-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathutil.py
351 lines (273 loc) · 13.8 KB
/
util.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
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
import os
import sys
import glob
import pickle
import numpy as np
import pandas as pd
import librosa
from datetime import datetime
from skimage.transform import resize
def load_audio_data(filepath, orig_sr=44100, target_sr=16000, framing=False, frame_size=30):
y, sr = librosa.load(filepath, sr=orig_sr)
y_ds = librosa.resample(y, orig_sr, target_sr)
length = frame_size*target_sr
if framing:
output = librosa.util.frame(y_ds, frame_length=length, hop_length=length)
else:
output = librosa.util.fix_length(y_ds, length)
return output
def get_X_and_Y():
pass
def build_vectors(s, l):
X_samples = []
Y_samples = []
if np.ndim(s["bass"]) > 1:
for idx in range(s["bass"].shape[1]):
X_samples.append(np.dstack((s["bass"][:,idx], s["drums"][:,idx], s["other"][:,idx], s["vocals"][:,idx])))
Y_samples.append(np.array((float(l['drums ratio']), float(l['other ratio']), float(l['vocals ratio']))))
else:
X_samples.append(np.dstack((s["bass"], s["drums"], s["other"], s["vocals"])))
Y_samples.append(np.array((float(l['drums ratio']), float(l['other ratio']), float(l['vocals ratio']))))
return X_samples, Y_samples
def load_dataset(augmented_data=False):
X_train = [] # (n, 4, 480000)
Y_train = [] # (n, 3)
lev = pd.read_csv("data/level_analysis.csv")
# load training audio and level data
train_path = os.path.join("DSD100", "Sources", "train", "*")
for idx, song in enumerate(glob.glob(train_path)):
track_id = int(os.path.basename(song)[0:4])
s = {"bass" : None, "drums" : None, "other" : None, "vocals" : None}
l = lev[(lev["augment type"] == "None") & (lev["track id"] == track_id)]
sys.stdout.write(" Loading song {: <20}\r".format(track_id))
sys.stdout.flush()
for stem in glob.glob(os.path.join(song, "normalized", "*.wav")):
stem_class = os.path.basename(stem).replace(".wav", "")
s[stem_class] = load_audio_data(stem)
X_samples, Y_samples = build_vectors(s, l)
X_train += X_samples
Y_train += Y_samples
if augmented_data:
for augment in glob.glob(os.path.join(song, "augmented", "*")):
augment_type = os.path.basename(augment)
l = lev[(lev["augment type"] == augment_type) & (lev["track id"] == track_id)]
for stem in glob.glob(os.path.join(augment, "normalized", "*.wav")):
stem_class = os.path.basename(stem).replace(".wav", "")
s[stem_class] = load_audio_data(stem)
X_samples, Y_samples = build_vectors(s, l)
X_train += X_samples
Y_train += Y_samples
X_train = np.vstack(X_train)
X_train = np.reshape(X_train, (X_train.shape[0], X_train.shape[2], X_train.shape[1]))
Y_train = np.vstack(Y_train)
X_test = []
Y_test = []
# load testing audio and level data
train_path = os.path.join("DSD100", "Sources", "test", "*")
for idx, song in enumerate(glob.glob(train_path)):
track_id = int(os.path.basename(song)[0:4])
s = {"bass" : None, "drums" : None, "other" : None, "vocals" : None}
l = lev[(lev["augment type"] == "None") & (lev["track id"] == track_id)]
sys.stdout.write(" Loading song {: <20}\r".format(track_id))
sys.stdout.flush()
for stem in glob.glob(os.path.join(song, "normalized", "*.wav")):
stem_class = os.path.basename(stem).replace(".wav", "")
s[stem_class] = load_audio_data(stem)
X_samples, Y_samples = build_vectors(s, l)
X_test += X_samples
Y_test += Y_samples
if augmented_data:
for augment in glob.glob(os.path.join(song, "augmented", "*")):
augment_type = os.path.basename(augment)
l = lev[(lev["augment type"] == augment_type) & (lev["track id"] == track_id)]
for stem in glob.glob(os.path.join(augment, "normalized", "*.wav")):
stem_class = os.path.basename(stem).replace(".wav", "")
s[stem_class] = load_audio_data(stem)
X_samples, Y_samples = build_vectors(s, l)
X_test += X_samples
Y_test += Y_samples
X_test = np.vstack(X_test)
X_test = np.reshape(X_test, (X_test.shape[0], X_test.shape[2], X_test.shape[1]))
Y_test = np.vstack(Y_test)
print(X_train.shape)
print(Y_train.shape)
return X_train, Y_train, X_test, Y_test
def load_song_data(window_size):
#key = "{0} {1} {2}".format(spect_type, spect_size, hop_size)
key = "mel 1024 1024"
# set data split
train_ids = np.arange(21, 100+1)
val_ids = np.arange(11, 20+1)
test_ids = np.arange(1, 10+1)
# set silence threshold
lim = 1e-6
discarded = 0
song_data = []
# load dataset
for idx, song in enumerate(glob.glob("data/*.pkl")):
# data holders
X = []
Y = []
track_id = int(os.path.basename(song).split("_")[2].strip(".pkl"))
# load song data
row = pickle.load(open(song, "rb"))
n_frames = np.floor((row['bass ' + key].shape[1])/window_size).astype('int')
for frame in range(n_frames):
start_idx = frame*window_size
end_idx = start_idx+window_size
bass_spect = row['bass ' + key][:, start_idx:end_idx]
drums_spect = row['drums ' + key][:, start_idx:end_idx]
other_spect = row['other ' + key][:, start_idx:end_idx]
vocals_spect = row['vocals ' + key][:, start_idx:end_idx]
b_mean = np.mean(bass_spect, axis=(0,1))
d_mean = np.mean(drums_spect, axis=(0,1))
o_mean = np.mean(other_spect, axis=(0,1))
v_mean = np.mean(vocals_spect, axis=(0,1))
if b_mean > lim and b_mean > lim and o_mean > lim and v_mean > lim:
if track_id in train_ids:
data_type = "train"
X.append(np.dstack((bass_spect, drums_spect, other_spect, vocals_spect)))
Y.append(np.array((row['drums ratio'], row['other ratio'], row['vocals ratio'])))
elif track_id in val_ids:
data_type = "val"
X.append(np.dstack((bass_spect, drums_spect, other_spect, vocals_spect)))
Y.append(np.array((row['drums ratio'], row['other ratio'], row['vocals ratio'])))
elif track_id in test_ids:
data_type = "test"
X.append(np.dstack((bass_spect, drums_spect, other_spect, vocals_spect)))
Y.append(np.array((row['drums ratio'], row['other ratio'], row['vocals ratio'])))
else:
discarded += 1
if len(X) > 0:
song_data.append({"track id" : track_id, "type" : data_type,
"X" : np.array(X), "Y" : np.array(Y)})
sys.stdout.write("Loaded songs: {:03d}\r".format(idx+1))
sys.stdout.flush()
print("\nDiscarded {0:d} frames with energy below the threshold.".format(discarded))
return song_data
def load_data(spect_type='mel', spect_size='1024', hop_size='1024', framing=True, window_size=128, resizing=False):
key = "{0} {1} {2}".format(spect_type, spect_size, hop_size)
train_ids = np.arange(21, 100+1)
val_ids = np.arange(11, 20+1)
test_ids = np.arange(1, 10+1)
# set silence threshold
lim = 1e-6
x_train = []
y_train = []
x_val = []
y_val = []
x_test = []
y_test = []
discarded = 0 # number of frames discarded
if framing:
for idx, song in enumerate(glob.glob("data/*.pkl")):
row = pickle.load(open(song, "rb"))
n_frames = np.floor((row['bass ' + key].shape[1])/window_size).astype('int')
track_id = int(os.path.basename(song).split("_")[2].strip(".pkl"))
for frame in range(n_frames):
start_idx = frame*window_size
end_idx = start_idx+window_size
bass_spect = row['bass ' + key][:, start_idx:end_idx]
drums_spect = row['drums ' + key][:, start_idx:end_idx]
other_spect = row['other ' + key][:, start_idx:end_idx]
vocals_spect = row['vocals ' + key][:, start_idx:end_idx]
if resizing:
bass_spect = resize(bass_spect, (128, 128), anti_aliasing=True)
drums_spect = resize(drums_spect, (128, 128), anti_aliasing=True)
other_spect = resize(other_spect, (128, 128), anti_aliasing=True)
vocals_spect = resize(vocals_spect, (128, 128), anti_aliasing=True)
b_mean = np.mean(bass_spect, axis=(0,1))
d_mean = np.mean(drums_spect, axis=(0,1))
o_mean = np.mean(other_spect, axis=(0,1))
v_mean = np.mean(vocals_spect, axis=(0,1))
if b_mean > lim and b_mean > lim and o_mean > lim and v_mean > lim:
if track_id in train_ids:
x_train.append(np.dstack((bass_spect, drums_spect, other_spect, vocals_spect)))
y_train.append(np.array((row['drums ratio'], row['other ratio'], row['vocals ratio'])))
elif track_id in val_ids:
x_val.append(np.dstack((bass_spect, drums_spect, other_spect, vocals_spect)))
y_val.append(np.array((row['drums ratio'], row['other ratio'], row['vocals ratio'])))
elif track_id in test_ids:
x_test.append(np.dstack((bass_spect, drums_spect, other_spect, vocals_spect)))
y_test.append(np.array((row['drums ratio'], row['other ratio'], row['vocals ratio'])))
else:
discarded += 1
sys.stdout.write("Loaded songs: {:03d}\r".format(idx+1))
sys.stdout.flush()
print("\nDiscarded {0:d} frames with energy below the threshold.".format(discarded))
else:
for idx, song in enumerate(glob.glob("data/*.pkl")):
row = pickle.load(open(song, "rb"))
y_rows.append(np.array((row['drums ratio'], row['other ratio'], row['vocals ratio'])))
bass_spect = row['bass ' + key][:, :window_size]
drums_spect = row['drums ' + key][:, :window_size]
other_spect = row['other ' + key][:, :window_size]
vocals_spect = row['vocals ' + key][:, :window_size]
x_rows.append(np.dstack((bass_spect, drums_spect, other_spect, vocals_spect)))
# transform into numpy arrays
X_train = np.array(x_train)
Y_train = np.array(y_train)
X_val = np.array(x_val)
Y_val = np.array(y_val)
X_test = np.array(x_test)
Y_test = np.array(y_test)
# remove nans
#X = np.nan_to_num(X)
input_shape = (X_train.shape[1], X_train.shape[2], 4) # four instruments - 1 per channel
print("\n=============== Training Data ==============")
print("Loaded inputs with shape:", X_train.shape)
print("Loaded outputs with shape:", Y_train.shape)
print("\n============== Validation Data =============")
print("Loaded inputs with shape:", X_val.shape)
print("Loaded outputs with shape:", Y_val.shape)
print("\n=============== Testing Data ===============")
print("Loaded inputs with shape:", X_test.shape)
print("Loaded outputs with shape:", Y_test.shape)
return X_train, Y_train, X_val, Y_val, X_test, Y_test, input_shape
def standardize(X_train, X_val, X_test):
#X_train_mean = np.mean(X_train, axis=0)
#X_train_std = np.std(X_train, axis=0)
#X_train -= X_train_mean # zero-center
#X_train /= X_train_std # normalize
#X_val -= X_train_mean # zero-center
#X_val /= X_train_std # normalize
#X_test -= X_train_mean # zero-center
#X_test /= X_train_std # normalize
X_max = np.max(X_train)
X_min = np.min(X_train)
X_train -= X_min
X_train /= (X_max - X_min)
X_val -= X_min
X_val /= (X_max - X_min)
X_test -= X_min
X_test /= (X_max - X_min)
return X_train, X_val, X_test
def generate_report(report_dir, r):
with open(os.path.join(report_dir, "report_summary.txt"), 'w') as results:
results.write("--- RUNTIME ---\n")
results.write("Start time: {}\n".format(r["start time"]))
results.write("End time: {}\n".format(r["end time"]))
results.write("Runtime: {}\n\n".format(r["elapsed time"]))
results.write("--- MSE RESULTS ---\n")
val_losses = []
for track_id, fold in r["training history"].items():
results.write("Validation results for track {}\n".format(track_id))
for epoch, val_loss in enumerate(fold["val_loss"]):
results.write("Epoch {0}: {1:0.6f}\n".format(epoch+1, val_loss))
val_losses.append(val_loss)
final_loss = np.mean(val_losses)
results.write("Final average validation loss: {}\n".format(final_loss))
results.write("\n--- TRAINING DETAILS ---\n")
results.write("Batch size: {0}\n".format(r["batch size"]))
results.write("Epochs: {0}\n".format(r["epochs"]))
results.write("Input shape: {0}\n".format(r["input shape"]))
results.write("Training type: {0}\n".format(r["train"]))
results.write("Folds: {0:d}\n".format(r["folds"]))
#results.write("Training split: {0:d}/{1:d}\n".format(split, X.shape[0]-split))
results.write("Learning rate: {0:f}\n".format(r["learning rate"]))
results.write("Spectrogram type: {0}\n".format(r["spect type"]))
results.write("Spectrogram size: {0}\n".format(r["spect size"]))
results.write("Standardize: {0}\n\n".format(r["standard"]))
results.write("\n--- NETWORK ARCHITECTURE ---\n")
r["model"].summary(print_fn=lambda x: results.write(x + '\n'))
return final_loss