Exact timm Train arguments
I'm trying to reproduce the results based on the 'How to train your vit' paper, but I'm not getting good performance. Could you provide the exact timm train.py arguments?
I know this post is more than a year old, but I need the exact same thing. Idk how hard it would but I think training script arguments for each timm model would be a huge help for reproducibility purposes
@ryanramos it's not trained in timm ... it's as per https://arxiv.org/abs/2106.10270 and trained on google infra although the hparams can be translated.
These are the most recent/interesting timm trained vit models https://huggingface.co/collections/timm/searching-for-better-vit-baselines-663eb74f64f847d2f35a9c19
Related hparams https://gist.github.com/rwightman/fb37c339efd2334177ff99a8083ebbc4
It's a lot of time to fully match train hparams to models and keep them consistent across lib changes ... so tend to share a few key ones over time