How to tell which augmentations were used in pretrained weights?

[guydav]

Hi folks,

Thank you all (and especially Ross) for all of the work put into this library. I searched through the documentation and code before posting this question, but couldn't find an answer:

Is there a way to know (or infer) which augmentations were used to train each set of pretrained weights supplied in timm? I'm trying to look for models trained with and without particular augmentations (say, horizontal flipping), of pretty much any architecture, and I couldn't find a way to tell from the pretrained weights and code how the models were augmented in trained.

Thank you very much!

[rwightman]

@guydav unfortunately, that is not possible. For horizontal flipping specifically, I'm pretty sure almost every single weight was trained with that enabled. But for the others, dropout, stochastic depth, label smoothing, auto-augn/rand-aug, cutmix/mixup, random erasing, there is no metadata.

Where models came from academic paper releases, the hparams match those in the papers (ie tf_ prefixed efficientnets, deit, convnext, etc). But again, I feel horizontal flipping + the base RandomResizeCrop is used pretty much everywhere.

[guydav]

Thanks, @rwightman . That all makes sense. I understand my use case is pretty rare, so I'm not surprised it's not documented.

I wonder if there's any use case to documenting augmentations (and other training settings?) in the long term, to increase reproducibility, but I imagine you all would know better than I do if that's helpful.