:orphan:

:mod:`torchfilter.train._train_dynamics`
========================================

.. py:module:: torchfilter.train._train_dynamics

.. autoapi-nested-parse::

   Private module; avoid importing from directly.



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


Functions
~~~~~~~~~

.. autoapisummary::

   torchfilter.train._train_dynamics.train_dynamics_single_step
   torchfilter.train._train_dynamics.train_dynamics_recurrent


.. function:: train_dynamics_single_step(buddy: fannypack.utils.Buddy, dynamics_model: torchfilter.base.DynamicsModel, dataloader: DataLoader, *, loss_function: str = 'nll', log_interval: int = 10) -> None

   Optimizes a dynamics model's single-step prediction accuracy. This is roughly
   equivalent to training with ``train_dynamics_recurrent()`` with a subsequence length
   of 2.

   :param buddy: Training helper.
   :type buddy: fannypack.utils.Buddy
   :param dynamics_model: Model to train.
   :type dynamics_model: torchfilter.base.DynamicsModel
   :param dataloader: Loader for a SingleStepDataset.
   :type dataloader: DataLoader

   :keyword loss_function: Either "nll" for negative log-likelihood or "mse"
                           for mean-squared error. Defaults to "nll".
   :kwtype loss_function: str, optional
   :keyword log_interval: Minibatches between each Tensorboard log.
   :kwtype log_interval: int, optional


.. function:: train_dynamics_recurrent(buddy: fannypack.utils.Buddy, dynamics_model: torchfilter.base.DynamicsModel, dataloader: DataLoader, *, loss_function: str = 'nll', log_interval: int = 10) -> None

   Trains a dynamics model via backpropagation through time.

   :param buddy: Training helper.
   :type buddy: fannypack.utils.Buddy
   :param dynamics_model: Model to train.
   :type dynamics_model: torchfilter.base.DynamicsModel
   :param dataloader: Loader for a SubsequenceDataset.
   :type dataloader: DataLoader

   :keyword loss_function: Either "nll" for negative log-likelihood or "mse"
                           for mean-squared error. Defaults to "nll".
   :kwtype loss_function: str, optional
   :keyword log_interval: Minibatches between each Tensorboard log.
   :kwtype log_interval: int, optional