02.06 · UNIT 02 · PyTorch in One Hour: models and training · Lesson

Data loaders

A Dataset defines one record. A DataLoader shuffles records and groups them into batches.

PLAIN-LANGUAGE INTRODUCTION

What is this?

A Dataset defines one record. A DataLoader shuffles records and groups them into batches.

One simple example

A loader with batch_size=2 over 5 examples gives batches of 2, 2, and 1.

Feature and label tensors, plus a batch size and a shuffle choice.

One batch of features and labels per loop step.

Separating record lookup from batching keeps both jobs simple and reusable.

drop_last=True on a test loader removes real examples from evaluation.

WORK THROUGH THE IDEA

A custom Dataset needs __init__, __getitem__, and __len__.
__getitem__ returns one record. __len__ returns the number of records.
shuffle=True mixes the order before each epoch. One complete pass is one epoch.
drop_last=True drops a short final batch. num_workers greater than 0 loads data in parallel processes.
Common mistake: labels must start at 0 and use torch.long. Keep features as floats.