How to train the cifar data using this repo?

[Ema1997]

Hi rwightman,
your repo is such a splendid and comprehensive work, and I am wondering how to modify the train.py in your codes to train these models on cifar10 dataset, or are there any pretrained checkpoints for cifar10? I found your excellent result on https://paperswithcode.com/sota/image-classification-on-cifar-10 and I am trying to re-implement your result based on your repo.

Thanks,
Ema1997.

[rwightman]

@Ema1997 this collection is focused on imagenet and larger datasets and image sizes. There are no cifar specific models, checkpoints, or training code. The default model cfgs are all rather large for cifar. However, lots of people do apply these models to cifar by scaling up the images during train/eval.

[Ema1997]

Ok, thanks. I would try to modify the images modules to fit in the cifar data.

[rwightman]

@Ema1997 I'm pretty sure the default data pipeline will work as is, it resizes short edge to the network target size and crops. Normally it'd be downscaling for ImageNet, but it'd just upscale in this case. You probably just need to plop the cifar10/100 dataset into the train script (there's on in torchvision) and then use --num-classes 10 (or 100) in the args and maybe disable the random random cropping, or limit it

[mrT23]

@Ema1997
this code:
https://github.com/Alibaba-MIIL/ImageNet21K/blob/main/Transfer_learning.md
achieves using timm top results on CIFAR-100, and should work as well on CIFAR-10

[Ema1997]

Thank you all for your sincere answers. : )