bambezius/dissertation-tgn

tgn

JAX checkpoint for the Tgn temporal graph model with separate intention and outcome heads.

Checkpoint

• Model class: Tgn
• Framework: jax
• Run timestamp: 20260430_135010
• Parameters: model_params.joblib
• Hyperparameters: config.json
• Metrics: metrics.json

Validation Metrics

{
  "val/loss": 1.1337319612503052,
  "val/intention_top1_accuracy": 0.6611314415931702,
  "val/intention_top2_accuracy": 0.8483576774597168,
  "val/intention_top3_accuracy": 0.9177007675170898,
  "val/intention_mrr": 0.7940667867660522,
  "val/outcome_top1_accuracy": 0.5961679220199585,
  "val/outcome_top2_accuracy": 0.7636861801147461,
  "val/outcome_top3_accuracy": 0.8319343328475952,
  "val/outcome_mrr": 0.7304578423500061
}

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")