phase2_corrected/code/sna_rescore.py

#!/usr/bin/env python3
"""RESCORING N-BEST SUR LE SHONA — jamais testé (le shona = la moitié du test).
Le KenLM DÉGRADE le shona (il déforme), mais le rescoring ne fait que CHOISIR parmi des
hypothèses déjà produites par sna_ps : mécanisme différent, non invalidé.

score(h) = ac_snaps(h) + somme_i w_i*ac_i(h) + gamma*nb_mots(h)

Gate = devhard-sna (433 clips), FIABLE pour le shona d'après §7 de l'AUTOPILOT.
Baseline à battre : sna_ps greedy = 0.1281 (combine).
"""
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"          # modèle principal shona
NBEST = 10
R = "/scratch/restore"
RESCORERS = [
    ("cont2", R + "/joint_cont2_best"),
    ("cont",  R + "/joint_cont_best"),
    ("sna_r2", R + "/sna_r2_best"),
]
AUD = "/root/devhard_audio"


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]
    w = jiwer.wer(a, b); c = jiwer.cer(a, b)
    return w, c, 0.5 * w + 0.5 * c


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)
    loss = 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)
    return -float(loss)


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:                                  # réécrire les chemins vers /root
        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" % len(sub), flush=True)

    # --- logits du modèle principal (cache) ---
    CACHE = "/scratch/lm/logits_sna.pkl"
    os.makedirs("/scratch/lm", exist_ok=True)
    if os.path.exists(CACHE):
        L1 = pickle.load(open(CACHE, "rb"))
    else:
        _, L1 = compute_logits(M1, sub)
        pickle.dump(L1, open(CACHE, "wb"))
    print("logits sna_ps OK", flush=True)

    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 sna_ps : %.4f" % REF, flush=True)

    # --- N-best par beam CTC PUR (aucun LM : il dégrade le shona) ---
    dec = build_ctcdecoder(lab)                    # pas de kenlm_model_path
    with Pool(8) as p:
        allbeams = dec.decode_beams_batch(p, L1, beam_width=64)
    with Pool(8) as p:
        db = [" ".join(x.split()) for x in dec.decode_batch(p, L1, beam_width=64)]
    _, _, BEAM = comb(refs, db)
    print("beam CTC pur (1-best)  : %.4f  (%+.4f vs greedy)" % (BEAM, BEAM - REF), flush=True)

    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]):           # injecter decode_batch ET le greedy
            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]))

    # --- ORACLE : plafond atteignable par simple sélection ---
    orc = []
    for i in range(len(cands)):
        best = min(cands[i], key=lambda h: comb([refs[i]], [h])[2] if refs[i].strip() else 0)
        orc.append(best)
    _, _, ORACLE = comb(refs, orc)
    print("ORACLE %d-best         : %.4f  (marge %+.4f)" % (NBEST, ORACLE, ORACLE - REF), flush=True)

    # --- scores des rescoreurs ---
    SC = {}
    for tag, mdl in RESCORERS:
        if not os.path.isdir(mdl):
            print("%-8s ABSENT %s" % (tag, mdl), 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 scores OK (|V|=%d)" % (tag, len(t2.get_vocab())), flush=True)

    def evaluate(W, gamma=0.0):
        hyps = []
        for i in range(len(cands)):
            tot = AC1[i] + gamma * NW[i]
            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)[2]

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

    print("\n--- (b) terme de longueur seul ---", flush=True)
    for gm in (-2.0, -1.0, 0.0, 1.0, 2.0, 4.0):
        m = evaluate({}, gm)
        print("  gamma=%+5.1f : %.4f  (%+.4f)" % (gm, m, m - REF), flush=True)

    print("\n--- (c) meilleur rescoreur + longueur ---", flush=True)
    best = (9.0, None, None, None)
    if solo:
        btag = min(solo, key=lambda t: solo[t][0])
        for w in (0.5, 1.0, 1.5, 2.5):
            for gm in (0.0, 1.0, 2.0, 4.0):
                m = evaluate({btag: w}, gm)
                if m < best[0]:
                    best = (m, btag, w, gm)
                print("  %s=%.1f gamma=%+5.1f : %.4f (%+.4f)" % (btag, w, gm, m, m - REF), flush=True)
    print("\nBEST_SNA %.4f  (%s w=%s gamma=%s)  baseline %.4f  gain %+.4f"
          % (best[0], best[1], best[2], best[3], REF, best[0] - REF), flush=True)
    json.dump({"best": best[0], "tag": best[1], "w": best[2], "gamma": best[3],
               "ref": REF, "oracle": ORACLE},
              open("/root/sna_rescore.json", "w"))
    print("SNA_RESCORE_DONE", flush=True)


if __name__ == "__main__":
    main()