rdangovs / essl Goto Github PK

Thx for sharing amazing work and codesets in advance while I have a single question on the code.
In 'eval_loop' function of 'main.py', bicubic interpolations are used for data preprocessing on both train and test sets.
Is there any background for this? Why bicubic interpolations are used (even upscaling for train data processing)?

Hi there! Thanks for sharing the codebase. Can you please provide the pretrained checkpoints of RN18 on cifar10 & cifar100 for the 800ep?

Thanks

Hi,
Very nice work. I have a small question regarding to appendix H. In there you stated that It is problematic if x and T_g(x) exist in the data. Why is it problematic exactly in a sense that the labels are different due to the orientation and the change in the representational space due to the equivariance should help the downstream task.