chainer-Fast-WaveNet/utils.py at master · dhgrs/chainer-Fast-WaveNet · 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
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
import random
import copy

import numpy
import librosa
import chainer
from chainer import configuration
from chainer import link


class MuLaw(object):
    def __init__(self, mu=256, int_type=numpy.int32, float_type=numpy.float32):
        self.mu = mu
        self.int_type = int_type
        self.float_type = float_type

    def transform(self, x):
        x = x.astype(self.float_type)
        y = numpy.sign(x) * numpy.log(1 + self.mu * numpy.abs(x)) / \
            numpy.log(1 + self.mu)
        y = numpy.digitize(y, 2 * numpy.arange(self.mu) / self.mu - 1) - 1
        return y.astype(self.int_type)

    def itransform(self, y):
        y = y.astype(self.float_type)
        y = 2 * y / self.mu - 1
        x = numpy.sign(y) / self.mu * ((self.mu) ** numpy.abs(y) - 1)
        return x.astype(self.float_type)


class Preprocess(object):
    def __init__(self, sr, n_fft, hop_length, n_mels, top_db, input_dim,
                 quantize, length, use_logistic):
        self.sr = sr
        self.n_fft = n_fft
        self.hop_length = hop_length
        self.n_mels = n_mels
        self.top_db = top_db
        if input_dim == 1:
            self.mu_law_input = False
        else:
            self.mu_law_input = True
            self.mu_law = MuLaw(quantize)
            self.quantize = quantize
        if length is None:
            self.length = None
        else:
            self.length = length + 1
        self.use_logistic = use_logistic
        if self.mu_law_input or not self.use_logistic:
            self.mu_law = MuLaw(quantize)
            self.quantize = quantize

    def __call__(self, path):
        # load data(trim and normalize)
        raw, _ = librosa.load(path, self.sr)
        raw, _ = librosa.effects.trim(raw, self.top_db)
        raw /= numpy.abs(raw).max()
        raw = raw.astype(numpy.float32)

        # mu-law transform
        if self.mu_law_input or not self.use_logistic:
            quantized = self.mu_law.transform(raw)

        # padding/triming
        if self.length is not None:
            if len(raw) <= self.length:
                # padding
                pad = self.length - len(raw)
                raw = numpy.concatenate(
                    (raw, numpy.zeros(pad, dtype=numpy.float32)))
                if self.mu_law_input or not self.use_logistic:
                    quantized = numpy.concatenate(
                        (quantized, self.quantize // 2 * numpy.ones(pad)))
                    quantized = quantized.astype(numpy.int32)
            else:
                # triming
                start = random.randint(0, len(raw) - self.length - 1)
                raw = raw[start:start + self.length]
                if self.mu_law_input or not self.use_logistic:
                    quantized = quantized[start:start + self.length]

        # make mel spectrogram
        spectrogram = librosa.feature.melspectrogram(
            raw, self.sr, n_fft=self.n_fft, hop_length=self.hop_length,
            n_mels=self.n_mels)
        spectrogram = librosa.power_to_db(
            spectrogram, ref=numpy.max)
        spectrogram += 40
        spectrogram /= 40
        if self.length is not None:
            spectrogram = spectrogram[:, :self.length // self.hop_length]
        spectrogram = spectrogram.astype(numpy.float32)

        # expand dimensions
        if self.mu_law_input:
            one_hot = numpy.identity(
                self.quantize, dtype=numpy.float32)[quantized]
            one_hot = numpy.expand_dims(one_hot.T, 2)
        raw = numpy.expand_dims(raw, 0)  # expand channel
        raw = numpy.expand_dims(raw, -1)  # expand height
        spectrogram = numpy.expand_dims(spectrogram, 2)
        if not self.use_logistic:
            quantized = numpy.expand_dims(quantized, 1)

        # return
        inputs = ()
        if self.mu_law_input:
            inputs += (one_hot[:, :-1],)
        else:
            inputs += (raw[:, :-1],)
        inputs += (spectrogram,)
        if self.use_logistic:
            inputs += (raw[:, 1:],)
        else:
            inputs += (quantized[1:],)
        return inputs


class ExponentialMovingAverage(link.Chain):

    def __init__(self, target, decay=0.999):
        super(ExponentialMovingAverage, self).__init__()
        self.decay = decay
        with self.init_scope():
            self.target = target
            self.ema = copy.deepcopy(target)

    def __call__(self, *args, **kwargs):
        if configuration.config.train:
            ys = self.target(*args, **kwargs)
            xp = chainer.cuda.get_array_module(ys)
            if xp != numpy:
                xp.cuda.Device(ys.array.device).use()
            for target_name, target_param in self.target.namedparams():
                for ema_name, ema_param in self.ema.namedparams():
                    if target_name == ema_name:
                        if not target_param.requires_grad \
                                or ema_param.array is None:
                            new_average = target_param.array
                        else:
                            new_average = self.decay * target_param.array + \
                                (1 - self.decay) * ema_param.array
                        ema_param.array = new_average
        else:
            ys = self.ema(*args, **kwargs)
        return ys