Datasets:

Size:
n<1K
ArXiv:
License:
monetjoe commited on
Commit
cb70f80
·
verified ·
1 Parent(s): 296ab1b

Update Guzheng_Tech99.py

Browse files
Files changed (1) hide show
  1. Guzheng_Tech99.py +97 -48
Guzheng_Tech99.py CHANGED
@@ -1,8 +1,10 @@
1
  import os
2
  import csv
3
  import random
 
4
  import datasets
5
  import numpy as np
 
6
  from glob import glob
7
 
8
  _NAMES = {
@@ -64,9 +66,15 @@ class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
64
  if self.config.name == "default"
65
  else datasets.Features(
66
  {
67
- "data": datasets.features.Array3D(
 
 
 
68
  dtype="float32", shape=(88, 258, 1)
69
  ),
 
 
 
70
  "label": datasets.features.Array2D(
71
  dtype="float32", shape=(7, 258)
72
  ),
@@ -83,7 +91,7 @@ class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
83
  square_sum = 0
84
  tfle = 0
85
  for i in range(len(data)):
86
- tfle += (data[i].sum(-1).sum(0) != 0).astype("int").sum()
87
  common_sum += data[i].sum(-1).sum(-1)
88
  square_sum += (data[i] ** 2).sum(-1).sum(-1)
89
 
@@ -92,20 +100,14 @@ class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
92
  std = np.sqrt(square_avg - common_avg**2)
93
  return common_avg, std
94
 
95
- def _norm(self, avg, std, data, size):
 
 
96
  avg = np.tile(avg.reshape((1, -1, 1, 1)), (size[0], 1, size[2], size[3]))
97
  std = np.tile(std.reshape((1, -1, 1, 1)), (size[0], 1, size[2], size[3]))
98
- data = (data - avg) / std
99
- return data
100
-
101
- def _load(self, wav_dir, csv_dir, groups, avg=None, std=None):
102
- # Return all [(audio address, corresponding to csv file address), ( , ), ...] list
103
- if std is None:
104
- std = np.array([None])
105
-
106
- if avg is None:
107
- avg = np.array([None])
108
 
 
109
  def files(wav_dir, csv_dir, group):
110
  flacs = sorted(glob(os.path.join(wav_dir, group, "*.flac")))
111
  if len(flacs) == 0:
@@ -122,12 +124,18 @@ class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
122
 
123
  return result
124
 
125
- # Returns the CQT of the input audio
126
- def logCQT(file):
127
- import librosa
 
 
 
 
 
 
128
 
129
- sr = _SAMPLE_RATE
130
- y, sr = librosa.load(file, sr=sr)
131
  # 帧长为32ms (1000ms/(16000/512) = 32ms), D2的频率是73.418
132
  cqt = librosa.cqt(
133
  y,
@@ -141,10 +149,21 @@ class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
141
  (1.0 / 80.0) * librosa.core.amplitude_to_db(np.abs(cqt), ref=np.max)
142
  ) + 1.0
143
 
 
 
 
 
 
 
 
 
 
 
 
144
  def chunk_data(f):
145
- s = int(_SAMPLE_RATE * _TIME_LENGTH / _HOP_LENGTH)
146
- xdata = np.transpose(f)
147
  x = []
 
 
148
  length = int(np.ceil((int(len(xdata) / s) + 1) * s))
149
  app = np.zeros((length - xdata.shape[0], xdata.shape[1]))
150
  xdata = np.concatenate((xdata, app), 0)
@@ -154,16 +173,16 @@ class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
154
 
155
  return np.array(x)
156
 
157
- def load_all(audio_path, csv_path):
158
  # Load audio features: The shape of cqt (88, 8520), 8520 is the number of frames on the time axis
159
- cqt = logCQT(audio_path)
 
 
 
160
  # Load the ground truth label
161
- hop = _HOP_LENGTH
162
  n_steps = cqt.shape[1]
163
- n_IPTs = 7
164
- technique = _NAMES
165
  IPT_label = np.zeros([n_IPTs, n_steps], dtype=int)
166
- with open(csv_path, "r") as f: # csv file for each audio
167
  reader = csv.DictReader(f, delimiter=",")
168
  for label in reader: # each note
169
  onset = float(label["onset_time"])
@@ -175,7 +194,12 @@ class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
175
  IPT_label[IPT, left:frame_right] = 1
176
 
177
  return dict(
178
- audiuo_path=audio_path, csv_path=csv_path, cqt=cqt, IPT_label=IPT_label
 
 
 
 
 
179
  )
180
 
181
  data = []
@@ -184,29 +208,33 @@ class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
184
  for input_files in files(wav_dir, csv_dir, group):
185
  data.append(load_all(*input_files))
186
 
187
- i = 0
188
- for dic in data:
189
- x = dic["cqt"]
190
- x = chunk_data(x)
191
  y_i = dic["IPT_label"]
192
  y_i = chunk_data(y_i)
193
  if i == 0:
194
- Xtr = x
 
 
195
  Ytr_i = y_i
196
- i += 1
197
 
198
  else:
199
- Xtr = np.concatenate([Xtr, x], axis=0)
 
 
200
  Ytr_i = np.concatenate([Ytr_i, y_i], axis=0)
201
 
202
  # Transform the shape of the input
203
- Xtr = np.expand_dims(Xtr, axis=3)
204
- # Calculate the mean and variance of the input
205
- if avg.all() == None and std.all() == None:
206
- avg, std = self._RoW_norm(Xtr)
207
  # Normalize
208
- Xtr = self._norm(avg, std, Xtr, Xtr.shape)
209
- return list(Xtr), list(Ytr_i)
 
 
210
 
211
  def _parse_csv_label(self, csv_file):
212
  label = []
@@ -259,19 +287,40 @@ class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
259
  testset.append(item)
260
 
261
  else:
262
- audio_dir = audio_files + "\\audio"
263
- csv_dir = csv_files + "\\label"
264
  X_train, Y_train = self._load(audio_dir, csv_dir, ["train"])
265
  X_valid, Y_valid = self._load(audio_dir, csv_dir, ["validation"])
266
  X_test, Y_test = self._load(audio_dir, csv_dir, ["test"])
267
- for i in range(len(X_train)):
268
- trainset.append({"data": X_train[i], "label": Y_train[i]})
 
 
 
 
 
 
 
269
 
270
- for i in range(len(X_valid)):
271
- validset.append({"data": X_valid[i], "label": Y_valid[i]})
 
 
 
 
 
 
 
272
 
273
- for i in range(len(X_test)):
274
- testset.append({"data": X_test[i], "label": Y_test[i]})
 
 
 
 
 
 
 
275
 
276
  random.shuffle(trainset)
277
  random.shuffle(validset)
 
1
  import os
2
  import csv
3
  import random
4
+ import librosa
5
  import datasets
6
  import numpy as np
7
+ from tqdm import tqdm
8
  from glob import glob
9
 
10
  _NAMES = {
 
66
  if self.config.name == "default"
67
  else datasets.Features(
68
  {
69
+ "mel": datasets.features.Array3D(
70
+ dtype="float32", shape=(128, 258, 1)
71
+ ),
72
+ "cqt": datasets.features.Array3D(
73
  dtype="float32", shape=(88, 258, 1)
74
  ),
75
+ "chroma": datasets.features.Array3D(
76
+ dtype="float32", shape=(12, 258, 1)
77
+ ),
78
  "label": datasets.features.Array2D(
79
  dtype="float32", shape=(7, 258)
80
  ),
 
91
  square_sum = 0
92
  tfle = 0
93
  for i in range(len(data)):
94
+ tfle += (data[i].sum(-1).sum(0) != 0).astype("float").sum()
95
  common_sum += data[i].sum(-1).sum(-1)
96
  square_sum += (data[i] ** 2).sum(-1).sum(-1)
97
 
 
100
  std = np.sqrt(square_avg - common_avg**2)
101
  return common_avg, std
102
 
103
+ def _norm(self, data):
104
+ size = data.shape
105
+ avg, std = self._RoW_norm(data)
106
  avg = np.tile(avg.reshape((1, -1, 1, 1)), (size[0], 1, size[2], size[3]))
107
  std = np.tile(std.reshape((1, -1, 1, 1)), (size[0], 1, size[2], size[3]))
108
+ return (data - avg) / std
 
 
 
 
 
 
 
 
 
109
 
110
+ def _load(self, wav_dir, csv_dir, groups):
111
  def files(wav_dir, csv_dir, group):
112
  flacs = sorted(glob(os.path.join(wav_dir, group, "*.flac")))
113
  if len(flacs) == 0:
 
124
 
125
  return result
126
 
127
+ def logMel(y, sr=_SAMPLE_RATE):
128
+ # 帧长为32ms (1000ms/(16000/512) = 32ms), D2的频率是73.418
129
+ mel = librosa.feature.melspectrogram(
130
+ y=y,
131
+ sr=sr,
132
+ hop_length=_HOP_LENGTH,
133
+ fmin=27.5,
134
+ )
135
+ return librosa.power_to_db(mel, ref=np.max)
136
 
137
+ # Returns the CQT of the input audio
138
+ def logCQT(y, sr=_SAMPLE_RATE):
139
  # 帧长为32ms (1000ms/(16000/512) = 32ms), D2的频率是73.418
140
  cqt = librosa.cqt(
141
  y,
 
149
  (1.0 / 80.0) * librosa.core.amplitude_to_db(np.abs(cqt), ref=np.max)
150
  ) + 1.0
151
 
152
+ def logChroma(y, sr=_SAMPLE_RATE):
153
+ # 帧长为32ms (1000ms/(16000/512) = 32ms), D2的频率是73.418
154
+ chroma = librosa.feature.chroma_stft(
155
+ y=y,
156
+ sr=sr,
157
+ hop_length=_HOP_LENGTH,
158
+ )
159
+ return (
160
+ (1.0 / 80.0) * librosa.core.amplitude_to_db(np.abs(chroma), ref=np.max)
161
+ ) + 1.0
162
+
163
  def chunk_data(f):
 
 
164
  x = []
165
+ xdata = np.transpose(f)
166
+ s = _SAMPLE_RATE * _TIME_LENGTH // _HOP_LENGTH
167
  length = int(np.ceil((int(len(xdata) / s) + 1) * s))
168
  app = np.zeros((length - xdata.shape[0], xdata.shape[1]))
169
  xdata = np.concatenate((xdata, app), 0)
 
173
 
174
  return np.array(x)
175
 
176
+ def load_all(audio_path, csv_path, hop=_HOP_LENGTH, n_IPTs=7, technique=_NAMES):
177
  # Load audio features: The shape of cqt (88, 8520), 8520 is the number of frames on the time axis
178
+ y, sr = librosa.load(audio_path, sr=_SAMPLE_RATE)
179
+ mel = logMel(y, sr)
180
+ cqt = logCQT(y, sr)
181
+ chroma = logChroma(y, sr)
182
  # Load the ground truth label
 
183
  n_steps = cqt.shape[1]
 
 
184
  IPT_label = np.zeros([n_IPTs, n_steps], dtype=int)
185
+ with open(csv_path, "r", encoding="utf-8") as f: # csv file for each audio
186
  reader = csv.DictReader(f, delimiter=",")
187
  for label in reader: # each note
188
  onset = float(label["onset_time"])
 
194
  IPT_label[IPT, left:frame_right] = 1
195
 
196
  return dict(
197
+ audio_path=audio_path,
198
+ csv_path=csv_path,
199
+ mel=mel,
200
+ cqt=cqt,
201
+ chroma=chroma,
202
+ IPT_label=IPT_label,
203
  )
204
 
205
  data = []
 
208
  for input_files in files(wav_dir, csv_dir, group):
209
  data.append(load_all(*input_files))
210
 
211
+ for i, dic in tqdm(enumerate(data), total=len(data), desc="Feature extracting"):
212
+ x_mel = chunk_data(dic["mel"])
213
+ x_cqt = chunk_data(dic["cqt"])
214
+ x_chroma = chunk_data(dic["chroma"])
215
  y_i = dic["IPT_label"]
216
  y_i = chunk_data(y_i)
217
  if i == 0:
218
+ Xtr_mel = x_mel
219
+ Xtr_cqt = x_cqt
220
+ Xtr_chroma = x_chroma
221
  Ytr_i = y_i
 
222
 
223
  else:
224
+ Xtr_mel = np.concatenate([Xtr_mel, x_mel], axis=0)
225
+ Xtr_cqt = np.concatenate([Xtr_cqt, x_cqt], axis=0)
226
+ Xtr_chroma = np.concatenate([Xtr_chroma, x_chroma], axis=0)
227
  Ytr_i = np.concatenate([Ytr_i, y_i], axis=0)
228
 
229
  # Transform the shape of the input
230
+ Xtr_mel = np.expand_dims(Xtr_mel, axis=3)
231
+ Xtr_cqt = np.expand_dims(Xtr_cqt, axis=3)
232
+ Xtr_chroma = np.expand_dims(Xtr_chroma, axis=3)
 
233
  # Normalize
234
+ Xtr_mel = self._norm(Xtr_mel)
235
+ Xtr_cqt = self._norm(Xtr_cqt)
236
+ Xtr_chroma = self._norm(Xtr_chroma)
237
+ return [list(Xtr_mel), list(Xtr_cqt), list(Xtr_chroma)], list(Ytr_i)
238
 
239
  def _parse_csv_label(self, csv_file):
240
  label = []
 
287
  testset.append(item)
288
 
289
  else:
290
+ audio_dir = os.path.join(audio_files, "audio")
291
+ csv_dir = os.path.join(csv_files, "label")
292
  X_train, Y_train = self._load(audio_dir, csv_dir, ["train"])
293
  X_valid, Y_valid = self._load(audio_dir, csv_dir, ["validation"])
294
  X_test, Y_test = self._load(audio_dir, csv_dir, ["test"])
295
+ for i in range(len(Y_train)):
296
+ trainset.append(
297
+ {
298
+ "mel": X_train[0][i],
299
+ "cqt": X_train[1][i],
300
+ "chroma": X_train[2][i],
301
+ "label": Y_train[i],
302
+ }
303
+ )
304
 
305
+ for i in range(len(Y_valid)):
306
+ validset.append(
307
+ {
308
+ "mel": X_valid[0][i],
309
+ "cqt": X_valid[1][i],
310
+ "chroma": X_valid[2][i],
311
+ "label": Y_valid[i],
312
+ }
313
+ )
314
 
315
+ for i in range(len(Y_test)):
316
+ testset.append(
317
+ {
318
+ "mel": X_test[0][i],
319
+ "cqt": X_test[1][i],
320
+ "chroma": X_test[2][i],
321
+ "label": Y_test[i],
322
+ }
323
+ )
324
 
325
  random.shuffle(trainset)
326
  random.shuffle(validset)