bambezius/dissertation-tgn-sym-dynamic-pc-cnn

tgn-sym-dynamic-pc-cnn

JAX checkpoint for the TgnSymDynamicPcCnn D2-invariant temporal graph model with pitch-control CNN fusion and separate intention and outcome heads.

Checkpoint

• Model class: TgnSymDynamicPcCnn
• Framework: jax
• Run timestamp: 20260430_160136
• Parameters: model_params.joblib
• Hyperparameters: config.json
• Metrics: metrics.json

Validation Metrics

{
  "val/loss": 0.9380925893783569,
  "val/intention_top1_accuracy": 0.7312043905258179,
  "val/intention_top2_accuracy": 0.8857664465904236,
  "val/intention_top3_accuracy": 0.9348540306091309,
  "val/intention_mrr": 0.8379348516464233,
  "val/outcome_top1_accuracy": 0.6596715450286865,
  "val/outcome_top2_accuracy": 0.8060219287872314,
  "val/outcome_top3_accuracy": 0.8671533465385437,
  "val/outcome_mrr": 0.7761701941490173
}

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")