I tried your autoaugment policies to my image-classification learning(imagenet-pretrained fine tuning).

[bemoregt]

Hi, @DeepVoltaire

I tried your autoaugment policies to my image-classification learning(imagenet-pretrained fine tuning).

But, accuracy is not good than default augmentation.

What's wrong to me?

Autoaugmentation is not suitable for transfer learning or finetuning?

Thanks at any rate.

[Jongchan]

@bemoregt
It seems that you are fine-tuning an ImageNet-pretrained model on ImageNet dataset. But AutoAugment is never validated on transfer learning situations. Only training from scratch.
I've done training from scratch, and this repository's implementation indeed outperforms the baseline if trained from scratch.

Possibly, you need longer epochs to train, or just use AutoAugment when you train from scratch.