Train test validation splits and augmentation

Question

I am dealing with an image dataset of 400, and split it into 70% train, 15%test, 15%validation. I would like to do some data augmentation (rotations/flips) to increase the amount of train data I have directly in tensor flow. I read that you typically only augment the training set. If I do this, my test and validation data would be much smaller than my train. Is that fine? Also, just to provide some context, I plan to train an end-to-end cnn as a binary classifier.

I'm an hs student who is really new to machine learning, and would appreciate any advice! I have a deadline in like an hour and am panicking a little heheh.

Answer 1

If I do this, my test and validation data would be much smaller than my train. Is that fine?

It's fine.

The main thing you need to have in a train/test/validation split is a holdout that's large enough to give the desired level of precision. A statistic measured on a holdout set is a statistic, so it's subject to random variation. Too small a partition is associated with a larger variance in the estimator of whatever statistic you're interested in.

When the holdout partition is too small to give reasonable estimates, cross-validation can be used instead.