File size: 4,946 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
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
#!/usr/bin/env python3
"""Generateur de soumission parametrable : lin = <modele> + beam KenLM, sna = sna_ps greedy.
Casse du 1er caractere toujours copiee du greedy (gain valide).
Usage : MODEL=... ARPA=... ALPHA=.. BETA=.. LSB=0|1 OUT=... python gen_sub.py
"""
import csv
import glob
import json
import os

import soundfile as sf
import torch
from multiprocessing import Pool
from pyctcdecode import build_ctcdecoder
from transformers import AutoModelForCTC, AutoProcessor

SR = 16000
CACHE = "/scratch/p2_16k"
LINM = os.environ.get("MODEL", "/root/models/joint_cont_best")
ARPA = os.environ.get("ARPA", "/scratch/lm/lin_5g.arpa")
ALPHA = float(os.environ.get("ALPHA", "0.5"))
BETA = float(os.environ.get("BETA", "0.5"))
LSB = os.environ.get("LSB", "0") == "1"
OUT = os.environ.get("OUT", "/root/sub_gen.csv")


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


def batches(sel, budget):
    durs = {f: sf.info(f).duration for f in sel}
    sel = sorted(sel, key=lambda f: -durs[f])
    bs, cur, acc = [], [], 0.0
    for f in sel:
        if cur and acc + durs[f] > budget:
            bs.append(cur)
            cur, acc = [], 0.0
        cur.append(f)
        acc += durs[f]
    if cur:
        bs.append(cur)
    return bs


def main():
    lang = json.load(open(os.environ.get("LANGF", "/root/test_lang.json")))
    files = sorted(glob.glob(os.path.join(CACHE, "*.wav")))
    ids = [os.path.splitext(os.path.basename(f))[0] for f in files]
    lin = [f for f in files if lang[os.path.splitext(os.path.basename(f))[0]] == "lin"]
    sna = [f for f in files if lang[os.path.splitext(os.path.basename(f))[0]] == "sna"]
    print("lin=%d (%s + KenLM a=%.2f b=%.2f lsb=%s) | sna=%d (sna_ps greedy)"
          % (len(lin), os.path.basename(LINM), ALPHA, BETA, LSB, len(sna)), flush=True)
    out = {}

    # ---- lin : beam + KenLM ----
    proc = AutoProcessor.from_pretrained(LINM)
    tok = proc.tokenizer
    m = AutoModelForCTC.from_pretrained(LINM, dtype=torch.float32).cuda().eval()
    logs, order, greedy = [], [], []
    with torch.inference_mode():
        for b in batches(lin, 90):
            au = [sf.read(f, dtype="float32")[0] for f in b]
            x = proc(au, sampling_rate=SR, 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, f in enumerate(b):
                logs.append(lg[j])
                order.append(f)
                greedy.append(norm(tok.decode(lg[j].argmax(-1))))
    del m
    torch.cuda.empty_cache()

    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] = ""
    dec = build_ctcdecoder(lab, kenlm_model_path=ARPA, alpha=ALPHA, beta=BETA,
                           lm_score_boundary=LSB)
    with Pool(8) as p:
        hyps = dec.decode_batch(p, logs, beam_width=64)
    for f, h, g in zip(order, hyps, greedy):
        h = norm(h)
        if h and g:
            h = g[:1] + h[1:]          # casse du 1er caractere = celle du modele acoustique
        out[os.path.splitext(os.path.basename(f))[0]] = h
    print("lin OK", flush=True)

    # ---- sna : greedy (le LM degrade, mesure) ----
    proc2 = AutoProcessor.from_pretrained("/root/models/sna_ps_best")
    m2 = AutoModelForCTC.from_pretrained("/root/models/sna_ps_best",
                                         dtype=torch.bfloat16).cuda().eval()
    with torch.inference_mode():
        for b in batches(sna, 140):
            au = [sf.read(f, dtype="float32")[0] for f in b]
            x = proc2(au, sampling_rate=SR, return_tensors="pt", padding=True)
            x = {k: v.to("cuda", dtype=torch.bfloat16 if v.dtype == torch.float32 else v.dtype)
                 for k, v in x.items()}
            pid = m2(**x).logits.float().argmax(-1).cpu().numpy()
            for f, s in zip(b, proc2.batch_decode(pid)):
                out[os.path.splitext(os.path.basename(f))[0]] = norm(s)
    del m2
    torch.cuda.empty_cache()
    print("sna OK", flush=True)

    # ---- ecriture, en comblant les vides depuis la meilleure soumission connue ----
    fb = {}
    ref = "/root/sub_p2_KENLM.csv"
    if os.path.exists(ref):
        fb = {r["ID"]: r["Target"] for r in csv.DictReader(open(ref, encoding="utf-8"))}
    filled = 0
    for i in ids:
        if not out.get(i, "").strip() and fb.get(i, "").strip():
            out[i] = fb[i]
            filled += 1
    with open(OUT, "w", newline="", encoding="utf-8") as fo:
        w = csv.writer(fo)
        w.writerow(["ID", "Target"])
        for i in ids:
            w.writerow([i, out.get(i) or "a"])
    empty = sum(1 for i in ids if not out.get(i, "").strip())
    print("GEN_DONE %s | %d IDs | vides=%d | combles=%d" % (OUT, len(ids), empty, filled), flush=True)


if __name__ == "__main__":
    main()