gfn.estimators

Module Contents

Classes

FunctionEstimator

Base class for function estimators.

LogEdgeFlowEstimator

Container for estimators $(s rightarrow s') mapsto log F(s rightarrow s')$.

LogStateFlowEstimator

Container for estimators $s mapsto log F(s)$.

LogZEstimator

Container for the estimator $log Z$.

LogitPBEstimator

Container for estimators $s mapsto (u(s' - a mid s'))_{a in mathbb{A}}$,

LogitPFEstimator

Container for estimators $s mapsto (u(s + a mid s))_{a in mathbb{A}}$,

Attributes

OutputTensor

class gfn.estimators.FunctionEstimator(env, module=None, output_dim=None, module_name=None, **nn_kwargs)

Bases: abc.ABC

Base class for function estimators.

Parameters

  • env (gfn.envs.Env) –

  • module (Optional[gfn.modules.GFNModule]) –

  • output_dim (Optional[int]) –

  • module_name (Optional[Literal[gfn.modules.NeuralNet, gfn.modules.Uniform, gfn.modules.Tabular, Zero]]) –

__call__(states)
Parameters

states (gfn.containers.States) –

Return type

OutputTensor

__repr__()

Return repr(self).

load_state_dict(state_dict)
Parameters

state_dict (dict) –

named_parameters()
Return type

dict

class gfn.estimators.LogEdgeFlowEstimator(env, module=None, module_name=None, **nn_kwargs)

Bases: FunctionEstimator

Container for estimators \((s \rightarrow s') \mapsto \log F(s \rightarrow s')\). The way it’s coded is a function \(s \mapsto (\log F(s \rightarrow (s + a)))_{a \in \mathbb{A}}\), where \(s+a\) is the state obtained by performing action \(a\) in state \(s\).

Parameters

  • env (gfn.envs.Env) –

  • module (Optional[gfn.modules.GFNModule]) –

  • module_name (Optional[Literal[gfn.modules.NeuralNet, gfn.modules.Uniform, gfn.modules.Tabular, Zero]]) –

class gfn.estimators.LogStateFlowEstimator(env, module=None, module_name=None, forward_looking=False, **nn_kwargs)

Bases: FunctionEstimator

Container for estimators \(s \mapsto \log F(s)\).

Parameters

  • env (gfn.envs.Env) –

  • module (Optional[gfn.modules.GFNModule]) –

  • module_name (Optional[Literal[gfn.modules.NeuralNet, gfn.modules.Uniform, gfn.modules.Tabular, Zero]]) –

__call__(states)
Parameters

states (gfn.containers.States) –

Return type

OutputTensor

class gfn.estimators.LogZEstimator(tensor)

Container for the estimator \(\log Z\).

Parameters

tensor (torchtyping.TensorType[0, float]) –

__repr__()

Return repr(self).

Return type

str

load_state_dict(state_dict)
Parameters

state_dict (dict) –

named_parameters()
Return type

dict

class gfn.estimators.LogitPBEstimator(env, module=None, module_name=None, **nn_kwargs)

Bases: FunctionEstimator

Container for estimators \(s \mapsto (u(s' - a \mid s'))_{a \in \mathbb{A}}\), such that \(P_B(s' - a \mid s') = \frac{e^{u(s' - a \mid s')}}{\sum_{a' \in \mathbb{A}} e^{u(s' - a' \mid s')}}\).

Parameters

  • env (gfn.envs.Env) –

  • module (Optional[gfn.modules.GFNModule]) –

  • module_name (Optional[Literal[gfn.modules.NeuralNet, gfn.modules.Uniform, gfn.modules.Tabular, Zero]]) –

class gfn.estimators.LogitPFEstimator(env, module=None, module_name=None, **nn_kwargs)

Bases: FunctionEstimator

Container for estimators \(s \mapsto (u(s + a \mid s))_{a \in \mathbb{A}}\), such that \(P_F(s + a \mid s) = \frac{e^{u(s + a \mid s)}}{\sum_{a' \in \mathbb{A}} e^{u(s + a' \mid s)}}\).

Parameters

  • env (gfn.envs.Env) –

  • module (Optional[gfn.modules.GFNModule]) –

  • module_name (Optional[Literal[gfn.modules.NeuralNet, gfn.modules.Uniform, gfn.modules.Tabular, Zero]]) –

gfn.estimators.OutputTensor