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
|
import unittest
from distutils.version import StrictVersion
import torch
import torchaudio.functional as F
from parameterized import param
from torchaudio._internal.module_utils import is_module_available
LIBROSA_AVAILABLE = is_module_available("librosa")
if LIBROSA_AVAILABLE:
import librosa
import numpy as np
from torchaudio_unittest.common_utils import get_spectrogram, get_whitenoise, nested_params, TestBaseMixin
@unittest.skipIf(not LIBROSA_AVAILABLE, "Librosa not available")
class Functional(TestBaseMixin):
"""Test suite for functions in `functional` module."""
dtype = torch.float64
@nested_params([0, 0.99])
def test_griffinlim(self, momentum):
# FFT params
n_fft = 400
win_length = n_fft
hop_length = n_fft // 4
window = torch.hann_window(win_length, device=self.device)
power = 1
# GriffinLim params
n_iter = 8
waveform = get_whitenoise(device=self.device, dtype=self.dtype)
specgram = get_spectrogram(
waveform, n_fft=n_fft, hop_length=hop_length, power=power, win_length=win_length, window=window
)
result = F.griffinlim(
specgram,
window=window,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
power=power,
n_iter=n_iter,
momentum=momentum,
length=waveform.size(1),
rand_init=False,
)
expected = librosa.griffinlim(
specgram[0].cpu().numpy(),
n_iter=n_iter,
hop_length=hop_length,
momentum=momentum,
init=None,
length=waveform.size(1),
pad_mode="reflect",
)[None, ...]
self.assertEqual(result, torch.from_numpy(expected), atol=5e-5, rtol=1e-07)
@nested_params(
[
param(),
param(n_mels=128, sample_rate=44100),
param(n_mels=128, fmin=2000.0, fmax=5000.0),
param(n_mels=56, fmin=100.0, fmax=9000.0),
param(n_mels=56, fmin=800.0, fmax=900.0),
param(n_mels=56, fmin=1900.0, fmax=900.0),
param(n_mels=10, fmin=1900.0, fmax=900.0),
],
[param(norm=n) for n in [None, "slaney"]],
[param(mel_scale=s) for s in ["htk", "slaney"]],
)
def test_create_mel_fb(
self, n_mels=40, sample_rate=22050, n_fft=2048, fmin=0.0, fmax=8000.0, norm=None, mel_scale="htk"
):
if norm == "slaney" and StrictVersion(librosa.__version__) < StrictVersion("0.7.2"):
self.skipTest("Test is known to fail with older versions of librosa.")
if self.device != "cpu":
self.skipTest("No need to run this test on CUDA")
expected = librosa.filters.mel(
sr=sample_rate, n_fft=n_fft, n_mels=n_mels, fmax=fmax, fmin=fmin, htk=mel_scale == "htk", norm=norm
).T
result = F.melscale_fbanks(
sample_rate=sample_rate,
n_mels=n_mels,
f_max=fmax,
f_min=fmin,
n_freqs=(n_fft // 2 + 1),
norm=norm,
mel_scale=mel_scale,
)
self.assertEqual(result, torch.from_numpy(expected), atol=7e-5, rtol=1.3e-6)
def test_amplitude_to_DB_power(self):
amin = 1e-10
db_multiplier = 0.0
top_db = 80.0
multiplier = 10.0
spec = get_spectrogram(get_whitenoise(device=self.device, dtype=self.dtype), power=2)
result = F.amplitude_to_DB(spec, multiplier, amin, db_multiplier, top_db)
expected = librosa.core.power_to_db(spec[0].cpu().numpy())[None, ...]
self.assertEqual(result, torch.from_numpy(expected))
def test_amplitude_to_DB(self):
amin = 1e-10
db_multiplier = 0.0
top_db = 80.0
multiplier = 20.0
spec = get_spectrogram(get_whitenoise(device=self.device, dtype=self.dtype), power=1)
result = F.amplitude_to_DB(spec, multiplier, amin, db_multiplier, top_db)
expected = librosa.core.amplitude_to_db(spec[0].cpu().numpy())[None, ...]
self.assertEqual(result, torch.from_numpy(expected))
@unittest.skipIf(not LIBROSA_AVAILABLE, "Librosa not available")
class FunctionalComplex(TestBaseMixin):
@nested_params([0.5, 1.01, 1.3])
def test_phase_vocoder(self, rate):
hop_length = 256
num_freq = 1025
num_frames = 400
# Due to cummulative sum, numerical error in using torch.float32 will
# result in bottom right values of the stretched sectrogram to not
# match with librosa.
spec = torch.randn(num_freq, num_frames, device=self.device, dtype=torch.complex128)
phase_advance = torch.linspace(0, np.pi * hop_length, num_freq, device=self.device, dtype=torch.float64)[
..., None
]
stretched = F.phase_vocoder(spec, rate=rate, phase_advance=phase_advance)
expected_stretched = librosa.phase_vocoder(spec.cpu().numpy(), rate=rate, hop_length=hop_length)
self.assertEqual(stretched, torch.from_numpy(expected_stretched))
|
