TensorDictBase.to_pandas#

TensorDictBase.to_pandas(*, separator: str | None = None)#

Converts the tensordict to a pandas DataFrame, with one column per leaf.

Each leaf must hold one value per row: a tensor of shape [num_rows], or non-tensor data. Tensors are copied to the CPU and converted to NumPy arrays. Non-tensor data becomes a column of Python values. Requires pandas.

A nested tensorclass gives one column per leaf with a separator, as a nested tensordict does. Without a separator, where a nested tensordict raises, a nested tensorclass becomes one column of objects that holds the tensorclass of each row. It also stays one such column with a separator if it has no leaves, or a leaf of more than one dimension.

Keyword Arguments:: separator (str, optional) – if given, joins the keys of nested entries with it to build the column names: with separator=".", the entry ("obs", "x") becomes the column "obs.x". Required when the tensordict has nested tensordicts. Defaults to None.
Returns:: A pd.DataFrame with one row per element of the tensordict.
Raises:: ValueError – if the tensordict has a nested tensordict and no separator, or if a tensor has more than one dimension.

Examples

>>> import torch
>>> from tensordict import TensorDict
>>> td = TensorDict({"a": torch.arange(3), "b": torch.zeros(3)}, [3])
>>> df = td.to_pandas()
>>> print(df)
   a    b
0  0  0.0
1  1  0.0
2  2  0.0

A nested tensorclass without a separator:

>>> from tensordict import tensorclass
>>> @tensorclass
... class Point:
...     x: torch.Tensor
...     y: torch.Tensor
>>> td = TensorDict(p=Point(x=torch.zeros(3), y=torch.ones(3), batch_size=[3]), batch_size=[3])
>>> df = td.to_pandas()
>>> df["p"].dtype, type(df["p"][0]).__name__, df["p"][0].y
(dtype('O'), 'Point', tensor(1.))
>>> list(td.to_pandas(separator=".").columns)
['p.x', 'p.y']

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

TensorDictBase.to_pandas#

Docs

Tutorials

Resources