This is the official PyTorch Implementation of "SoTTA: Robust Test-Time Adaptation on Noisy Data Streams (NeurIPS '23)" by Taesik Gong*, Yewon Kim*, Taeckyung Lee*, Sorn Chottananurak, and Sung-Ju Lee (* Equal contribution).
Hi, author,
Thanks for your code and paper.
I have a few questions about the evaluation protocol. Let's take CIFAR-10-C as an example, for the ''Near'' scenario, the online data stream should consists of the original target samples (i.e., Benign) of CIFAR-10-C and the CIFAR-100 test set (if I understand right).
So the model need to classify only the Benign samples or it has to both classify Benign samples and identify samples from CIFAR-100 test set?
And accordingly, the dimension of the classifier is unchanged or the the classifier is expanded to ''C+1''-dim? C for number of categories and 1 for anomaly detection.
Hopefully you can explain more about the new setting. Thx.
Hi, This is a great job! When I downloaded the dataset of CIFAR-10-C, it showed that "HTTP request sent, awaiting response... 403 FORBIDDEN", could you please send the datasets to me by email? Thanks.