about augmentation setting

[jcy132]

1. Any reason for changing the augmentation setting from the original code?
   (for example in imagenet-R, crop range(0.08, 1.0) instead of (0.05, 1.0) in original code )

2. In the implemented code, same augmentation is applied for cifar100 and imagenet-R for train, which is different with origianl code.
   Maybe the bug?

[JH-LEE-KR]

Hi,
thanks for your comment.

1. It is my mistake.
   I failed to check the value of the original code and used the default value of torchvision.transforms.RandomResizedCrop.
   I'll modify this part.

2. What is the difference? It is the same as I understand it.
   Both CIFAR-100 and ImageNet-R use preprocess.train_preprocess for train augmentation.
   Let me know if you find anything the difference between them.

Best,
Jaeho Lee.

[gqk]

Hi, Jaeho,

I found the inference interpolation algorithm is different from the Jax code. The performance on CIFAR100 drops about 1.5% when changing it from BICUBIC to BILINEAR (i.e., 3 -> 2). What about the results on your machine?

dualprompt-pytorch/datasets.py

Line 211 in 2c15502

transforms.Resize(size, interpolation=3), # to maintain same ratio w.r.t. 224 images