File size: 4,767 Bytes
7ee3434
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2023 Amphion.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

# This code is modified from https://github.com/svc-develop-team/so-vits-svc/blob/4.0/preprocess_hubert_f0.py

import os
import librosa
import torch
import numpy as np
from fairseq import checkpoint_utils
from tqdm import tqdm
import torch


def load_hubert_model(hps):
    # Load model
    ckpt_path = hps.hubert_file
    print("Load Hubert Model...")

    models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
        [ckpt_path],
        suffix="",
    )
    model = models[0]
    model.eval()

    if torch.cuda.is_available():
        model = model.cuda()

    return model


def get_hubert_content(hmodel, wav_16k_tensor):
    feats = wav_16k_tensor
    if feats.dim() == 2:  # double channels
        feats = feats.mean(-1)
    assert feats.dim() == 1, feats.dim()
    feats = feats.view(1, -1)
    padding_mask = torch.BoolTensor(feats.shape).fill_(False)
    inputs = {
        "source": feats.to(wav_16k_tensor.device),
        "padding_mask": padding_mask.to(wav_16k_tensor.device),
        "output_layer": 9,  # layer 9
    }
    with torch.no_grad():
        logits = hmodel.extract_features(**inputs)
        feats = hmodel.final_proj(logits[0]).squeeze(0)

    return feats


def content_vector_encoder(model, audio_path, default_sampling_rate=16000):
    """
    # content vector default sr: 16000
    """

    wav16k, sr = librosa.load(audio_path, sr=default_sampling_rate)
    device = next(model.parameters()).device
    wav16k = torch.from_numpy(wav16k).to(device)

    # (1, 256, frame_len)
    content_feature = get_hubert_content(model, wav_16k_tensor=wav16k)

    return content_feature.cpu().detach().numpy()


def repeat_expand_2d(content, target_len):
    """
    content : [hubert_dim(256), src_len]
    target: [hubert_dim(256), target_len]
    """
    src_len = content.shape[-1]
    target = torch.zeros([content.shape[0], target_len], dtype=torch.float).to(
        content.device
    )
    temp = torch.arange(src_len + 1) * target_len / src_len
    current_pos = 0
    for i in range(target_len):
        if i < temp[current_pos + 1]:
            target[:, i] = content[:, current_pos]
        else:
            current_pos += 1
            target[:, i] = content[:, current_pos]

    return target


def get_mapped_features(raw_content_features, mapping_features):
    """
    Content Vector: frameshift = 20ms, hop_size = 480 in 24k

    Now it's only used for mapping to bigvgan's mels (sr = 24k, hop_size = 256, frameshift ~= 10.7 ms)
    """
    source_hop = 480
    target_hop = 256

    factor = np.gcd(source_hop, target_hop)
    source_hop //= factor
    target_hop //= factor
    print(
        "Mapping source's {} frames => target's {} frames".format(
            target_hop, source_hop
        )
    )

    results = []
    for index, mapping_feat in enumerate(tqdm(mapping_features)):
        # mappping_feat: (mels_frame_len, n_mels)
        target_len = len(mapping_feat)

        # (source_len, 256)
        raw_feats = raw_content_features[index][0].cpu().numpy().T
        source_len, width = raw_feats.shape

        # const ~= target_len * target_hop
        const = source_len * source_hop // target_hop * target_hop

        # (source_len * source_hop, dim)
        up_sampling_feats = np.repeat(raw_feats, source_hop, axis=0)
        # (const, dim) -> (const/target_hop, target_hop, dim) -> (const/target_hop, dim)
        down_sampling_feats = np.average(
            up_sampling_feats[:const].reshape(-1, target_hop, width), axis=1
        )

        err = abs(target_len - len(down_sampling_feats))
        if err > 3:
            print("index:", index)
            print("mels:", mapping_feat.shape)
            print("raw content vector:", raw_feats.shape)
            print("up_sampling:", up_sampling_feats.shape)
            print("down_sampling_feats:", down_sampling_feats.shape)
            exit()
        if len(down_sampling_feats) < target_len:
            # (1, dim) -> (err, dim)
            end = down_sampling_feats[-1][None, :].repeat(err, axis=0)
            down_sampling_feats = np.concatenate([down_sampling_feats, end], axis=0)

        # (target_len, dim)
        feats = down_sampling_feats[:target_len]
        results.append(feats)

    return results


def extract_hubert_features_of_dataset(datasets, model, out_dir):
    for utt in tqdm(datasets):
        uid = utt["Uid"]
        audio_path = utt["Path"]

        content_vector_feature = content_vector_encoder(model, audio_path)  # (T, 256)

        save_path = os.path.join(out_dir, uid + ".npy")
        np.save(save_path, content_vector_feature)