forked from marcoppasini/musika
-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils_encode.py
214 lines (172 loc) · 8.65 KB
/
utils_encode.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
import os
import numpy as np
import tensorflow as tf
from pydub import AudioSegment
from glob import glob
from tqdm import tqdm
from utils import Utils_functions
class UtilsEncode_functions:
def __init__(self, args):
self.args = args
self.U = Utils_functions(args)
self.paths = sorted(glob(self.args.files_path + "/*"))
def audio_generator(self):
for p in self.paths:
try:
tp, ext = os.path.splitext(p)
bname = os.path.basename(tp)
wvo = AudioSegment.from_file(p, format=ext[1:])
wvo = wvo.set_frame_rate(self.args.sr)
wvls = wvo.split_to_mono()
wvls = [s.get_array_of_samples() for s in wvls]
wv = np.array(wvls).T.astype(np.float32)
wv /= np.iinfo(wvls[0].typecode).max
yield np.squeeze(wv), bname
except Exception as e:
print(e)
print("Exception ignored! Continuing...")
pass
# def create_dataset(self):
# self.ds = (
# tf.data.Dataset.from_generator(
# self.audio_generator, output_signature=(tf.TensorSpec(shape=(None, 2), dtype=tf.float32))
# )
# .prefetch(tf.data.experimental.AUTOTUNE)
# .apply(tf.data.experimental.ignore_errors())
# )
def compress_files(self, models_ls=None):
critic, gen, enc, dec, enc2, dec2, gen_ema, [opt_dec, opt_disc], switch = models_ls
# self.create_dataset()
os.makedirs(self.args.save_path, exist_ok=True)
c = 0
time_compression_ratio = 16 # TODO: infer time compression ratio
shape2 = self.args.shape
pbar = tqdm(self.audio_generator(), position=0, leave=True, total=len(self.paths))
for (wv,bname) in pbar:
try:
if wv.shape[0] > self.args.hop * self.args.shape * 2 + 3 * self.args.hop:
split_limit = (
5 * 60 * self.args.sr
) # split very long waveforms (> 5 minutes) and process separately to avoid out of memory errors
nsplits = (wv.shape[0] // split_limit) + 1
wvsplits = []
for ns in range(nsplits):
if wv.shape[0] - (ns * split_limit) > self.args.hop * self.args.shape * 2 + 3 * self.args.hop:
wvsplits.append(wv[ns * split_limit : (ns + 1) * split_limit, :])
for wv in wvsplits:
wv = tf.image.random_crop(
wv,
size=[
(((wv.shape[0] - (3 * self.args.hop)) // (self.args.shape * self.args.hop)))
* self.args.shape
* self.args.hop
+ 3 * self.args.hop,
2,
],
)
chls = []
for channel in range(2):
x = wv[:, channel]
x = tf.expand_dims(tf.transpose(self.U.wv2spec(x, hop_size=self.args.hop), (1, 0)), -1)
ds = []
num = x.shape[1] // self.args.shape
rn = 0
for i in range(num):
ds.append(
x[:, rn + (i * self.args.shape) : rn + (i * self.args.shape) + self.args.shape, :]
)
del x
ds = tf.convert_to_tensor(ds, dtype=tf.float32)
lat = self.U.distribute_enc(ds, enc)
del ds
lat = tf.split(lat, lat.shape[0], 0)
lat = tf.concat(lat, -2)
lat = tf.squeeze(lat)
switch = False
if lat.shape[0] > (self.args.max_lat_len * time_compression_ratio):
switch = True
ds2 = []
num2 = lat.shape[-2] // shape2
rn2 = 0
for j in range(num2):
ds2.append(lat[rn2 + (j * shape2) : rn2 + (j * shape2) + shape2, :])
ds2 = tf.convert_to_tensor(ds2, dtype=tf.float32)
lat = self.U.distribute_enc(tf.expand_dims(ds2, -3), enc2)
del ds2
lat = tf.split(lat, lat.shape[0], 0)
lat = tf.concat(lat, -2)
lat = tf.squeeze(lat)
chls.append(lat)
if lat.shape[0] > self.args.max_lat_len and switch:
lat = tf.concat(chls, -1)
del chls
latc = lat[: (lat.shape[0] // self.args.max_lat_len) * self.args.max_lat_len, :]
latc = tf.split(latc, latc.shape[0] // self.args.max_lat_len, 0)
for el in latc:
np.save(self.args.save_path + f"/{bname}_{c}.npy", el)
c += 1
pbar.set_postfix({"Saved Files": c})
np.save(self.args.save_path + f"/{bname}_{c}.npy", lat[-self.args.max_lat_len :, :])
c += 1
pbar.set_postfix({"Saved Files": c})
del lat
del latc
except Exception as e:
print(e)
print("Exception ignored! Continuing...")
pass
def compress_whole_files(self, models_ls=None):
critic, gen, enc, dec, enc2, dec2, gen_ema, [opt_dec, opt_disc], switch = models_ls
# self.create_dataset()
os.makedirs(self.args.save_path, exist_ok=True)
c = 0
time_compression_ratio = 16 # TODO: infer time compression ratio
shape2 = self.args.shape
pbar = tqdm(self.audio_generator(), position=0, leave=True, total=len(self.paths))
for (wv,bname) in pbar:
try:
# wv_len_orig = wv.shape[0]
if wv.shape[0] > self.args.hop * self.args.shape * 2 + 3 * self.args.hop:
rem = (wv.shape[0] - (3 * self.args.hop)) % (self.args.shape * self.args.hop)
if rem != 0:
wv = tf.concat([wv, tf.zeros([rem,2], dtype=tf.float32)], 0)
chls = []
for channel in range(2):
x = wv[:, channel]
x = tf.expand_dims(tf.transpose(self.U.wv2spec(x, hop_size=self.args.hop), (1, 0)), -1)
ds = []
num = x.shape[1] // self.args.shape
rn = 0
for i in range(num):
ds.append(
x[:, rn + (i * self.args.shape) : rn + (i * self.args.shape) + self.args.shape, :]
)
del x
ds = tf.convert_to_tensor(ds, dtype=tf.float32)
lat = self.U.distribute_enc(ds, enc)
del ds
lat = tf.split(lat, lat.shape[0], 0)
lat = tf.concat(lat, -2)
lat = tf.squeeze(lat)
ds2 = []
num2 = lat.shape[-2] // shape2
rn2 = 0
for j in range(num2):
ds2.append(lat[rn2 + (j * shape2) : rn2 + (j * shape2) + shape2, :])
ds2 = tf.convert_to_tensor(ds2, dtype=tf.float32)
lat = self.U.distribute_enc(tf.expand_dims(ds2, -3), enc2)
del ds2
lat = tf.split(lat, lat.shape[0], 0)
lat = tf.concat(lat, -2)
lat = tf.squeeze(lat)
chls.append(lat)
lat = tf.concat(chls, -1)
del chls
np.save(self.args.save_path + f"/{bname}.npy", lat)
c += 1
pbar.set_postfix({"Saved Files": c})
del lat
except Exception as e:
print(e)
print("Exception ignored! Continuing...")
pass