`gfn.losses.trajectory_balance`

Implementations of the [Trajectory Balance loss](https://arxiv.org/abs/2201.13259) and the [Log Partition Variance loss](https://arxiv.org/abs/2302.05446).

Module Contents

Classes

`LogPartitionVarianceLoss`
`TBParametrization`	$mathcal{O}_{PFZ} = mathcal{O}_1 times mathcal{O}_2 times mathcal{O}_3$, where
`TrajectoryBalance`

Attributes

`LossTensor`
`ScoresTensor`

class gfn.losses.trajectory_balance.LogPartitionVarianceLoss(parametrization, log_reward_clip_min=-12, on_policy=False)

Bases: gfn.losses.base.TrajectoryDecomposableLoss

Parameters

parametrization (gfn.losses.base.PFBasedParametrization) –
log_reward_clip_min (float) –
on_policy (bool) –

__call__(trajectories)

Parameters: trajectories (gfn.containers.Trajectories) –
Return type: LossTensor

gfn.losses.trajectory_balance.LossTensor

gfn.losses.trajectory_balance.ScoresTensor

class gfn.losses.trajectory_balance.TBParametrization

Bases: gfn.losses.base.PFBasedParametrization

$\mathcal{O}_{PFZ} = \mathcal{O}_1 \times \mathcal{O}_2 \times \mathcal{O}_3$, where $\mathcal{O}_1 = \mathbb{R}$ represents the possible values for logZ, and $\mathcal{O}_2$ is the set of forward probability functions consistent with the DAG. $\mathcal{O}_3$ is the set of backward probability functions consistent with the DAG, or a singleton thereof, if self.logit_PB is a fixed LogitPBEstimator. Useful for the Trajectory Balance Loss.

logZ :gfn.estimators.LogZEstimator

class gfn.losses.trajectory_balance.TrajectoryBalance(parametrization, log_reward_clip_min=-12, on_policy=False)