Test vs Validation

Dev set or validation set:

Same data as train, use in train model, maybe compare two different models' performance

Test set:

Same or different data as train, use in fine-tuning (after train model) unbiased estimate performance of model (can skip test set)

Mismatch train & test: cropped photo vs photo in the wild (low resolution, weird angle)

How to split data:

	Train	Dev	Test
Small dataset	60% (6,000)	20% (2,000)	20% (2,000)
Big dataset	98% (1,000,000)	1% (10,000)	1% (10,000)