phase2_corrected/code/sna_rescore_v2.py

#!/usr/bin/env python3
"""SHONA v2 — pousser le levier qui a payé (+0.001127 au LB).
Constat v1 : seul un SPÉCIALISTE shona marche comme rescoreur (sna_r2 −0.0029 ; cont2 −0.0002 ;
cont +0.0011 = nuit). Or 4 autres modèles shona n'ont JAMAIS été testés.
Oracle 10-best = 0.0949 (marge −0.0332) : la marge est dans la SÉLECTION.
Ici : (1) tous les rescoreurs shona en solo, (2) N-best élargi 25, (3) combinaison des 2 meilleurs
(avec garde-fou : on ne retient la combinaison que si elle bat nettement le meilleur solo,
sinon sur-apprentissage sur 433 clips — leçon §4c).
"""
import json, os, pickle
import jiwer, numpy as np, soundfile as sf, torch
from multiprocessing import Pool
from pyctcdecode import build_ctcdecoder
from transformers import AutoModelForCTC, AutoProcessor

M1 = "/root/models/sna_ps_best"
R = "/scratch/restore"
NBEST = int(os.environ.get("NBEST", "25"))
AUD = "/root/devhard_audio"
CANDS_R = [("sna_r2", R + "/sna_r2_best"), ("sna_r", R + "/sna_r_best"),
           ("sna_s1", R + "/sna_s1_best"), ("sna_s2", R + "/sna_s2_best"),
           ("sna_ws", R + "/sna_ws_best")]


def comb(refs, hyps):
    pr = [(r, h) for r, h in zip(refs, hyps) if r.strip()]
    a = [x for x, _ in pr]; b = [y for _, y in pr]
    return 0.5 * jiwer.wer(a, b) + 0.5 * jiwer.cer(a, b)


def encode_for(tok, text):
    v = tok.get_vocab()
    delim = getattr(tok, "word_delimiter_token", "|")
    s = text.replace(" ", delim)
    keep = "".join(c for c in s if c in v)
    if not keep:
        keep = "".join(c for c in text.lower().replace(" ", delim) if c in v)
    return [v[c] for c in keep if v[c] != tok.pad_token_id]


def ctc_score(logp, ids, blank):
    T = logp.shape[0]
    if not ids or len(ids) > T:
        return -1e9
    lp = torch.from_numpy(logp).unsqueeze(1)
    return -float(torch.nn.functional.ctc_loss(
        lp, torch.tensor(ids).unsqueeze(0), torch.tensor([T]), torch.tensor([len(ids)]),
        blank=blank, reduction="sum", zero_infinity=True))


def compute_logits(model_dir, rows):
    proc = AutoProcessor.from_pretrained(model_dir)
    m = AutoModelForCTC.from_pretrained(model_dir, dtype=torch.float32).cuda().eval()
    out = []
    with torch.inference_mode():
        for i in range(0, len(rows), 4):
            b = rows[i:i + 4]
            au = [sf.read(r["audio"], dtype="float32")[0] for r in b]
            x = proc(au, sampling_rate=16000, return_tensors="pt", padding=True)
            x = {k: v.cuda() for k, v in x.items()}
            lg = m(**x).logits.log_softmax(-1).float().cpu().numpy()
            for j in range(len(b)):
                out.append(lg[j])
    del m; torch.cuda.empty_cache()
    return proc, out


def main():
    rows = [json.loads(l) for l in open("/root/devhard/devhard_linsna.jsonl", encoding="utf-8")]
    sub = [r for r in rows if r["lang"] == "sna"]
    for r in sub:
        r["audio"] = os.path.join(AUD, os.path.basename(r["audio"]))
    sub = [r for r in sub if os.path.exists(r["audio"])]
    refs = [r["text"] for r in sub]
    print("devhard-sna %d clips | NBEST=%d" % (len(sub), NBEST), flush=True)

    CACHE = "/scratch/lm/logits_sna.pkl"
    L1 = pickle.load(open(CACHE, "rb")) if os.path.exists(CACHE) else compute_logits(M1, sub)[1]
    tok = AutoProcessor.from_pretrained(M1).tokenizer
    v = tok.get_vocab()
    lab = [None] * len(v)
    for t, i in v.items():
        lab[i] = t
    lab[tok.word_delimiter_token_id] = " "
    lab[tok.unk_token_id] = "⁇"
    lab[tok.pad_token_id] = ""
    greedy = [" ".join(tok.decode(l.argmax(-1)).replace("|", " ").split()) for l in L1]
    REF = comb(refs, greedy)
    print("baseline greedy %.4f" % REF, flush=True)

    dec = build_ctcdecoder(lab)
    with Pool(8) as p:
        allbeams = dec.decode_beams_batch(p, L1, beam_width=128)
    with Pool(8) as p:
        db = [" ".join(x.split()) for x in dec.decode_batch(p, L1, beam_width=128)]

    cands, AC1, NW = [], [], []
    for i, bs in enumerate(allbeams):
        c = [" ".join(b[0].split()) for b in bs[:NBEST]]
        a = [(b[3] if len(b) > 3 else 0.0) for b in bs[:NBEST]]
        for extra in (db[i], greedy[i]):
            if extra and extra not in c:
                c.append(extra)
                a.append(ctc_score(L1[i], encode_for(tok, extra), tok.pad_token_id))
        cands.append(c); AC1.append(np.array(a))
        NW.append(np.array([float(len(x.split())) for x in c]))
    orc = [min(cands[i], key=lambda h: comb([refs[i]], [h]) if refs[i].strip() else 0)
           for i in range(len(cands))]
    print("ORACLE %d-best %.4f (marge %+.4f)" % (NBEST, comb(refs, orc), comb(refs, orc) - REF), flush=True)

    SC = {}
    for tag, mdl in CANDS_R:
        if not os.path.isdir(mdl):
            print("%-8s ABSENT" % tag, flush=True); continue
        proc, LG = compute_logits(mdl, sub)
        t2 = proc.tokenizer
        SC[tag] = [np.array([ctc_score(LG[i], encode_for(t2, x), t2.pad_token_id)
                             for x in cands[i]]) for i in range(len(cands))]
        print("%-8s OK" % tag, flush=True)

    def ev(W):
        hyps = []
        for i in range(len(cands)):
            tot = AC1[i].copy()
            for t, w in W.items():
                if w:
                    tot = tot + w * SC[t][i]
            hyps.append(cands[i][int(np.argmax(tot))])
        return comb(refs, hyps)

    print("\n--- solo (ref %.4f) ---" % REF, flush=True)
    solo = {}
    for t in SC:
        bb = (9.0, 0.0)
        for w in (0.3, 0.5, 1.0, 1.5, 2.5, 4.0):
            m = ev({t: w})
            if m < bb[0]:
                bb = (m, w)
        solo[t] = bb
        print("  %-8s %.4f (w=%.1f)  %+.4f" % (t, bb[0], bb[1], bb[0] - REF), flush=True)

    ranked = sorted(solo, key=lambda t: solo[t][0])
    best_solo = solo[ranked[0]]
    print("\n--- combinaison des 2 meilleurs (%s + %s) ---" % (ranked[0], ranked[1]), flush=True)
    bc = (9.0, None, None)
    for w1 in (0.5, 1.0, 1.5, 2.5):
        for w2 in (0.0, 0.3, 0.5, 1.0, 1.5):
            m = ev({ranked[0]: w1, ranked[1]: w2})
            if m < bc[0]:
                bc = (m, w1, w2)
    print("  best %.4f (%s=%.1f %s=%.1f)  %+.4f vs solo" % (bc[0], ranked[0], bc[1], ranked[1], bc[2], bc[0] - best_solo[0]), flush=True)
    keep_combo = bc[0] < best_solo[0] - 0.0015          # garde-fou anti sur-apprentissage
    print("\nRETENU : %s" % ("COMBINAISON" if keep_combo else "SOLO %s w=%.1f" % (ranked[0], best_solo[1])), flush=True)
    json.dump({"ref": REF, "solo": {k: list(v) for k, v in solo.items()},
               "combo": list(bc), "keep_combo": bool(keep_combo), "nbest": NBEST},
              open("/root/sna_v2.json", "w"))
    print("SNA_V2_DONE", flush=True)


if __name__ == "__main__":
    main()