Gradient Estimation Methods

Baselines

class storch.method.baseline.Baseline[source]

Bases: ABC, Module

abstract compute_baseline(tensor: StochasticTensor, cost_node: CostTensor) Tensor[source]
training: bool
class storch.method.baseline.BatchAverageBaseline[source]

Bases: Baseline

Uses the average over the other samples as baseline. Introduced by https://arxiv.org/abs/1602.06725

compute_baseline(tensor: StochasticTensor, costs: CostTensor) Tensor[source]
training: bool
class storch.method.baseline.MovingAverageBaseline(exponential_decay=0.95)[source]

Bases: Baseline

Takes the (unconditional) average over the different costs.

compute_baseline(tensor: StochasticTensor, cost_node: CostTensor) Tensor[source]
training: bool

Multi-sample estimators

RELAX

Unordered set estimator