vita-epfl / ttt-plus-plus

[NeurIPS21] TTT++: When Does Self-supervised Test-time Training Fail or Thrive?
MIT License
56 stars 6 forks source link

Clarification on how test time accuracy is calculated #6

Closed devavratTomar closed 2 years ago

devavratTomar commented 2 years ago

Dear Authors, Many thanks for the great work. While going through the paper I could not understand how you compute the final accuracy on the test samples. Is the accuracy computed in an online manner (evaluate the model on the current batch as it trains on) or offline manner (evaluate the whole test set after training is finished)?

Thanks.

YuejiangLIU commented 2 years ago

Thanks for your interest! We evaluated TTA methods offline, i.e., the accuracy of the whole test set after adaptation. Please refer to the following lines of our code: https://github.com/vita-epfl/ttt-plus-plus/blob/e429a194ec3dbccfdd63b8b85c417f69830e7002/cifar/ttt%2B%2B.py#L327 https://github.com/vita-epfl/ttt-plus-plus/blob/e429a194ec3dbccfdd63b8b85c417f69830e7002/cifar/shot.py#L177

devavratTomar commented 2 years ago

Thank you!