Skip to main content

Weighted Sampler

If 99% of your data is one class, random batches teach the model just one trick: always guess that class.

Key Insight​

A sampler decides the order in which a DataLoader visits examples. A WeightedRandomSampler gives each example its own sampling probability, so you can draw rare classes more often and build balanced batches from an imbalanced dataset.

Why This Matters​

On imbalanced data, a model can reach high accuracy by always predicting the majority class while learning nothing useful. Balanced sampling forces the model to see minority classes often enough to actually learn them.