TensorDictBase.contiguous#

TensorDictBase.contiguous(*, canonical: bool = False, inplace: bool = False) → Self#

Returns a new tensordict with contiguous leaves.

Each leaf is replaced by torch.Tensor.contiguous(): a leaf that is already contiguous is shared with the tensordict, the others are copied.

Keyword Arguments:

canonical (bool, optional) – if True, a leaf whose strides are not the row-major strides of its shape is copied too, even if torch.Tensor.is_contiguous() returns True for it (as it can with dimensions of size 1). Defaults to False.
inplace (bool, optional) – if True, the tensordict keeps its identity and keys, each leaf that is not contiguous is replaced by a contiguous copy, and self is returned. See In-place operations. Defaults to False.

Note

On a LazyStackedTensorDict, a PersistentTensorDict (h5, zarr) or a store, the result is a TensorDict. If the tensordicts of a lazy stack have no names, the result has no names either, even when the stack dimension has a name (from stack_dim_name or the names setter). densify() and clone() keep that name. To keep it here, name the dimension again with refine_names().

Examples

>>> import torch
>>> from tensordict import TensorDict, lazy_stack
>>> td = TensorDict(a=torch.zeros(4, 3).T, batch_size=[3])
>>> td.is_contiguous(), td.contiguous().is_contiguous()
(False, True)

On a lazy stack, the result loses the name of the stack dimension:

>>> stack = lazy_stack(
...     [TensorDict(a=torch.zeros(3), batch_size=[3]) for _ in range(2)],
...     stack_dim_name="s",
... )
>>> stack.names, stack.contiguous().names
(['s', None], [None, None])
>>> stack.contiguous().refine_names("s", ...).names
['s', None]
>>> stack.densify().names
['s', None]

Every tensordict class and every tensorclass has this method. For the other methods, see Methods by task.

TensorDictBase.contiguous#

Docs

Tutorials

Resources