-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdataset.py
167 lines (148 loc) · 7.23 KB
/
dataset.py
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
157
158
159
160
161
162
163
164
165
166
167
import os
import random
import matplotlib.pyplot as plt
import sounddevice as sd
import torch
import torchaudio
import torchaudio.transforms as T
from torch.utils.data import DataLoader, Subset
from torchaudio.datasets import LIBRISPEECH
import torchaudio.functional as F
from tqdm import tqdm
class GetDataset:
"""
Handles dataset loading, STFT transformations, noise addition, and data preprocessing for speech enhancement.
"""
def __init__(self, root="data/", url="train-clean-100", sample_rate=16000, n_fft=254, hop_length=32, win_length=128, max_length_seconds=1, device="cpu"):
os.makedirs(root, exist_ok=True)
self.dataset = LIBRISPEECH(root=root, url=url, download=True)
self.device = device
self.sample_rate = sample_rate
self.n_fft = n_fft
self.hop_length = hop_length
self.win_length = win_length
self.max_length = int(max_length_seconds * sample_rate) + (12 * hop_length) - 1
self.alpha = 0.5
self.beta = 1
self.snr_db = 1
self.max_length_seconds = max_length_seconds
self.stft = T.Spectrogram(n_fft=n_fft, hop_length=hop_length, win_length=win_length, power=None, onesided=True)
self.istft = T.InverseSpectrogram(n_fft=n_fft, hop_length=hop_length, win_length=win_length, onesided=True)
self.example_spec, _ = self[10]
self.complex_shape = (1, *self.example_spec.shape[1:])
self.real_shape = (2, *self.example_spec.shape[1:])
def __len__(self):
return len(self.dataset)
def __getitem__(self, idx):
waveform, sr, _, _, _, _ = self.dataset[idx]
if sr != self.sample_rate:
resampler = T.Resample(sr, self.sample_rate)
waveform = resampler(waveform)
if waveform.shape[1] < self.max_length:
pad = self.max_length - waveform.shape[1]
waveform = torch.nn.functional.pad(waveform, (0, pad))
else:
max_start = waveform.shape[1] - int(self.max_length)
start = torch.randint(0, max_start + 1, (1,)).item()
waveform = waveform[:, start : start + self.max_length]
noise = torch.randn_like(waveform) * (torch.std(waveform) / (self.snr_db + 1e-10))
signal_power = torch.mean(waveform ** 2)
noise_power = torch.mean(noise ** 2)
snr_linear = 10 ** (self.snr_db / 10)
eps = 1e-10
noise = noise * torch.sqrt(signal_power / (snr_linear * (noise_power + eps)))
noised_waveform = waveform + noise
spectrogram = self.complex_normalize(self.stft(waveform))
noised_spectrogram = self.complex_normalize(self.stft(noised_waveform))
return spectrogram, noised_spectrogram
def get_dataloader(self, batch_size=16, shuffle=True):
#clamped = Subset(self, range(2))
#return DataLoader(clamped, batch_size=batch_size, shuffle=shuffle)
return DataLoader(self, batch_size=batch_size, shuffle=shuffle)
def complex_to_real(self, tensor):
assert tensor.dim() == 4
real_tensor = torch.view_as_real(tensor).squeeze(1)
return real_tensor.permute(0, 3, 1, 2)
def real_to_complex(self, tensor):
assert tensor.dim() == 4
tensor = tensor.permute(0, 2, 3, 1).contiguous()
return torch.view_as_complex(tensor)
def complex_normalize(self, tensor):
magnitude, phase = torch.abs(tensor), torch.angle(tensor)
return torch.polar(self.beta * magnitude ** self.alpha, phase)
def complex_denormalize(self, tensor):
magnitude, phase = torch.abs(tensor), torch.angle(tensor)
return torch.polar((magnitude / self.beta) ** (1 / self.alpha), phase)
def reconstruct_phase_istft(self, sampled_spectrogram):
return self.istft(self.complex_denormalize(self.real_to_complex(sampled_spectrogram)))
def reconstruct_phase_griffinlim(self, sampled_spectrogram):
"""
Uses Griffin-Lim algorithm to reconstruct phase from magnitude.
"""
sampled_complex = self.real_to_complex(sampled_spectrogram)
sampled_complex = self.complex_denormalize(sampled_complex)
magnitude = torch.abs(sampled_complex)
reconstructed_waveform = F.griffinlim(
magnitude,
n_fft=self.n_fft,
win_length=self.win_length,
hop_length=self.hop_length,
window=torch.hann_window(self.win_length, device=magnitude.device),
power=1.0,
n_iter=64,
momentum=0.99,
length=None,
rand_init=True
)
return reconstructed_waveform
def reconstruct_phase_noisy(self, sampled_spectrogram, noisy_spectrogram, threshold=.001):
"""
Uses noisy input's phase for bins where the predicted magnitude is above a threshold.
"""
sampled_complex = self.real_to_complex(sampled_spectrogram)
sampled_complex = self.complex_denormalize(sampled_complex)
noisy_complex = self.real_to_complex(noisy_spectrogram)
noisy_complex = self.complex_denormalize(noisy_complex)
sampled_magnitude = torch.abs(sampled_complex)
noisy_phase = torch.angle(noisy_complex)
reconstructed_complex = torch.polar(sampled_magnitude, noisy_phase)
return self.istft(reconstructed_complex)
def get_test_batch(self, batch_size):
length = self.sample_rate * 2 + (self.hop_length * 23)
inputs, targets = [], []
for i in [random.randint(0, len(self.dataset)) for _ in range(batch_size)]:
waveform, sr, _, _, _, _ = self.dataset[i]
if waveform.shape[1] < length:
waveform = torch.nn.functional.pad(waveform, (0, length - waveform.shape[1]))
else:
max_start = waveform.shape[1] - int(length)
start = torch.randint(0, max_start + 1, (1,)).item()
waveform = waveform[:, start : start + length]
noise = torch.randn_like(waveform) * (torch.std(waveform) / (self.snr_db + 1e-10))
signal_power = torch.mean(waveform ** 2)
noise_power = torch.mean(noise ** 2)
snr_linear = 10 ** (self.snr_db / 10)
eps = 1e-10
noise = noise * torch.sqrt(signal_power / (snr_linear * (noise_power + eps)))
noised_waveform = waveform + noise
spectrogram = self.complex_normalize(self.stft(waveform))
noised_spectrogram = self.complex_normalize(self.stft(noised_waveform))
inputs.append(noised_spectrogram)
targets.append(spectrogram)
return torch.stack(targets), torch.stack(inputs)
def print_info(self):
out = (
"\nSTFT parameters:\n"
f"\tSample Rate: {self.sample_rate}\n"
f"\tn_fft: {self.n_fft}\n"
f"\tWindow: {self.win_length}\n"
f"\tHop Length: {self.hop_length}\n\n"
"Data:\n"
f"\tComplex shape: {self.complex_shape}\n"
f"\tReal shape: {self.real_shape}\n"
f"\tNumber of datapoints: {len(self)}\n"
f"\tDatapoint length (seconds): {self.max_length_seconds}\n"
f"\tSample max: {self.example_spec.detach().numpy().max()}\n"
f"\tSample min: {self.example_spec.detach().numpy().min()}\n"
)
print(out)