reference/generated/tensordict.nn.mappings

Run in Google Colab

Colab

Download Notebook

Notebook

View on GitHub

GitHub

mappings

tensordict.nn.mappings(key: str) → Callable

Returns the function named by key, which maps a tensor elementwise, for example to positive values.

Modules such as NormalParamExtractor use it to turn the output of a network into a scale. All the functions except "none" (the identity) map real numbers to non-negative numbers. "relu" can return 0, the others return positive numbers.

Parameters:

key (str) – one of "softplus", "exp", "relu", "expln" (see expln()), "biased_softplus" (a biased_softplus with bias=1.0), "none", or a name added with add_custom_mapping(). "biased_softplus_{bias}" and "biased_softplus_{bias}_{min_val}" return a biased_softplus with these arguments, for example "biased_softplus_2.0_0.1".

Returns:

A callable that takes a tensor and returns a tensor.

Raises:

• NotImplementedError – if key is not a known name.

• ValueError – if a "biased_softplus_..." key does not have the form "biased_softplus_{bias}" or "biased_softplus_{bias}_{min_val}".

Examples

>>> import torch
>>> from tensordict.nn import mappings
>>> mappings("exp")(torch.zeros(2))
tensor([1., 1.])
>>> mappings("biased_softplus_2.0_0.1")(torch.zeros(2))
tensor([2., 2.])