acc

[plzhang4321]

what means "it may be in std" . Your awesome implementation in pytorch, but I cnnnot understand its big drop~

[senya-ashukha]

Hi,

Thank you for the comment!

It means that results can differ depending on a random seed. This is an expensive setup that takes 16v100 and 170h, and we tried it only once. Also, there is a mismatch in default PyTorch and TF augmentation of ImageNet images; in PyTorch, to the best of my knowledge, there is no min_object_covered parameter https://github.com/google-research/simclr/blob/3fb622131d1b6dee76d0d5f6aac67db84dab3800/data_util.py#L269.

Let us know if you have any further questions!

[plzhang4321]

Thank you~