Learning Test Time Augmentation with Cascade Loss Prediction

Siyang Pan, Jiaqian Yu, Dongwook Lee, Qiang Wang, ChangBeom Park, ByungIn Yoo

Abstract: Data augmentation has been a successful common practice for improving the performance of deep neural network during training stage. In recent years, studies on test time augmentation (TTA) have also been promising due to its effectiveness on improving the robustness against out-of-distribution data at inference. Instead of simply adopting pre-defined handcrafted geometric operations such as croping and flipping, recent TTA methods learn predictive transformations which are supposed to provide the best performance gain on each test sample. However, the desired iteration number of transformation is proportional to the inference time of the predictor, and the gain by ensembling multiple augmented inputs still requires additional forward pass of the target model. In this paper, we propose a cascade method for test time augmentation prediction. It only requires a single forward pass of the transformation predictor, while can output multiple desirable transformations iteratively. These transformations will then be adopted sequentially on the test sample at once before the target model inference. The experimental results show that our method provides a better trade-off between computational cost and overall performance at test time, and shows significant improvement compared to existing methods.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)