stadlmax / Graph-Posterior-Network

Graph Posterior Network: Bayesian Predictive Uncertainty for Node Classification (NeurIPS 2021)
https://www.daml.in.tum.de/graph-postnet
MIT License
39 stars 12 forks source link

Epistemic uncertainty with and without effects #4

Open vietngth opened 1 year ago

vietngth commented 1 year ago

Hi, I ran the experiments for ood_locgpn{10,16} to reproduce the results. The metrics are logged after training as follows:

val test
accuracy 0.923077 0.889306
brier_score 0.280527 0.269254
ECE 0.10638 0.0702577
confidence_aleatoric_apr 0.966722 0.980251
confidence_epistemic_apr 0.95776 0.962611
confidence_structure_apr nan nan
confidence_aleatoric_auroc 0.706276 0.864282
confidence_epistemic_auroc 0.635031 0.772384
confidence_structure_auroc nan nan
CE 0.37979 0.339846
avg_prediction_confidence_aleatoric 0.816697 0.819579
avg_prediction_confidence_epistemic 14919.8 14410.5
avg_sample_confidence_aleatoric 0.816697 0.819579
avg_sample_confidence_epistemic 16556.9 15927.3
avg_sample_confidence_features 15835.7 16779.6
avg_sample_confidence_neighborhood nan nan
average_entropy 0.513569 0.501767
ood_detection_aleatoric_apr 0.801985 0.78415
ood_detection_aleatoric_auroc 0.833289 0.836685
ood_detection_epistemic_apr 0.80048 0.788122
ood_detection_epistemic_auroc 0.852889 0.855747
ood_detection_features_apr 0.75896 0.748484
ood_detection_features_auroc 0.81794 0.8237
ood_detection_neighborhood_apr nan nan
ood_detection_neighborhood_auroc nan nan
ood_detection_structure_apr nan nan
ood_detection_structure_auroc nan nan
ood_accuracy 0 0
ood_avg_prediction_confidence_aleatoric 0.556919 0.558297
ood_avg_prediction_confidence_epistemic 2879.92 2808.5
ood_avg_sample_confidence_aleatoric 0.556919 0.558297
ood_avg_sample_confidence_epistemic 4068.24 4109.26
ood_avg_sample_confidence_neighborhood nan nan
ood_avg_sample_confidence_features 2063.81 1759.3
ood_average_entropy 1.03177 1.03767
id_accuracy 0.923077 0.889306
id_avg_prediction_confidence_aleatoric 0.816697 0.819579
id_avg_prediction_confidence_epistemic 14919.8 14410.5
id_avg_sample_confidence_aleatoric 0.816697 0.819579
id_avg_sample_confidence_epistemic 16556.9 15927.3
id_avg_sample_confidence_features 15835.7 16779.6
id_average_entropy 0.513569 0.501767

Is this experiment exclusive for the With Effect case? I would like to reproduce 3 cases: aleatoric w/Net, Epist w/Net and Epist w/o Net as table 1 in the paper. Could you provide me with some guidance? Thank you.