gfn.gym.helpers.bayesian_structure.evaluation

The code is adapted from: https://github.com/larslorch/dibs/blob/master/dibs/metrics.py

Functions

expected_edges(posterior_samples)

Compute the expected number of edges.

expected_shd(posterior_samples, gt_graph)

Compute the Expected Structural Hamming Distance.

posterior_estimate(gflownet, env[, num_samples, ...])

Get the posterior estimate of DAG-GFlowNet as a collection of graphs

threshold_metrics(posterior_samples, gt_graph)

Compute threshold metrics (e.g. AUROC, Precision, Recall, etc...).

Module Contents

gfn.gym.helpers.bayesian_structure.evaluation.expected_edges(posterior_samples)

Compute the expected number of edges.

This function computes the expected number of edges in graphs sampled from the posterior approximation.

Parameters:

posterior_samples (torch.Tensor) – Samples from the posterior. The tensor must have size (B, N, N), where B is the number of sample graphs from the posterior approximation, and N is the number of variables in the graphs.

Returns:

The expected number of edges.

Return type:

e_edges

gfn.gym.helpers.bayesian_structure.evaluation.expected_shd(posterior_samples, gt_graph)

Compute the Expected Structural Hamming Distance.

This function computes the Expected SHD between a posterior approximation given as a collection of samples from the posterior, and the ground-truth graph used in the original data generation process.

Parameters:

  • posterior_samples (torch.Tensor) – Samples from the posterior. The tensor must have size (B, N, N), where B is the number of sample graphs from the posterior approximation, and N is the number of variables in the graphs.

  • gt_graph (torch_geometric.data.Data) – GeometricData instance representing the ground-truth graph.

Returns:

The Expected SHD.

Return type:

e_shd

gfn.gym.helpers.bayesian_structure.evaluation.posterior_estimate(gflownet, env, num_samples=1000, batch_size=100, verbose=True)

Get the posterior estimate of DAG-GFlowNet as a collection of graphs sampled from the GFlowNet.

Parameters:

  • gflownet (gfn.gflownet.GFlowNet) – GFlowNet instance.

  • env (gfn.gym.bayesian_structure.BayesianStructure) – BayesianStructure environment.

  • rng – Optional random generator instance.

  • num_samples – The number of samples in the posterior approximation.

  • verbose – If True, display a progress bar for the sampling process.

Returns:

torch.Tensor with shape (B, N, N), where B is the number of sample

graphs in the posterior approximation, and N is the number of variables in a graph.

Return type:

posterior

gfn.gym.helpers.bayesian_structure.evaluation.threshold_metrics(posterior_samples, gt_graph)

Compute threshold metrics (e.g. AUROC, Precision, Recall, etc…).

Parameters:

  • posterior_samples (torch.Tensor) – Samples from the posterior. The tensor must have size (B, N, N), where B is the number of sample graphs from the posterior approximation, and N is the number of variables in the graphs.

  • gt_graph (torch_geometric.data.Data) – GeometricData instance representing the ground-truth graph.

Returns:

  • False Positive Rate

  • True Positive Rate

  • Area Under the Receiver Operating Characteristic Curve

  • Precision

  • Recall

  • Area Under the Precision-Recall Curve

  • Average Precision

Return type:

A dictionary containing the following metrics