File size: 4,666 Bytes
6351b36
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Extract per-layer embeddings from SSL speech models.

For wav2vec2, HuBERT, WavLM, XLS-R, and Whisper, the "last hidden state" used
in the main extraction is known to underperform for speaker tasks relative to
intermediate layers (see SUPERB benchmark). This script saves the mean-pooled
embedding from EVERY transformer layer, enabling layer-wise analysis.

Output format: for each audio file, a numpy array of shape (n_layers, dim).

Usage:
    python3 extract_ssl_layers.py --model wav2vec2 --device cpu
    python3 extract_ssl_layers.py --model hubert --device cpu
    python3 extract_ssl_layers.py --model wavlm --device cpu
    python3 extract_ssl_layers.py --model xlsr --device cpu
    python3 extract_ssl_layers.py --model whisper --device cpu
"""

import argparse
import numpy as np
import torch
from extraction_utils import load_audio, extract_all


HF_CONFIGS = {
    'wav2vec2': {
        'model_id': 'facebook/wav2vec2-base',
        'model_cls': 'Wav2Vec2Model',
        'fe_cls': 'Wav2Vec2FeatureExtractor',
    },
    'hubert': {
        'model_id': 'facebook/hubert-base-ls960',
        'model_cls': 'HubertModel',
        'fe_cls': 'Wav2Vec2FeatureExtractor',
    },
    'wavlm': {
        'model_id': 'microsoft/wavlm-base-plus',
        'model_cls': 'WavLMModel',
        'fe_cls': 'Wav2Vec2FeatureExtractor',
    },
    'xlsr': {
        'model_id': 'facebook/wav2vec2-xls-r-300m',
        'model_cls': 'Wav2Vec2Model',
        'fe_cls': 'Wav2Vec2FeatureExtractor',
    },
}


def build_hf_model_fn(model_name, device):
    cfg = HF_CONFIGS[model_name]
    import transformers
    ModelCls = getattr(transformers, cfg['model_cls'])
    FeCls = getattr(transformers, cfg['fe_cls'])
    print(f"Loading {model_name} ({cfg['model_id']}) on {device}...")
    feature_extractor = FeCls.from_pretrained(cfg['model_id'])
    model = ModelCls.from_pretrained(cfg['model_id']).to(device)
    model.eval()

    def model_fn(audio_path):
        audio = load_audio(audio_path, target_sr=16000)
        inputs = feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
        inputs = {k: v.to(device) for k, v in inputs.items()}
        with torch.no_grad():
            outputs = model(**inputs, output_hidden_states=True)
        # outputs.hidden_states: tuple of (n_layers+1) tensors, each (1, T, D)
        # Index 0 is the CNN feature projection output; indices 1..n are transformer layers
        all_layers = torch.stack(outputs.hidden_states, dim=0)  # (L+1, 1, T, D)
        # Mean pool over time for each layer
        pooled = all_layers.mean(dim=2).squeeze(1)  # (L+1, D)
        return pooled.cpu().numpy()

    return model_fn


def build_whisper_model_fn(device, size='base'):
    import whisper
    print(f"Loading Whisper {size} encoder on {device}...")
    model = whisper.load_model(size, device=device)

    def model_fn(audio_path):
        audio = whisper.load_audio(str(audio_path))
        audio = whisper.pad_or_trim(audio)
        mel = whisper.log_mel_spectrogram(audio).to(device)
        # Whisper's encoder returns only the final output by default.
        # To get per-layer outputs, register a forward hook.
        layer_outputs = []

        def hook(module, input_, output):
            # output is typically a tensor (B, T, D) or tuple
            out = output[0] if isinstance(output, tuple) else output
            layer_outputs.append(out.mean(dim=1).squeeze().cpu().numpy())  # (D,)

        handles = []
        for block in model.encoder.blocks:
            handles.append(block.register_forward_hook(hook))

        with torch.no_grad():
            _ = model.encoder(mel.unsqueeze(0))

        for h in handles:
            h.remove()

        # Also include final layer norm output
        with torch.no_grad():
            final = model.encoder(mel.unsqueeze(0)).mean(dim=1).squeeze().cpu().numpy()
        layer_outputs.append(final)

        return np.stack(layer_outputs, axis=0)  # (L, D)

    return model_fn


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", required=True,
                        choices=['wav2vec2', 'hubert', 'wavlm', 'xlsr', 'whisper'])
    parser.add_argument("--device", default="cpu")
    parser.add_argument("--base-dir", default=None)
    parser.add_argument("--output-dir", default=None)
    args = parser.parse_args()

    if args.model == 'whisper':
        model_fn = build_whisper_model_fn(args.device)
    else:
        model_fn = build_hf_model_fn(args.model, args.device)

    extract_all(model_fn, f"{args.model}_layers", args.base_dir, args.output_dir)


if __name__ == "__main__":
    main()