-
Notifications
You must be signed in to change notification settings - Fork 0
/
gsc_dataset.py
166 lines (133 loc) · 6.42 KB
/
gsc_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
import torch
import torchaudio
import torch.nn.functional as F
from torch.utils.data import Dataset
import random
import numpy as np
class PDMEncodeur_seq(torch.nn.Module):
def __init__(self, pdm_factor=10, orig_freq=16000):
super().__init__()
self.pdm_factor = pdm_factor
self.upsampler = torchaudio.transforms.Resample(orig_freq=orig_freq, new_freq=orig_freq*pdm_factor)
def to(self, dest):
super().to(dest)
self.upsampler = self.upsampler.to(dest)
return self
def forward(self, waveform):
waveform = (waveform/2)+0.5
if self.pdm_factor !=1: waveform = self.upsampler(waveform)
spikes = torch.zeros_like(waveform)
error = torch.zeros_like(waveform[:,0])
for i in range(waveform.shape[1]):
error += waveform[:,i]
spikes[:,i] = error>0
error -= spikes[:,i]
return spikes
class PDMEncodeur(torch.nn.Module):
def __init__(self, pdm_factor=10, orig_freq=16000):
super().__init__()
self.pdm_factor = pdm_factor
self.upsampler = torchaudio.transforms.Resample(orig_freq=orig_freq, new_freq=orig_freq*pdm_factor)
self.th = 1.
def to(self, dest):
super().to(dest)
self.upsampler = self.upsampler.to(dest)
return self
def forward(self, waveform):
waveform = (waveform/2)+0.5
if self.pdm_factor !=1: waveform = self.upsampler(waveform)
spikes = torch.zeros_like(waveform)
waveform = waveform.to(torch.float64)
waveform_cumsum = torch.cumsum(waveform, dim=1)
waveform_div = waveform_cumsum//self.th
waveform_div_diff = waveform_div-F.pad(waveform_div[:,:-1], (1,0), value=-1)
spikes[waveform_div_diff>0] = 1.
return spikes
class DataAug(torch.nn.Module):
def __init__(self, shift_factor=0.1, sample_rate=16000):
super().__init__()
self.shift_factor = int(shift_factor*sample_rate)
def shift(self, waveform):
shift_factor = random.randint(-self.shift_factor, self.shift_factor)
if shift_factor>0: return F.pad(waveform[:,:-shift_factor], (shift_factor,0))
elif shift_factor<0: return F.pad(waveform[:,-shift_factor:], (0,-shift_factor))
else: return waveform
def forward(self, waveform):
waveform = self.shift(waveform)
return waveform
"""
For all subsets (train, valid, test) the individuals waveforms are loaded with torchaudio, stacked and saved as a pytorch tensor (train_data.pt, valid_data.pt, test_data.pt)
For all subsets (train, valid, test) the corresponding labels are stacked and saved as numpy arrays (train_targets.npy, valid_targets.npy, test_targets.npy)
"""
class InMemoryGSCDataset(Dataset):
def __init__(self, subset="train", root="./Data/", transform=None, n_examples=None, pdm_factor=10, device=None, **kwargs):
super().__init__()
self.labels = ['backward', 'bed', 'bird', 'cat', 'dog', 'down', 'eight', 'five', 'follow', 'forward', 'four', 'go', 'happy', 'house', 'learn', 'left', 'marvin',
'nine', 'no', 'off', 'on', 'one', 'right', 'seven', 'sheila', 'six', 'stop', 'three', 'tree', 'two', 'up', 'visual', 'wow', 'yes', 'zero']
self.sc_sample_rate = 16000
self.sc_duration_s = 1
self.device = device
self.pdm_factor = pdm_factor
self.pdm_encodeur = None
self.transform = transform
self.n_examples = n_examples
self.waveforms = torch.load(root+f"SpeechCommands/{subset}_data.pt")
labels_ = np.load(root+f"SpeechCommands/{subset}_targets.npy")
self.label_indexes = torch.tensor([self.labels.index(l) for l in labels_])
if self.pdm_factor>0:
self.pdm_encodeur = PDMEncodeur(self.pdm_factor, self.sc_sample_rate).to(device)
if self.n_examples and self.n_examples<1 and self.n_examples>0: self.reduce_size()
self.label_indexes = self.label_indexes.to(device)
def reduce_size(self):
part_len = int(self.n_examples*len(self.label_indexes))
ind = np.arange(len(self.label_indexes))
rng = np.random.default_rng(0)
rng.shuffle(ind)
ind = ind[:part_len]
self.waveforms = torch.stack([self.waveforms[i] for i in ind])
self.label_indexes = torch.stack([self.label_indexes[i] for i in ind])
def to(self, device):
self.device = device
if self.transform: self.transform = self.transform.to(device)
if self.pdm_encodeur: self.pdm_encodeur = self.pdm_encodeur.to(device)
if hasattr(self, 'waveforms'): self.waveforms = self.waveforms.to(device)
if hasattr(self, 'label_indexes'): self.label_indexes = self.label_indexes.to(device)
return self
def __getitem__(self, n):
waveform = self.waveforms[n]
waveform = waveform.to(self.device)
if self.transform:
waveform = self.transform(waveform)
if self.pdm_encodeur:
waveform = self.pdm_encodeur(waveform)
return waveform, self.label_indexes[n]
def __len__(self):
return len(self.label_indexes)
if __name__=='__main__':
import os
os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--n_examples', type=int, default=None)
parser.add_argument("--data_dir", type=str, default="./Data/")
parser.add_argument("--transform", type=str, default=None)
parser.add_argument("--pdm_factor", type=int, default=0)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
args = parser.parse_args()
dataset = InMemoryGSCDataset(root=args.data_dir, subset="test",
transform=args.transform, pdm_factor=args.pdm_factor,
n_examples=args.n_examples, device=device)
len(dataset)
len(dataset.labels)
class_weights = torch.unique(dataset.label_indexes, return_counts=True)[1]/len(dataset)
sample_weights = class_weights[dataset.label_indexes].tolist()
def plot_waveform(x, ax):
ax.plot(x.cpu().squeeze())
ax.set_xlabel('Sample Index')
import matplotlib.pylab as plt
for example_index in torch.randint(len(dataset), size=(2,)):
waveform, index = dataset[example_index]
fig, ax = plt.subplots(1, 1)
plot_waveform(waveform, ax)
#ax.set_title(dataset.labels[index])
plt.show()