search

kdiAAA / TDA

[CVPR 2024] Official Repository for "Efficient Test-Time Adaptation of Vision-Language Models"
https://kdiaaa.github.io/tda/
MIT License
58 stars 5 forks source link

Question About Data Augmentation #1

Closed Yushu-Li closed 6 months ago

Yushu-Li commented 6 months ago

Hello,

Firstly, I wanted to express my gratitude for your outstanding paper and for generously sharing your code.

Upon reviewing the TDA experiments on the OOD benchmark (A/V/R/S), I noticed that they run rather slowly. My understanding was that AugMix augmentation is employed with 63 views for each test image in the OOD datasets. However, it was mentioned in the paper that only original test images are utilized.

In the "utils.py" file, the "build_test_data_loader" function utilizes the "get_ood_preprocess" function for OOD datasets. aug_preprocess = AugMixAugmenter(base_transform, preprocess, n_views=63, augmix=True) This command appears to generate 63 views of augmentation for each individual test image.

I'd like to confirm if I've misunderstood something related to this part. Thank you very much.

Best, Yushu

kdiAAA commented 6 months ago

Hello Yushu,

Thank you for your interest in our work! For the OOD benchmark datasets (A/V/R/S), we used the same augmentations as described in TPT. We did not detail this in the paper as it’s not a primary contribution. Also, the paper does not specify that only original test images are used; the augmentations seen in "utils.py" are indeed part of our intended approach.

Hope this helps clarify your queries!

Best, Adilbek

Yushu-Li commented 6 months ago

Dear Adilbek,

Thank you for your prompt response and clarification. Your clarification has addressed my queries effectively.

Best, Yushu