keras-spec-augment/spec_aug_generator.py at master · T-Sumida/keras-spec-augment · GitHub

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
# coding:utf-8
import os
import glob
import random
import librosa
import numpy as np
from tensorflow.keras.utils import Sequence
from tensorflow.keras.utils import to_categorical


class SpecAugGenerator(Sequence):
    def __init__(self, dataset_path, cfg, seed=1024):
        self.cfg = cfg
        self.random_state = np.random.RandomState(seed)
        self._create_class_dict(dataset_path)
        self.queue = []
        self.index_per_class = [0] * self.class_num

    def __getitem__(self, idx):
        i = 0
        x, y = [], []
        while i < self.cfg['batch_size']:
            # class選択キューの補給
            if len(self.queue) == 0:
                self.queue = self._extend_queue(self.queue)

            # queueからクラス番号を抜き出し、ファイルパスを得る
            class_id = self.queue.pop(0)
            data_index = self.index_per_class[class_id]
            data_path = self.data_dict[class_id][data_index]

            # wavファイルを読み込み、メルスペクトログラムを作成
            signal, _ = librosa.core.load(data_path, self.cfg['sr'], mono=True)
            signal = self._clip_or_padding(signal)
            spec = librosa.feature.melspectrogram(
                signal, sr=self.cfg['sr'], n_mels=self.cfg['mel_bins'],
                n_fft=self.cfg['n_fft'], hop_length=self.cfg['hop_size']
            )
            spec_db = librosa.power_to_db(spec, ref=np.max)
            spec_db = spec_db - np.mean(spec_db)

            # spec_augmentを実施
            spec_masked = self._augment(spec_db)

            self.index_per_class[class_id] += 1
            if self.index_per_class[class_id] >= len(self.data_dict[class_id]):
                self.index_per_class[class_id] = 0
                self.random_state.shuffle(self.data_dict[class_id])
            x.append(np.expand_dims(spec_masked, axis=2))
            y.append(to_categorical(class_id, num_classes=self.class_num))
            i += 1
        return np.array(x), np.array(y)

    def _clip_or_padding(self, signal):
        # オーディオを指定の長さにクリッピング or パディングする
        limit_audio_samples = self.cfg['sr'] * self.cfg['audio_length']
        if signal.shape[0] < limit_audio_samples:
            signal = np.concatenate(
                (signal, np.zeros(limit_audio_samples - signal.shape[0])),
                axis=0
            )
        elif signal.shape[0] > limit_audio_samples:
            clipping_start_index = signal.shape[0]//2 - limit_audio_samples//2
            signal = signal[clipping_start_index:clipping_start_index+limit_audio_samples]
        return signal

    def _extend_queue(self, queue):
        # queueを補給する
        class_set = list(np.arange(self.class_num))
        self.random_state.shuffle(class_set)
        queue = class_set
        return queue

    def _augment(self, spec):
        # time warp is not implemented
        NFrame = spec.shape[1]
        NBin = spec.shape[0]

        # check
        if NFrame < self.cfg['spec_aug']['T_width'] * 2 + 1:
            T = NFrame//self.cfg['spec_aug']['T_line_num']
        else:
            T = self.cfg['spec_aug']['T_width']
        if NBin < self.cfg['spec_aug']['F_width'] * 2 + 1:
            F = NBin // self.cfg['spec_aug']['F_line_num']
        else:
            F = self.cfg['spec_aug']['F_width']
        t = np.random.randint(T-1, size=self.cfg['spec_aug']['T_line_num']) + 1
        f = np.random.randint(F-1, size=self.cfg['spec_aug']['F_line_num']) + 1
        mask_t = np.ones((NFrame, 1))
        mask_f = np.ones((NBin, 1))

        index = 0
        t_tmp = t.sum() + self.cfg['spec_aug']['T_line_num']
        for _t in t:
            k = random.randint(index, NFrame-t_tmp)
            mask_t[k:k+_t] = 0
            index += k + _t + 1
            t_tmp = t_tmp - (_t + 1)
        mask_t[index:] = 1
        index = 0
        f_tmp = f.sum() + self.cfg['spec_aug']['F_line_num']
        for _f in f:
            k = random.randint(index, NBin-f_tmp)
            mask_f[k:k+_f] = 0
            index += k + _f + 1
            f_tmp = f_tmp - (_f + 1)
        mask_f[index:] = 1

        spec_masked = ((spec * mask_t.T) * mask_f)
        return spec_masked

    def __len__(self):
        return self.data_num // self.cfg['batch_size']

    def on_epoch_end(self):
        for i in range(self.class_num):
            self.random_state.shuffle(self.data_dict[i])
        self.index_per_class = [0] * self.class_num
        return

    def _create_class_dict(self, dataset_path):
        self.data_dict = {}
        self.data_num = 0
        classes = glob.glob(os.path.join(dataset_path, "*"))
        self.class_num = len(classes)
        for i, cla in enumerate(classes):
            wav_files = glob.glob(os.path.join(cla, "*.wav"))
            self.data_num += len(wav_files)
            self.data_dict[i] = wav_files