BatchSampler

A batchsampler is a component used in data loading systems to produce batches of data indices. In PyTorch, the BatchSampler class (torch.utils.data.BatchSampler) wraps a base sampler that yields individual indices (such as RandomSampler or SubsetRandomSampler) and groups those indices into batches of a specified size. This allows customizing how data are batched without altering the underlying sampling strategy.

A batchsampler is initialized with a base sampler, a batch_size, and a drop_last flag. When iterated, it pulls indices from the underlying sampler and returns them as lists of length batch_size. If the remaining elements are fewer than batch_size and drop_last is False, the final batch is yielded with the smaller size; if drop_last is True, the final incomplete batch is discarded. The exact behavior regarding length can depend on the implementation, but the general contract is to produce fixed-size batches or a final potentially smaller batch as configured.

In the PyTorch data loading workflow, a DataLoader can be configured to use a batch_sampler directly. When batch_sampler is provided, the DataLoader ignores its own batch_size, shuffle, and drop_last settings, and uses the batches produced by the batch_sampler to fetch and collate data. This separation allows complex batching strategies to be combined with multi-process loading and custom collation.

Use cases for batch samplers include implementing specialized batching schemes, such as stratified or balanced batches, combining with distributed sampling to ensure proper data division across workers, and experiments with non-standard batch formation. Overall, batchsamplers give fine-grained control over how data indices are grouped into batches while integrating with the broader data-loading ecosystem.