TensorDictBase.pad#

TensorDictBase.pad(pad_size: Sequence[int], value: float = 0.0, inplace: bool = False, safe: bool = True) → Self#

Pads this tensordict along the batch dimensions with a constant value.

Parameters:

pad_size (Sequence[int]) – The padding size, applied to the batch dimensions starting from the first. For a tensordict of ndim batch dimensions, pad_size is a sequence of even length up to 2 * ndim formatted as (dim0_left, dim0_right, dim1_left, dim1_right, ...).
value (float, optional) – The fill value. Defaults to 0.0.
inplace (bool, optional) –
If True, this tensordict’s object identity and key set are preserved; each leaf storage is replaced by its padded counterpart one leaf at a time, keeping peak memory close to 1x instead of 2x. The leaf tensors themselves are still freshly allocated because padding necessarily grows shapes. Defaults to False.

Warning

If inplace=True and a later leaf’s pad fails (e.g. OOM), the tensordict is left in an inconsistent state. safe=True (the default) catches the user-error class of failures before any mutation occurs.
safe (bool, optional) – If True, validate that the operation would succeed for every leaf before any mutation occurs. Set to False to skip the pre-flight walk when the inputs are known to be valid. Defaults to True.

Returns:

The padded tensordict. When inplace=True this is self.

Note

On a lazy stack whose tensordicts have no names, the result of inplace=False has no names, even when the stack dimension has a name. See pad().

Examples

>>> import torch
>>> from tensordict import TensorDict
>>> td = TensorDict({"a": torch.zeros(3, 4)}, batch_size=[3, 4])
>>> td.pad([0, 0, 0, 1]).batch_size
torch.Size([3, 5])
>>> td.pad([0, 0, 0, 1], inplace=True) is td
True
>>> td.batch_size
torch.Size([3, 5])

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

TensorDictBase.pad#

Docs

Tutorials

Resources