training hyperparameter combinations

[lswzjuer]

Timm provides a lot of pre-trained models and corresponding accuracies, but in most cases I can't train to the accuracies in the list.
Can you provide the training hyperparameter combinations corresponding to the models included in the "results/results-imagenet.csv" list?

[rwightman]

@lswzjuer those are not generally available for all models, no. It's additional work to maintain and keep those relevant, and answer all the questiosn why the accuracies can't be reproduced when they're using this recipe or that... so it hasn't been a focus.

I'm going to try to improve that at some point, but the best resource right now is ResNet Strikes Back paper (https://arxiv.org/abs/2110.00476) which outlines all major recipes that I use in enough detail to adapt and try with different model architectures... you can also look at the recipes in Swin and DeiT repositories and papers which are based on timm and they are directly translateable here and work well on other models.

Moving to discussions.