TensorDictBase.to_namedtuple#

TensorDictBase.to_namedtuple(dest_cls: type | None = None) → Any#

Converts the tensordict to a named tuple, recursively.

Each nested tensordict becomes a nested named tuple. Without dest_cls, each level is an instance of a new named tuple class, GenericDict, whose fields are the keys. The None fields of tensorclasses are dropped.

On a tensorclass, the top level is a plain tuple, also with dest_cls: it holds the values of to_dict(retain_none=False), in the same order, without the field names. The nested tensordicts and tensorclasses in it are still named tuples. A plain tuple of tensors can be pickled and loaded with torch.load(weights_only=True), while a GenericDict cannot be pickled, since its class is made at each call. To keep the field names, call to_dict(), or to_tensordict().to_namedtuple().

Parameters:: dest_cls (type, optional) – the named tuple class of the result. A nested tensordict is converted to dest_cls too if dest_cls takes its keys, and to a GenericDict otherwise. Defaults to None.

See also

from_namedtuple(), which converts back.

Examples

>>> from tensordict import TensorDict
>>> import torch
>>> data = TensorDict({
...     "a_tensor": torch.zeros((3)),
...     "nested": {"a_tensor": torch.zeros((3)), "a_string": "zero!"}}, [3])
>>> data.to_namedtuple()
GenericDict(a_tensor=tensor([0., 0., 0.]), nested=GenericDict(a_tensor=tensor([0., 0., 0.]), a_string='zero!'))

On a tensorclass, the top level is a plain tuple:

>>> from tensordict import tensorclass
>>> @tensorclass
... class Data:
...     a_tensor: torch.Tensor
...     a_string: str
>>> Data(a_tensor=torch.zeros(3), a_string="zero!", batch_size=[3]).to_namedtuple()
(tensor([0., 0., 0.]), 'zero!')

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

TensorDictBase.to_namedtuple#

Docs

Tutorials

Resources