update

example/config_no_temporal.json

@@ -0,0 +1,44 @@
+{
+  "data_path": "../../dataset/hai/hai-21.03/train1.csv.gz",
+  "data_glob": "../../dataset/hai/hai-21.03/train*.csv.gz",
+  "split_path": "./feature_split.json",
+  "stats_path": "./results/cont_stats.json",
+  "vocab_path": "./results/disc_vocab.json",
+  "out_dir": "./results",
+  "device": "auto",
+  "timesteps": 600,
+  "batch_size": 128,
+  "seq_len": 128,
+  "epochs": 10,
+  "max_batches": 4000,
+  "lambda": 0.7,
+  "lr": 0.0005,
+  "seed": 1337,
+  "log_every": 10,
+  "ckpt_every": 50,
+  "ema_decay": 0.999,
+  "use_ema": true,
+  "clip_k": 5.0,
+  "grad_clip": 1.0,
+  "use_condition": true,
+  "condition_type": "file_id",
+  "cond_dim": 32,
+  "use_tanh_eps": false,
+  "eps_scale": 1.0,
+  "model_time_dim": 128,
+  "model_hidden_dim": 512,
+  "model_num_layers": 2,
+  "model_dropout": 0.1,
+  "model_ff_mult": 2,
+  "model_pos_dim": 64,
+  "model_use_pos_embed": true,
+  "disc_mask_scale": 0.9,
+  "cont_loss_weighting": "inv_std",
+  "cont_loss_eps": 1e-6,
+  "cont_target": "x0",
+  "cont_clamp_x0": 5.0,
+  "shuffle_buffer": 1024,
+  "use_temporal_stage1": false,
+  "sample_batch_size": 8,
+  "sample_seq_len": 128
+}

example/config_temporal_strong.json

@@ -0,0 +1,49 @@
+{
+  "data_path": "../../dataset/hai/hai-21.03/train1.csv.gz",
+  "data_glob": "../../dataset/hai/hai-21.03/train*.csv.gz",
+  "split_path": "./feature_split.json",
+  "stats_path": "./results/cont_stats.json",
+  "vocab_path": "./results/disc_vocab.json",
+  "out_dir": "./results",
+  "device": "auto",
+  "timesteps": 600,
+  "batch_size": 128,
+  "seq_len": 128,
+  "epochs": 10,
+  "max_batches": 4000,
+  "lambda": 0.7,
+  "lr": 0.0005,
+  "seed": 1337,
+  "log_every": 10,
+  "ckpt_every": 50,
+  "ema_decay": 0.999,
+  "use_ema": true,
+  "clip_k": 5.0,
+  "grad_clip": 1.0,
+  "use_condition": true,
+  "condition_type": "file_id",
+  "cond_dim": 32,
+  "use_tanh_eps": false,
+  "eps_scale": 1.0,
+  "model_time_dim": 128,
+  "model_hidden_dim": 512,
+  "model_num_layers": 2,
+  "model_dropout": 0.1,
+  "model_ff_mult": 2,
+  "model_pos_dim": 64,
+  "model_use_pos_embed": true,
+  "disc_mask_scale": 0.9,
+  "cont_loss_weighting": "inv_std",
+  "cont_loss_eps": 1e-6,
+  "cont_target": "x0",
+  "cont_clamp_x0": 5.0,
+  "shuffle_buffer": 1024,
+  "use_temporal_stage1": true,
+  "temporal_hidden_dim": 512,
+  "temporal_num_layers": 2,
+  "temporal_dropout": 0.0,
+  "temporal_epochs": 5,
+  "temporal_lr": 0.0005,
+  "sample_batch_size": 8,
+  "sample_seq_len": 128
+}

example/run_compare.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+"""Run two configs sequentially and print metric deltas."""
+
+import argparse
+import csv
+import subprocess
+import sys
+from pathlib import Path
+
+
+def run(cmd):
+    print("running:", " ".join(cmd))
+    subprocess.run(cmd, check=True)
+
+
+def read_last_metrics(path: Path):
+    if not path.exists():
+        return None
+    rows = list(csv.DictReader(path.open()))
+    return rows[-1] if rows else None
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Compare metrics for two configs.")
+    base_dir = Path(__file__).resolve().parent
+    parser.add_argument("--config-a", default=str(base_dir / "config_no_temporal.json"))
+    parser.add_argument("--config-b", default=str(base_dir / "config_temporal_strong.json"))
+    parser.add_argument("--device", default="auto")
+    args = parser.parse_args()
+
+    history = base_dir / "results" / "metrics_history.csv"
+    if history.exists():
+        history.unlink()
+
+    run([sys.executable, str(base_dir / "run_all.py"), "--config", args.config_a, "--device", args.device])
+    a = read_last_metrics(history)
+    run([sys.executable, str(base_dir / "run_all.py"), "--config", args.config_b, "--device", args.device])
+    b = read_last_metrics(history)
+
+    if not a or not b:
+        raise SystemExit("missing metrics_history.csv entries")
+
+    def f(v):
+        return float(v) if v is not None else None
+
+    print("baseline_avg_ks", a["avg_ks"])
+    print("temporal_avg_ks", b["avg_ks"])
+    print("delta_avg_ks", f(b["avg_ks"]) - f(a["avg_ks"]))
+    print("baseline_avg_jsd", a["avg_jsd"])
+    print("temporal_avg_jsd", b["avg_jsd"])
+    print("delta_avg_jsd", f(b["avg_jsd"]) - f(a["avg_jsd"]))
+    print("baseline_avg_lag1_diff", a["avg_lag1_diff"])
+    print("temporal_avg_lag1_diff", b["avg_lag1_diff"])
+    print("delta_avg_lag1_diff", f(b["avg_lag1_diff"]) - f(a["avg_lag1_diff"]))
+
+
+if __name__ == "__main__":
+    main()