File size: 8,088 Bytes
a001cf7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
"""Fine-tune a Whisper-family ASR (default PhoWhisper-medium) on train audio_text (final turn + history turns).

Dev ids are excluded so asr_eval / agent dev scores stay honest. Light augmentation: speed perturb and additive noise.

usage: python asr_ft.py --root DATA_ROOT --exclude dev300.txt --out ckpt/asr_phowm_ft [--model vinai/PhoWhisper-medium] [--epochs 5]
"""
from __future__ import annotations

import argparse
import json
import random
import time
from pathlib import Path

import librosa
import numpy as np
import soundfile as sf
import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor


def is_batch_c(r: dict) -> bool:
    """generation batch C (dialects, foreign names: highest WER), from context only - see analysis/batches.py"""
    t = r["context"]["current_time"]
    if t[11:16] == "09:00" and t[5:7] in ("09", "10"):
        return False  # batch B
    return int(r["id"].split("_")[1]) >= 900


def load_pairs(root: Path, exclude: set[str], oversample_c: int = 1, holdout_every: int = 0):
    """returns (train pairs, held-out pairs). holdout_every=N keeps every N-th batch-C sample out of training, as a
    guard set: when the whole train set is used, it is the only way to check the new model is not worse."""
    pairs, held, nc = [], [], 0
    for line in open(root / "train" / "train.jsonl", encoding="utf-8"):
        r = json.loads(line)
        if r["id"] in exclude:
            continue
        mine = [(root / "train" / r["audio"], r["audio_text"])]
        mine += [(root / "train" / t["audio"], t["audio_text"]) for t in r["history"] if t.get("audio")]
        if is_batch_c(r):
            nc += 1
            if holdout_every and nc % holdout_every == 0:
                held += mine
                continue
        pairs += mine * (oversample_c if is_batch_c(r) else 1)
    return pairs, held


def wer_of(model, proc, audio, refs, bs=8):
    import re
    import unicodedata
    norm = lambda t: re.findall(r"[^\W_]+", unicodedata.normalize("NFC", t.lower()))
    model.eval()
    errs = words = 0
    for i in range(0, len(audio), bs):
        feats = proc.feature_extractor(audio[i:i + bs], sampling_rate=16000, return_tensors="pt").input_features.cuda().to(model.dtype)
        with torch.no_grad():
            out = proc.batch_decode(model.generate(feats, num_beams=1, max_new_tokens=220), skip_special_tokens=True)
        for h, r in zip(out, refs[i:i + bs]):
            a, b = norm(r), norm(h)
            d = list(range(len(b) + 1))
            for x in range(1, len(a) + 1):
                p, d[0] = d[0], x
                for y in range(1, len(b) + 1):
                    p, d[y] = d[y], min(d[y] + 1, d[y - 1] + 1, p + (a[x - 1] != b[y - 1]))
            errs += d[-1]
            words += len(a)
    return errs / max(1, words)


def augment(w: np.ndarray, rng: random.Random) -> np.ndarray:
    if rng.random() < 0.5:
        w = librosa.effects.time_stretch(w, rate=rng.uniform(0.9, 1.1))
    if rng.random() < 0.4:
        snr = rng.uniform(10, 30)
        noise = np.random.default_rng(rng.randrange(1 << 30)).standard_normal(len(w)).astype(np.float32)
        w = w + noise * (np.sqrt(np.mean(w ** 2) + 1e-8) / 10 ** (snr / 20))
    return w.astype(np.float32)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--root", required=True)
    ap.add_argument("--model", default="vinai/PhoWhisper-medium")
    ap.add_argument("--exclude", help="ids whose audio must not be used (dev)")
    ap.add_argument("--out", required=True)
    ap.add_argument("--epochs", type=int, default=5)
    ap.add_argument("--bs", type=int, default=16)
    ap.add_argument("--lr", type=float, default=1e-5)
    ap.add_argument("--oversample-c", type=int, default=1, help="repeat batch-C utterances (dialects, foreign names) this many times")
    ap.add_argument("--extra", help="jsonl {audio: abs wav path (16 kHz), text} added to training (TTS synthetic speech)")
    ap.add_argument("--holdout-every", type=int, default=20, help="keep every N-th batch-C sample out as a guard set (0 = none)")
    a = ap.parse_args()

    rng = random.Random(0)
    torch.manual_seed(0)
    exclude = {l.strip() for l in open(a.exclude)} if a.exclude else set()
    pairs, held = load_pairs(Path(a.root), exclude, a.oversample_c, a.holdout_every)
    if a.extra:
        ex = [json.loads(l) for l in open(a.extra, encoding="utf-8")]
        pairs += [(Path(r["audio"]), r["text"]) for r in ex]
        print(f"+ {len(ex)} extra (TTS) utterances")
    print(f"{len(pairs)} training utterances, {len(held)} held out (guard set)")
    audio = [sf.read(p, dtype="float32")[0] for p, _ in pairs]

    proc = WhisperProcessor.from_pretrained(a.model)
    model = WhisperForConditionalGeneration.from_pretrained(a.model, torch_dtype=torch.float32).cuda()
    model.config.forced_decoder_ids = None
    model.generation_config.forced_decoder_ids = None
    model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
    opt = torch.optim.AdamW(model.parameters(), lr=a.lr, weight_decay=0.01)
    steps = a.epochs * (len(pairs) // a.bs)
    sched = torch.optim.lr_scheduler.LambdaLR(opt, lambda s: min(1.0, (s + 1) / 100) * max(0.0, 1 - s / steps))
    prefix = proc.tokenizer("", add_special_tokens=True).input_ids[:-1]  # decoder start + task/lang tokens of this checkpoint

    step, t0 = 0, time.time()
    model.train()
    for ep in range(a.epochs):
        order = list(range(len(pairs)))
        rng.shuffle(order)
        for b in range(0, len(order) - a.bs + 1, a.bs):
            idx = order[b : b + a.bs]
            wavs = [augment(audio[i], rng) for i in idx]
            feats = proc.feature_extractor(wavs, sampling_rate=16000, return_tensors="pt").input_features.cuda()
            labels = [proc.tokenizer(pairs[i][1], add_special_tokens=True).input_ids for i in idx]
            n = max(map(len, labels))
            lab = torch.full((len(labels), n), -100, dtype=torch.long)
            for j, l in enumerate(labels):
                lab[j, : len(l)] = torch.tensor(l)
            dec_in = lab[:, :-1].clone()
            dec_in[dec_in == -100] = proc.tokenizer.pad_token_id
            target = lab[:, 1:].clone()
            target[:, : max(0, len(prefix) - 1)] = -100  # do not train on the fixed prefix tokens
            with torch.autocast("cuda", dtype=torch.bfloat16):
                out = model(input_features=feats, decoder_input_ids=dec_in.cuda())
            loss = torch.nn.functional.cross_entropy(out.logits.float().transpose(1, 2), target.cuda(), ignore_index=-100)
            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
            opt.step()
            sched.step()
            opt.zero_grad(set_to_none=True)
            step += 1
            if step % 50 == 0:
                print(f"ep {ep} step {step}/{steps} loss {loss.item():.4f} {time.time() - t0:.0f}s", flush=True)
        model.save_pretrained(a.out)
        proc.save_pretrained(a.out)
    print("saved", a.out)
    # guard: never ship an ASR that is worse than the one we started from
    if held:
        held_audio = [sf.read(p, dtype="float32")[0] for p, _ in held]
        refs = [t for _, t in held]
        new_wer = wer_of(model.half(), proc, held_audio, refs)
        del model
        torch.cuda.empty_cache()
        base = WhisperForConditionalGeneration.from_pretrained(a.model, torch_dtype=torch.float16).cuda()
        base.generation_config.forced_decoder_ids = None
        base_wer = wer_of(base, proc, held_audio, refs)
        print(f"GUARD held-out batch C: start model WER {100 * base_wer:.2f}% -> new model WER {100 * new_wer:.2f}%", flush=True)
        if new_wer >= base_wer:
            print("GUARD: new model is not better -> keeping the start model", flush=True)
            base.save_pretrained(a.out)
            WhisperProcessor.from_pretrained(a.model).save_pretrained(a.out)
            open(Path(a.out) / "GUARD_KEPT_START_MODEL", "w").write(f"{base_wer} {new_wer}\n")


if __name__ == "__main__":
    main()