배치 구성, 섞기와 num_workers

A dataset hands over one sample at a time, but training wants groups. The DataLoader wraps your dataset and serves it in convenient batches. 📦

Set batch_size and the loader stacks that many samples into one tensor. Bigger batches use your hardware better and smooth out noisy updates.

Each batch adds a new first dimension. Thirty-two samples of shape 784 become a single tensor shaped 32 by 784, ready for the model.

You iterate the loader like any Python sequence. Each turn of the loop yields one batch of inputs and labels for your training step.

Setting shuffle to True reorders samples each epoch. This breaks accidental ordering so the model cannot memorize the sequence of your data.

Turn shuffling on for the training set but off for validation and test. Evaluation just measures performance, so a stable order is fine there.

Reading and decoding data can stall the GPU. Setting num_workers above zero spawns helper processes that prepare the next batch while the model trains.

A common start for num_workers is the number of CPU cores you have. Too many can thrash memory, so measure rather than guess blindly.

When training on a GPU, set pin_memory to True. It places batches in page-locked memory so transfers to the device run noticeably faster.

The final batch is often smaller than the rest. Use drop_last True to discard it when your model needs every batch the same size.

In practice you build a separate loader for train, validation, and test. Each gets its own settings, like shuffle on only for training.

A DataLoader batches your dataset, shuffles training data, and uses num_workers to load batches in parallel. It keeps your model fed and fast. 🎉