train the student on the whisper voice

kaggle/notebook_radio3.py

@@ -0,0 +1,61 @@
+"""Kaggle notebook body: teach the student the whisper voice.
+
+Radio v2 showed the split cleanly: teacher 0.22 on the whisper line, student
+1.00. a09 was added after training, so the student has never seen it. This run
+adds 120 trajectories for a09 to the cached six voices and retrains at the same
+4000 steps, then evaluates a09 against a08 as an in-distribution control.
+
+Trajectories are unpacked from the cache first, so the a09 pass appends to the
+same directory instead of replacing it.
+"""
+import os
+import subprocess
+import sys
+import time
+import zipfile
+
+from notebook_eval import BUNDLE, MODEL, ROOT, WEIGHTS, pull, stage_deps, stage_repo
+
+BUNDLE_REPO = "idefinity/webfm-eval"
+TRAJ = ROOT / "train/traj"
+
+
+def run(tag: str, extra: dict, gpu: int = 0) -> int:
+    env = dict(os.environ, WORK_ROOT=str(ROOT), **extra)
+    env["CUDA_VISIBLE_DEVICES"] = str(gpu)
+    env["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+    log = open(ROOT / f"r3_{tag}.log", "w")
+    p = subprocess.Popen([sys.executable, str(BUNDLE / "kaggle/train_voices.py")],
+                         env=env, stdout=log, stderr=subprocess.STDOUT, cwd=str(ROOT))
+    rc = p.wait()
+    print(f"[{tag}] rc={rc}", flush=True)
+    if rc != 0:
+        print((ROOT / f"r3_{tag}.log").read_text()[-2500:])
+    return rc
+
+
+def main() -> None:
+    stage_repo()
+    stage_deps()
+    pull(WEIGHTS, MODEL)
+    pull(BUNDLE_REPO, BUNDLE, repo_type="dataset")
+
+    TRAJ.mkdir(parents=True, exist_ok=True)
+    cache = BUNDLE / "traj.zip"
+    if cache.exists():
+        with zipfile.ZipFile(cache) as z:
+            z.extractall(ROOT / "train")
+        print("unpacked cache:", len(list(TRAJ.glob("*/*.pt"))), "files", flush=True)
+
+    # One voice, one GPU: ~120 generations.
+    if run("a09", {"STAGE": "traj", "VOICES": "a09"}) != 0:
+        raise SystemExit("a09 trajectories failed")
+    print("total trajectories:", len(list(TRAJ.glob("*/*.pt"))), flush=True)
+
+    if run("train", {"STAGE": "all", "TRAIN_STEPS": "4000", "BATCH": "16",
+                     "EVAL_VOICES": "a09,a08"}) != 0:
+        raise SystemExit("train/eval failed")
+
+
+if __name__ == "__main__":
+    main()