File size: 3,123 Bytes
6eed659
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
#!/usr/bin/env python3
"""Analyse profonde : d'ou vient l'erreur des champions sur la validation ?
Distingue clips pathologiques FIXABLES (longs, mal decodes) vs NON FIXABLES
(refs cassees a la source : audio court / texte long, mots/sec implausible)."""
import json
import numpy as np
import soundfile as sf
import torch
import jiwer
from transformers import AutoModelForCTC, AutoProcessor

MODELS = {"lin": "lin_s4_best", "sna": "sna_ps_best", "lug": "lug_r_best"}


def norm(s):
    return " ".join(str(s).replace("|", " ").split())


for LG, name in MODELS.items():
    P = f"/root/models/{name}"
    rows = [json.loads(l) for l in open(f"/scratch/prep/manifests/waxal_{LG}_validation.jsonl", encoding="utf-8")]
    rows = [r for r in rows if r.get("text", "").strip()]
    proc = AutoProcessor.from_pretrained(P)
    m = AutoModelForCTC.from_pretrained(P, torch_dtype=torch.bfloat16).cuda().eval()
    rows.sort(key=lambda r: -r["duration"])
    out = {}
    cur, bud, bs = [], 0.0, []
    for r in rows:
        if cur and bud + r["duration"] > 120:
            bs.append(cur); cur, bud = [], 0.0
        cur.append(r); bud += r["duration"]
    if cur:
        bs.append(cur)
    with torch.inference_mode():
        for b in bs:
            au = [sf.read(r["audio"], dtype="float32")[0] for r in b]
            f = proc(au, sampling_rate=16000, return_tensors="pt", padding=True)
            f = {k: v.to("cuda", dtype=torch.bfloat16 if v.dtype == torch.float32 else v.dtype) for k, v in f.items()}
            ids = m(**f).logits.float().argmax(-1).cpu().numpy()
            for r, s in zip(b, proc.batch_decode(ids)):
                out[r["id"]] = norm(s)

    patho, wps_all = [], []
    for r in rows:
        ref = norm(r["text"]); hyp = out[r["id"]]
        rw, hw, dur = len(ref.split()), len(hyp.split()), r["duration"]
        wps = rw / max(dur, 0.1)
        wer = jiwer.wer([ref], [hyp]) if ref else 0
        wps_all.append(wps)
        if hw <= 2 or wer >= 1.0:
            patho.append((r["id"], round(dur, 1), rw, hw, round(wps, 1), round(wer, 2)))

    print(f"=== {LG} validation : {len(rows)} clips | pathologiques (hyp<=2 mots OU WER>=100%): "
          f"{len(patho)} ({100*len(patho)/len(rows):.1f}%)")
    if patho:
        durs = [p[1] for p in patho]; wpss = [p[4] for p in patho]; rws = [p[2] for p in patho]
        n_short = sum(1 for d in durs if d < 3)
        n_wps = sum(1 for w in wpss if w > 4)
        n_longok = sum(1 for _, d, _, _, w, _ in patho if d > 25 and w < 4)
        print(f"  duree patho med={np.median(durs):.1f}s max={max(durs):.1f}s | <3s: {n_short} | >25s&wps<4 (long fixable?): {n_longok}")
        print(f"  mots/sec patho med={np.median(wpss):.1f} | wps>4 (REF CASSEE): {n_wps}/{len(patho)} ({100*n_wps/len(patho):.0f}%)")
        print(f"  ref_mots patho med={int(np.median(rws))}")
    n_all_bad = sum(1 for w in wps_all if w > 4)
    print(f"  TOUS clips avec mots/sec>4 (refs artefacts, non transcriptibles): {n_all_bad} ({100*n_all_bad/len(wps_all):.1f}%)")
    print("  4 refs longues patho:", sorted(patho, key=lambda x: -x[2])[:4])
print("ANALYSE_DONE")