TensorDictBase.to_dict#

TensorDictBase.to_dict(*, retain_none: bool = True, convert_tensors: bool | Literal['numpy'] = False, tolist_first: bool = False) → dict[str, Any]#

Returns a nested dictionary with the entries of the tensordict.

Nested tensordicts and tensorclasses become nested dicts. The tensors are not copied unless convert_tensors is set. Non-tensor data is returned as its Python value.

Keyword Arguments:

retain_none (bool, optional) – if True, the None fields of tensorclasses are written in the dict as None. If False, they are dropped. Defaults to True.
convert_tensors (bool or str, optional) – if True, tensors are converted to (nested) Python lists with torch.Tensor.tolist(). If "numpy", they are converted to NumPy arrays. If False, they are kept as tensors. Defaults to False.
tolist_first (bool, optional) – if True, a nested tensordict is converted with tolist(), so that a nested tensordict with batch dimensions becomes a list of dicts, one per element. Defaults to False.

Returns:

A dict whose keys are those of the tensordict.

Raises:

RuntimeError – on a lazy stack whose tensordicts have different keys (see Variable-size data).

See also

tolist(), numpy() and from_dict(), which converts back.

Examples

>>> import torch
>>> from tensordict import TensorDict
>>>
>>> td = TensorDict(
...     a=torch.arange(6).view(2, 3),
...     b=TensorDict(c=torch.arange(4).reshape(2, 2), batch_size=(2, 2)),
...     batch_size=(2,)
... )
>>> print(td.to_dict())
{'a': tensor([[0, 1, 2],
        [3, 4, 5]]), 'b': {'c': tensor([[0, 1],
        [2, 3]])}}
>>> print(td.to_dict(convert_tensors=True))
{'a': [[0, 1, 2], [3, 4, 5]], 'b': {'c': [[0, 1], [2, 3]]}}

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

TensorDictBase.to_dict#

Docs

Tutorials

Resources