forked from hutchresearch/fine-grained-cad
-
Notifications
You must be signed in to change notification settings - Fork 0
/
data.py
191 lines (156 loc) · 6.9 KB
/
data.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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
import os
import pickle
import random
import numpy as np
import pandas as pd
import torch
from torch.utils.data import Dataset, Sampler
NINE_WAY_MAP = {0: 0, 1: 1, 2: 2, 3: 3, 4: 4, 5: 5, 6: 6, 7: 7, 8: 8}
FIVE_WAY_MAP = {0: 4, 1: 0, 2: 0, 3: 0, 4: 0, 5: 1, 6: 1, 7: 3, 8: 2}
FOUR_WAY_MAP = {0: 3, 1: 0, 2: 0, 3: 0, 4: 0, 5: 0, 6: 0, 7: 2, 8: 1}
class CAAMLRawFrameDataset(Dataset):
def __init__(self, split, root, classes=9, seq_len=10*100, sample_rate=16_000, mask=None, mask_mode=None,
spacing=0, split_csv='/research/hutchinson/data/2019_ml_teaching/split.csv', return_session=False,
features='pase'):
# Read data and filter to split
df = pd.read_csv(split_csv)
df['split'] = df['split'].str.upper()
df = df.loc[df['split'] == split.upper()]
df = df.reset_index()
# Construct wav path and labels
mel_root = '/research/hutchinson/data/2019_ml_teaching/mel_features/normalized/'
path = lambda r, f: os.path.join(root, '/'.join(r['session'].split('_')), f)
prep = lambda r: os.path.join(mel_root, r['session'] + '.npy')
df['wav'] = df.apply(lambda r: path(r, 'signal.npy'), axis=1)
df['pre'] = df.apply(lambda r: prep(r), axis=1)
df['lab'] = df.apply(lambda r: pickle.load(open(path(r, 'labels.pkl'), 'rb')), axis=1)
df['len'] = df.apply(lambda r: len(r['lab']), axis=1)
df['start'] = 0
df['end'] = 0
label_map = {
9: NINE_WAY_MAP,
5: FIVE_WAY_MAP,
4: FOUR_WAY_MAP
}[classes]
# Create a indexing scheme for all possible sequence start positions
padding = (seq_len-spacing) // 2
l = 0
wavs = {}
pres = {}
for i, r in df.iterrows():
df.loc[i, 'start'] = l
l += r['len'] + 2*padding
rounding = (spacing - l % spacing) - 1
l += rounding
df.loc[i, 'len'] = r['len'] + 2*padding + rounding
df.loc[i, 'end'] = l
labs = np.vectorize(label_map.__getitem__)(r['lab'])
labs = np.concatenate([-np.ones(padding, dtype=np.long), labs, -np.ones(seq_len+rounding+padding,
dtype=np.long)])
sigs = np.load(r['wav'])
if 'pase' in features:
prepend = sigs[:, :160*padding][::-1] if padding > 0 else None
pstpend = sigs[:, -160*(padding+rounding+seq_len):][::-1]
sigs = np.concatenate([prepend, sigs, pstpend], axis=1) if prepend is not None else np.concatenate([sigs, pstpend], axis=1)
else:
sigs = None
prec = np.load(r['pre'], allow_pickle=True)
if 'mels' in features and 'prosody' in features:
prec = np.concatenate(prec, axis=0)
elif 'mels' in features:
prec = prec[0]
elif 'prosody' in features:
prec = prec[1]
else:
prec = None
if prec is not None:
prepend = prec[:, :padding][::-1] if padding > 0 else None
pstpend = prec[:, -(padding+rounding+seq_len):][::-1]
prec = np.concatenate([prepend, prec, pstpend], axis=1) if prepend is not None else np.concatenate([prec, pstpend], axis=1)
prec = prec[:, :len(labs)]
df.at[i, 'lab'] = labs
wavs[r['wav']] = sigs
pres[r['wav']] = prec
l += 1
self.data = df
self.wavs = wavs
self.pres = pres
self.n_samples = l
self.seq_len = seq_len
self.mask = mask
self.mask_mode = mask_mode
self.return_session = return_session
def __len__(self):
return self.n_samples
def __getitem__(self, idx):
# Retrieve row from dataframe
r = self.data.loc[(self.data['start'] <= idx) & (self.data['end'] >= idx)]
wav = self.wavs[r['wav'].item()]
pre = self.pres[r['wav'].item()]
lab = r['lab'].item()
# Get appropriate sequence
seq_start = idx - r['start'].item()
seq_end = seq_start + self.seq_len
lab = lab[seq_start:seq_end] if lab is not None else np.empty(0)
pre = pre[:, seq_start:seq_end] if pre is not None else np.empty(0)
wav = wav[:, seq_start*160:seq_end*160] if wav is not None else np.empty(0)
if self.mask is not None and self.mask_mode is not None:
time_mask = torch.FloatTensor(wav.shape).uniform_() < self.mask
if self.mask_mode == 'silence':
wav[time_mask] = 0
elif self.mask_mode == 'noise':
wav[time_mask] = torch.tanh(torch.FloatTensor(wav.shape).normal_()[time_mask])
elif self.mask_mode == 'sample':
wav[time_mask] = torch.LongTensor(wav.shape).random_(0, wav.size(1))[time_mask]
return (wav, pre, lab) if not self.return_session else (wav, pre, lab, r['session'].item())
class SpacedSequentialSampler(Sampler):
def __init__(self, data_source, spacing=1):
self.data_source = data_source
self.spacing = spacing
def __iter__(self):
n = len(self.data_source)
indicies = [i for i in range(n)][::self.spacing]
return iter(indicies)
def __len__(self):
return len(self.data_source) // self.spacing
class SpacedRandomSampler(Sampler):
def __init__(self, data_source, spacing=1):
self.data_source = data_source
self.spacing = spacing
self.offset = 0
self.n_samples = len(self.data_source) // self.spacing
def __iter__(self):
n = len(self.data_source)
indicies = [i for i in range(n)][self.offset::self.spacing]
self.n_samples = len(indicies)
self.offset = random.randrange(self.spacing)
random.shuffle(indicies)
return iter(indicies)
def __len__(self):
return self.n_samples
class SessionBatchSampler(Sampler):
def __init__(self, dataset, sampler, batch_size, drop_last):
self.dataset = dataset
self.sampler = sampler
self.batch_size = batch_size
self.drop_last = drop_last
def __iter__(self):
batch = []
_, _, _, cur_session = self.dataset[0]
for idx in self.sampler:
_, _, _, session = self.dataset[idx]
if cur_session != session and len(batch) != 0:
yield batch
batch = []
batch.append(idx)
if len(batch) == self.batch_size:
yield batch
batch = []
cur_session = session
if len(batch) > 0 and not self.drop_last:
yield batch
def __len__(self):
if self.drop_last:
return len(self.sampler) // self.batch_size
else:
return (len(self.sampler) + self.batch_size - 1) // self.batch_size