default_collate

→Tensor or list or dict or tuple

default_collate(batch: list[object])

source

Collate a list of samples into a batched tensor or nested structure.

The default collate_fn used by DataLoader. Walks the first element of batch to decide how to combine the remaining elements, preserving the original nested container structure.

Parameters

batchlist of object

Samples (one per dataset item) to combine into a single batch. Every element must share the same structure / type as batch[0].

Returns

Tensor or list or dict or tuple

Tensor leaves → stacked along a new leading axis.
np.ndarray leaves → stacked then wrapped as Tensor (numpy bridge — DataLoader is an H4 carve-out).
int / float leaves → 1-D Tensor.
str / bytes leaves → kept as a Python list.
dict / list / tuple / NamedTuple containers → walked recursively; original container type is preserved.

Notes

The recursion is structural: a batch of {"x": Tensor, "y": int} becomes {"x": stacked_Tensor, "y": 1d_Tensor} of the same dict shape. Heterogeneous batches (different keys / shapes) are not supported — callers must normalise upstream.

Examples

>>> default_collate([(t1, 0), (t2, 1), (t3, 2)])
(<Tensor shape=(3, ...)>, <Tensor shape=(3,)>)