bambezius/dissertation-tgn-sym

tgn-sym

JAX checkpoint for the TgnSym D2-invariant temporal graph model with separate intention and outcome heads.

Checkpoint

• Model class: TgnSym
• Framework: jax
• Run timestamp: 20260430_130357
• Parameters: model_params.joblib
• Hyperparameters: config.json
• Metrics: metrics.json

Validation Metrics

{
  "val/loss": 1.0408462285995483,
  "val/intention_top1_accuracy": 0.6916058659553528,
  "val/intention_top2_accuracy": 0.8633211851119995,
  "val/intention_top3_accuracy": 0.9248175621032715,
  "val/intention_mrr": 0.8131301403045654,
  "val/outcome_top1_accuracy": 0.6186131834983826,
  "val/outcome_top2_accuracy": 0.782299280166626,
  "val/outcome_top3_accuracy": 0.8479927182197571,
  "val/outcome_mrr": 0.7483379244804382
}

Loading

import json
from pathlib import Path

import joblib

checkpoint_path = Path("path/to/checkpoint")

with open(checkpoint_path / "config.json") as f:
    config = json.load(f)

params = joblib.load(checkpoint_path / "model_params.joblib")