Any place I can find the exact training method/configuration used for the pretrained models?

[rezafuru]

I sincerely apologize, if there is an obvious place to look for them, but after spending some time I still cannot find any documentation on the training methodology used on the pretrained models. Specifically, I'm interested in the training configurations for the results obtained in https://github.com/rwightman/pytorch-image-models/tree/main/results.

I'd greatly appreciate if someone could point me in the direction.

[rwightman]

@rezafuru the checkpoints come from a WIDE variety of sources, even for my own training I didn't link to specific hparams. Keeping that sort of info synced and up to date is more than a 1 person job.

It's a two edge sword to do so, for every person that understand any set of hparam needs adjustments on different GPU counts, etc there are more that don't get this and file complaint issues "cannot reproduce this result exactly on my .. blah blah'

As I update the various models to support the new multi-weight pretrained configs, I am trying to add tags that differentiate some of the models a bit better by training type.. I have some internal 'codes' for my hparam sets that will be put in a glossary.

All that said, there are templates for all of the hparams I typically use in the appendix of this paper, https://arxiv.org/abs/2110.00476

[rwightman]

Some recent examples here: https://gist.github.com/rwightman/943c0fe59293b44024bbd2d5d23e6303

[rezafuru]

Right, makes sense. Thank you for your quick answer, this already helps me a lot!