holdout run: exclude a07 from training, add int4 eval stage

kaggle/notebook_holdout.py

@@ -0,0 +1,82 @@
+"""Kaggle notebook body: hold out one voice, then measure generalisation.
+
+The previous student fitted the seven voices it was trained on to within 1.19 pp
+of the teacher. That is in-distribution. The case that matters for user
+uploads is a voice the student has never heard, so this run trains on six and
+evaluates the seventh.
+
+a07 (伊芙琳) is held out. a08 stays in training as the in-distribution control,
+so the two columns separate "fitted" from "generalised".
+
+Also measures the int4 quantisation of the same student, since that decides
+which bundle ships.
+
+  traj       6 voices x 120 texts, 2 GPUs     ~55 min
+  train      4-step student                   ~25 min
+  export     ONNX fp32 + int4                 ~5 min
+  eval       a07 + a08, teacher vs student    ~15 min
+  eval_int4  same cells through the int4 graph ~12 min
+"""
+import os
+import subprocess
+import sys
+import time
+
+from notebook_eval import BUNDLE, MODEL, ROOT, WEIGHTS, pull, stage_deps, stage_repo
+
+BUNDLE_REPO = "idefinity/webfm-eval"
+SHARDS = [["a01", "a03", "a04"], ["a05", "a06", "a08"]]
+HOLDOUT = "a07"
+EVAL_VOICES = "a07,a08"
+
+
+def run(tag: str, extra: dict, gpu: int = 0) -> int:
+    env = dict(os.environ, WORK_ROOT=str(ROOT), HOLDOUT=HOLDOUT, **extra)
+    env["CUDA_VISIBLE_DEVICES"] = str(gpu)
+    env["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+    log = open(ROOT / f"holdout_{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"holdout_{tag}.log").read_text()[-2500:])
+    return rc
+
+
+def main() -> None:
+    stage_repo()
+    stage_deps()
+    pull(WEIGHTS, MODEL)
+    pull(BUNDLE_REPO, BUNDLE, repo_type="dataset")
+
+    # Guard: confirm the holdout filter works before spending an hour.
+    rc = run("smoke", {"STAGE": "traj", "VOICES": "a08", "SMOKE": "1", "N_TEXTS": "2"})
+    if rc != 0:
+        raise SystemExit("traj path broken")
+    if (ROOT / "train/traj/a07").exists() and any((ROOT / "train/traj/a07").iterdir()):
+        raise SystemExit("holdout voice leaked into trajectories")
+
+    print("== trajectories (6 voices) ==", flush=True)
+    procs = []
+    for gpu, shard in enumerate(SHARDS):
+        env = dict(os.environ, WORK_ROOT=str(ROOT), STAGE="traj", HOLDOUT=HOLDOUT,
+                   VOICES=",".join(shard), CUDA_VISIBLE_DEVICES=str(gpu))
+        log = open(ROOT / f"traj_gpu{gpu}.log", "w")
+        procs.append(subprocess.Popen([sys.executable, str(BUNDLE / "kaggle/train_voices.py")],
+                                      env=env, stdout=log, stderr=subprocess.STDOUT, cwd=str(ROOT)))
+    while any(p.poll() is None for p in procs):
+        time.sleep(60)
+        n = len(list((ROOT / "train/traj").glob("*/*.pt")))
+        print(f"[{time.strftime('%H:%M')}] trajectories={n}", flush=True)
+    print("traj exits:", [p.returncode for p in procs])
+
+    if run("final", {"STAGE": "all", "TRAIN_STEPS": "1500", "BATCH": "16",
+                     "EVAL_VOICES": EVAL_VOICES}) != 0:
+        raise SystemExit("train/export/eval failed")
+
+    run("int4", {"STAGE": "eval_int4", "EVAL_VOICES": EVAL_VOICES})
+
+
+if __name__ == "__main__":
+    main()