gfn.gym.bayesian_structure
==========================

.. py:module:: gfn.gym.bayesian_structure


Classes
-------

.. autoapisummary::

   gfn.gym.bayesian_structure.BayesianStructure


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

.. py:class:: BayesianStructure(n_nodes, state_evaluator, device = 'cpu', debug = False)

   Bases: :py:obj:`gfn.gym.graph_building.GraphBuilding`


   Environment for incrementally building a directed acyclic graph (DAG) for
   Bayesian structure learning (Deleu et al., 2022).

   The environment allows the following actions:
   - Adding edges between existing nodes with features
   - Terminating construction (EXIT)

   :param n_nodes: Number of nodes in the graph.
   :param state_evaluator: Callable that computes rewards for final states.
                           If None, uses default GCNConvEvaluator
   :param device_str: Device to run computations on ('cpu' or 'cuda')


   .. py:method:: backward_step(states, actions)

      Performs a backward step in the environment.

      :param states: The current states.
      :param actions: The actions to undo.

      :returns: The previous states.



   .. py:method:: is_action_valid(states, actions, backward = False)

      Check if actions are valid for the given states.

      :param states: Current graph states.
      :param actions: Actions to validate.
      :param backward: Whether this is a backward step.

      :returns: True if all actions are valid, False otherwise.



   .. py:method:: log_reward(final_states)

      The environment's reward given a state.
      This or log_reward must be implemented.

      :param final_states: A batch of final states.

      :returns: Tensor of shape "batch_shape" containing the rewards.
      :rtype: torch.Tensor



   .. py:method:: make_actions_class()

      Returns the GraphActions class for this environment.

      :returns: A type of a subclass of GraphActions with environment-specific
                functionalities.



   .. py:method:: make_random_states_tensor(batch_shape, conditions = None, device = None)

      Makes a batch of random DAG states with fixed number of nodes.

      :param batch_shape: Shape of the batch dimensions.
      :param conditions: Optional tensor of shape (*batch_shape, condition_dim) containing
                         condition vectors for conditional GFlowNets.
      :param device: The device to create the graph states on.

      :returns: A PyG Batch object containing random DAG states.



   .. py:method:: make_states_class()

      Creates a `GraphStates` class for this environment.



   .. py:attribute:: n_actions


   .. py:attribute:: n_nodes


   .. py:method:: reward(final_states)
      :abstractmethod:


      The environment's reward given a state.

      :param final_states: A batch of final states.

      :returns: A tensor of shape `(batch_size,)` containing the rewards.



   .. py:method:: step(states, actions)

      Performs a step in the environment.

      :param states: The current states.
      :param actions: The actions to take.

      :returns: The next states.