-
Notifications
You must be signed in to change notification settings - Fork 40
/
dataset.py
177 lines (150 loc) · 5.45 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
168
169
170
171
172
173
174
175
176
177
import json
import os
import numpy as np
from torch.utils.data import Dataset
from text import text_to_sequence
from utils.tools import pad_1D, pad_2D
import pickle
class Dataset(Dataset):
def __init__(
self, filename, preprocess_config, train_config, sort=False, drop_last=False
):
self.dataset_name = preprocess_config["dataset"]
self.preprocessed_path = preprocess_config["path"]["preprocessed_path"]
self.cleaners = preprocess_config["preprocessing"]["text"]["text_cleaners"]
self.batch_size = train_config["optimizer"]["batch_size"]
self.expanded_char_path = preprocess_config["path"]["expanded_char_path"]
self.basename, self.speaker, self.text, self.raw_text = self.process_meta(filename)
with open(os.path.join(self.preprocessed_path, "prosody_phone.json"), 'r') as f:
self.phones_prosody = json.load(f)
with open(os.path.join(self.preprocessed_path, "speakers.json")) as f:
self.speaker_map = json.load(f)
with open(self.expanded_char_path, 'rb') as f:
self.expanded_chars = pickle.load(f)
self.sort = sort
self.drop_last = drop_last
def __len__(self):
return len(self.text)
def __getitem__(self, idx):
basename = self.basename[idx]
speaker = self.speaker[idx]
speaker_id = self.speaker_map[speaker]
raw_text = self.raw_text[idx]
phone = np.array(text_to_sequence(self.text[idx], self.cleaners))
mel_path = os.path.join(
self.preprocessed_path,
"mel",
"{}-mel-{}.npy".format(speaker, basename),
)
mel = np.load(mel_path)
pitch_path = os.path.join(
self.preprocessed_path,
"pitch",
"{}-pitch-{}.npy".format(speaker, basename),
)
pitch = np.load(pitch_path)
energy_path = os.path.join(
self.preprocessed_path,
"energy",
"{}-energy-{}.npy".format(speaker, basename),
)
energy = np.load(energy_path)
duration_path = os.path.join(
self.preprocessed_path,
"duration",
"{}-duration-{}.npy".format(speaker, basename),
)
duration = np.load(duration_path)
# add prosody
prosody = np.array(self.phones_prosody[basename])
# add char embedding
char_vec = self.expanded_chars[basename].numpy()
sample = {
"id": basename,
"speaker": speaker_id,
"text": phone,
"raw_text": raw_text,
"mel": mel,
"pitch": pitch,
"energy": energy,
"duration": duration,
"prosody": prosody,
"char_vec": char_vec
}
return sample
def process_meta(self, filename):
with open(
os.path.join(self.preprocessed_path, filename), "r", encoding="utf-8"
) as f:
name = []
speaker = []
text = []
raw_text = []
for line in f.readlines():
n, s, t, r = line.strip("\n").split("|")
if n == '002365':
continue
name.append(n)
speaker.append(s)
text.append(t)
raw_text.append(r)
return name, speaker, text, raw_text
def reprocess(self, data, idxs):
ids = [data[idx]["id"] for idx in idxs]
speakers = [data[idx]["speaker"] for idx in idxs]
texts = [data[idx]["text"] for idx in idxs]
raw_texts = [data[idx]["raw_text"] for idx in idxs]
mels = [data[idx]["mel"] for idx in idxs]
pitches = [data[idx]["pitch"] for idx in idxs]
energies = [data[idx]["energy"] for idx in idxs]
durations = [data[idx]["duration"] for idx in idxs]
# add prosody
prosodys = [data[idx]["prosody"] for idx in idxs]
# add char embedding
char_vecs = [data[idx]["char_vec"] for idx in idxs]
text_lens = np.array([text.shape[0] for text in texts])
mel_lens = np.array([mel.shape[0] for mel in mels])
speakers = np.array(speakers)
texts = pad_1D(texts)
mels = pad_2D(mels)
pitches = pad_1D(pitches)
energies = pad_1D(energies)
durations = pad_1D(durations)
# add prosody
prosodys = pad_1D(prosodys)
# add char embedding
char_vecs = pad_2D(char_vecs)
return (
ids,
raw_texts,
speakers,
texts,
text_lens,
max(text_lens),
mels,
mel_lens,
max(mel_lens),
pitches,
energies,
durations,
char_vecs,
prosodys
)
def collate_fn(self, data):
data_size = len(data)
if self.sort:
len_arr = np.array([d["text"].shape[0] for d in data])
idx_arr = np.argsort(-len_arr)
else:
idx_arr = np.arange(data_size)
tail = idx_arr[len(idx_arr) - (len(idx_arr) % self.batch_size) :]
idx_arr = idx_arr[: len(idx_arr) - (len(idx_arr) % self.batch_size)]
idx_arr = idx_arr.reshape((-1, self.batch_size)).tolist()
if not self.drop_last and len(tail) > 0:
idx_arr += [tail.tolist()]
output = list()
for idx in idx_arr:
output.append(self.reprocess(data, idx))
return output
if __name__ == "__main__":
pass