#random-crop

Question 1: Single choice

For CIFAR-10 you build two datasets from the same images: the training set applies a random crop to every image each time it is loaded, while the validation set uses the images unchanged. Why is the augmentation left out of the validation set?

Validation must reflect the real, unaltered inputs; augmentation only regularizes training1
Random crops on validation images would leak the labels into the model's inputs2
Augmented validation images would make the validation loss drop below the training loss3
Data augmentation only works on tensors that require gradients, which validation data do not4

Timed test