Strabismus

Step 3 of 6

Splitting

The photos are divided 70/15/15 into training, validation and test sets before anything else touches them.

70 / 15 / 15

Three sets, chosen before anything else

The preprocessed photos were divided into training, validation and test sets in a 70:15:15 ratio. The split was made class by class, so every set keeps the same mix of the five classes.

360

training photos

76

validation photos

81

test photos

Training360Validation76Test81517 photos
one photo our esotropia photo, which lands in training

Three jobs

Why three sets and not two

A model can look good on photos it has studied and still fail on new ones. Keeping photos aside is how you measure what it actually learned.

Validation photos guide training while it runs. Test photos are never used for any decision, so the final score is a clean measure of how the model handles photos it has never seen.

  • 360TrainingThe only photos the models learn from.
  • 76ValidationChecked after every round of training, to decide when to slow down or stop.
  • 81TestLocked away until the very end. Every result on this site comes from these photos.

The order matters

Why we split before augmenting

The next step makes ten versions of every training photo. If we had made the copies first and split afterwards, a photo's flipped version could land in training while its brighter twin landed in the test set.

The model would then be tested on photos it had effectively already seen, and its score would look better than it really is. This is called data leakage. Splitting first keeps the test set honest: none of its photos, or any version of them, is ever used for training.

Split first, then copy

What we did

Training

AA flippedA brighterA grey

Test

B

Photo B stays unseen. The test is fair.

Copy first, then split

What we avoided

Training

AA flippedA grey

Test

A brighter

Leak: the test holds a near-copy of a training photo, so the score comes out too high.