Comments on ImageNet

[Equationliu]

Does this method scale to ImageNet?

Hope for your reply~

[okyksl]

Hi, thanks for your question.

If you have in mind 224 x 224 as resolution, I believe you would need huge amounts of GPU memory to apply this method. If you decrease to 64 x 64 and use a large-scale subset of ImageNet (compared to CIFAR-10), I am guessing it would be again costly due to the training time.

This is considering that you don't have access to multiple high-end GPUs (e.g., A100s). Then, I don't have a good answer.