2ba596643cbbbc20318224181fa46b28-AuthorFeedback.pdf

We trained our loss predictor for five crop areas. We trained our loss predictor on the searched GPS policies to choose ones specific foreachtestinstance. Ourmethodproperlyselects valid transforms from the candidates chosen greedily by GPS, and therefore further improves the performances over static ensemble fromGPS. As [17] shows, manually targeted30 image restoration can be harmful to robustness when the corruption of each test image is unknown at test-time. Wewillupdatethe36 experimental results with various train-time augmentations and more baselines, and revise the manuscript toreflect37 youradditionalcomments.