linmult.models.LinT
===================

.. py:module:: linmult.models.LinT

.. autoapi-nested-parse::

   LinT: unimodal linear-complexity transformer.



Classes
-------

.. autoapisummary::

   linmult.models.LinT.LinT


Module Contents
---------------

.. py:class:: LinT(config: linmult.core.config.LinTConfig | str | pathlib.Path)

   Bases: :py:obj:`torch.nn.Module`


   Linear-complexity Transformer for a single input modality.

   Processes one time-series input through a projection + self-attention
   pipeline and applies configurable output heads.

   :param config: Configuration object or path to a YAML file.
   :type config: LinTConfig | str | Path

   Initialize LinT.

   :param config: Configuration object or path to a YAML
                  file. When a path is given the file is loaded with
                  :meth:`LinTConfig.from_yaml`.
   :type config: LinTConfig | str | Path


   .. py:method:: extra_repr() -> str

      Return the model name for identification in repr output.



   .. py:method:: forward(x: torch.Tensor, mask: torch.Tensor | None = None, name: str | None = None) -> dict[str, torch.Tensor]

      Run the forward pass.

      :param x: Input of shape ``(B, T, F)``. May also be a
                single-element list ``[tensor]``.
      :type x: torch.Tensor
      :param mask: Bool mask of shape ``(B, T)``.
                   True = valid timestep. A fully-False mask is treated as ``None``.
      :type mask: torch.Tensor, optional
      :param name: Key used for special-handling lookup (e.g.
                   weighted-sum of layer activations). May also be a single-element list.
      :type name: str, optional

      :returns:

                Mapping from head name to output tensor.
                    Shape is ``(B, output_dim)`` when ``time_dim_reducer`` is set,
                    otherwise ``(B, T, output_dim)``.
      :rtype: dict[str, torch.Tensor]