-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
310 lines (240 loc) · 12.1 KB
/
utils.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
######################################################
# UTILITY FUNCTIONS #
# #
# Some utility functions that were written while #
# working on deep chroma estimation #
# #
# Johannes Zeitler (johannes.zeitler@fau.de) #
# Dec. 2020 #
######################################################
import numpy as np
import matplotlib.pyplot as plt
import librosa
import pretty_midi
import os
import LibFMP.B
import IPython.display as ipd
# split long audio files into shorter segments of a defined duration
def splitLargeFiles(files, cqt_dir, chroma_dir, frameRate=50, maxDuration_seconds=100, suffix='_short', saveAlways=False):
maxFrames = int(maxDuration_seconds*frameRate)
for f in files:
cqt = np.load(os.path.join(cqt_dir, f))
chroma = np.load(os.path.join(chroma_dir, f))
fileLen = chroma.shape[1]
if fileLen < maxFrames:
if saveAlways:
fname, ftype = f.split('.')
np.save(os.path.join(cqt_dir, fname+suffix+str(0)+'.'+ftype), cqt)
np.save(os.path.join(chroma_dir, fname+suffix+str(0)+'.'+ftype), cqt)
continue
else:
splitInds=np.arange(maxFrames, fileLen, maxFrames)
cqtSplits=np.split(cqt, splitInds, axis=1)
chromaSplits = np.split(chroma, splitInds, axis=1)
fname, ftype = f.split('.')
for i in range(len(cqtSplits)):
np.save(os.path.join(cqt_dir, fname+suffix+str(i)+'.'+ftype), cqtSplits[i])
np.save(os.path.join(chroma_dir, fname+suffix+str(i)+'.'+ftype), chromaSplits[i])
# sonify pitch activations and audio signal. Adds possibility to specify start and end time
def sonifyStereo(pitch, x, hopSize, fs, min_pitch, limTime=None, harmonics_weights=[1, .5]):
if limTime is not None:
x = x[np.max((0,int(limTime[0]*fs))) : np.min((len(x), int(limTime[1]*fs)))]
pitch = pitch[:, np.max((0, int(limTime[0]*fs/hopSize))):np.min((pitch.shape[1], int(limTime[1]*fs/hopSize)))]
_, out = LibFMP.B.sonify_pitch_activations_with_signal(pitch, x, fs/hopSize, fs, min_pitch=min_pitch, Fc=440, harmonics_weights=harmonics_weights, fading_msec=5, stereo=True)
return out
# sonify pitch activations. Adds possibility to specify start and end time
def sonifyMONO(pitch, hopSize, fs, min_pitch=60, limTime=None, harmonics_weights=[1, .5], chroma=False):
if limTime is not None:
pitch = pitch[:, np.max((0, int(limTime[0]*fs/hopSize))):np.min((pitch.shape[1], int(limTime[1]*fs/hopSize)))]
if chroma:
out = LibFMP.B.sonify_chromagram(pitch, int(pitch.shape[1]*hopSize),
fs/hopSize, fs, fading_msec=5)
else:
out = LibFMP.B.sonify_pitch_activations(pitch, int(pitch.shape[1]*hopSize), fs/hopSize, fs, min_pitch=min_pitch, Fc=440, harmonics_weights=harmonics_weights, fading_msec=5)
return out
# check if audio should be shifted to match a cqt better
# output: number of frames that the pitch activations should be shifted to match the audio (negative = to the left)
def checkFrameShift(cqt, pitch, maxShift=20, numFrames=None):
shiftInds = np.arange(-maxShift, maxShift+1, 1)
startFrame = maxShift
if numFrames is None:
endFrame = cqt.shape[1] - maxShift - 1
else:
endFrame = np.min((startFrame+numFrames, cqt.shape[1]-maxShift-1))
cqtSnip = cqt[:, startFrame:endFrame]
minInd = np.argmin([np.sum(cqtSnip*(1-pitch[:,startFrame-i:endFrame-i])) for i in shiftInds])
return shiftInds[minInd]
# shift pitch and chroma activations by specified number of frames.
# negative shift = to the left
def shiftFrames(pitch, chroma, shift):
if shift == 0:
return pitch, chroma
pitch_ = np.zeros_like(pitch)
chroma_ = np.zeros_like(chroma)
if shift > 0:
pitch_[:,shift:] = pitch[:,:-shift]
chroma_[:,shift:] = chroma[:,:-shift]
else:
pitch_[:,:shift] = pitch[:,-shift:]
chroma_[:,:shift] = chroma[:,-shift:]
return pitch_, chroma_
# check if pitch and chroma activations should be shifted to be synchronized to audio data, then shift accordingly
def checkAndShift(cqt, pitch, chroma, maxShift=20, numFrames=None, isCenteredCQT=False):
if isCenteredCQT:
binsPerKey = int(cqt.shape[0]/pitch.shape[0])
startInd = int((binsPerKey-1)/2)
cqt = cqt[np.arange(startInd, cqt.shape[0], binsPerKey),:]
if len(cqt.shape)>2:
cqt = cqt[:,:,0]
shift = checkFrameShift(cqt, pitch, maxShift, numFrames)
pitch, chroma = shiftFrames(pitch, chroma, shift)
return pitch, chroma, shift
# transform a list of note events (from csv or similar) to pitch and chroma activation matrices. Captures every note event.
def score_to_matrices(score, timeSteps, bottomPitch=0, topPitch=127, field_t0=0, field_t1=1, field_pitch=2, controlChanges=None):
''' OLD VERSION, doesn't recognize fast staccato
pitchMatrix = np.zeros((topPitch-bottomPitch+1, len(timeSteps)))
chromaMatrix = np.zeros((12, len(timeSteps)))
for i in range(score.shape[0]):
t0 = score[i, field_t0]
t1 = score[i, field_t1]
p = score[i, field_pitch]
chromaMatrix[int(np.mod(p-60,12)),:] += ((t0 < timeSteps)*1) * ((t1>timeSteps)*1)
if p < bottomPitch or p > topPitch:
continue
pitchMatrix[int(p-bottomPitch),:] += ((t0 < timeSteps)*1) * ((t1>timeSteps)*1)
return (pitchMatrix > 0)*1.0, (chromaMatrix > 0)*1.0
'''
pitchMatrix = np.zeros((topPitch-bottomPitch+1, len(timeSteps)))
chromaMatrix = np.zeros((12, len(timeSteps)))
if isinstance(score, np.ndarray):
score = [score[i,:] for i in range(score.shape[0])]
for i in range(len(score)):
t0 = score[i][field_t0]
t1 = score[i][field_t1]
p = score[i][field_pitch]
if p < bottomPitch or p > topPitch:
continue
fAct = np.arange(len(timeSteps))[((t0 < timeSteps) * (t1>timeSteps)).astype(bool)]
if len(fAct) < 1:
startFrame = np.argmin(np.abs(timeSteps-t0))
stopFrame = np.argmin(np.abs(timeSteps-t1))
if np.abs(timeSteps[startFrame] - t0) < np.abs(timeSteps[stopFrame] - t1):
fAct = startFrame
else:
fAct = stopFrame
pitchMatrix[int(p-bottomPitch), fAct] = 1
# process sustain pedal like pretty midi does
if controlChanges is not None:
fs = 1 / np.mean(timeSteps[1:]-timeSteps[:-1])
pedal_threshold = 64
CC_SUSTAIN_PEDAL = 64
time_pedal_on = 0
is_pedal_on = False
for cc in [_e for _e in controlChanges
if _e.number == CC_SUSTAIN_PEDAL]:
time_now = int(cc.time*fs)
is_current_pedal_on = (cc.value >= pedal_threshold)
if not is_pedal_on and is_current_pedal_on:
time_pedal_on = time_now
is_pedal_on = True
elif is_pedal_on and not is_current_pedal_on:
# For each pitch, a sustain pedal "retains"
# the maximum velocity up to now due to
# logarithmic nature of human loudness perception
subpr = pitchMatrix[:, time_pedal_on:time_now]
# Take the running maximum
pedaled = np.maximum.accumulate(subpr, axis=1)
pitchMatrix[:, time_pedal_on:time_now] = pedaled
is_pedal_on = False
for p in range(bottomPitch, topPitch+1):
chromaMatrix[int(np.mod(p,12)),:] += pitchMatrix[p-bottomPitch,:]
return (pitchMatrix>0)*1.0, (chromaMatrix>0)*1.0
# transform a midi file to pitch and chroma activation matrices
def midi_to_matrices(midi, timeSteps, bottomPitch=0, topPitch=127):
''' OLD VERSION
midi_data = pretty_midi.pretty_midi.PrettyMIDI(midi)
pnoRoll = np.zeros((topPitch-bottomPitch+1, len(timeSteps)))
chroma = np.zeros((12, len(timeSteps)))
for instrument in midi_data.instruments:
roll = instrument.get_piano_roll(fs=fs/hopLength, times=timeSteps)
pnoRoll += (roll[bottomPitch:topPitch+1,:] > 0)*1.0
chroma += (instrument.get_chroma(fs=fs/hopLength, times=timeSteps) > 0)*1
return (pnoRoll>0)*1.0, (chroma>0)*1.0
'''
pitchMatrix = np.zeros((topPitch-bottomPitch+1, len(timeSteps)))
chromaMatrix = np.zeros((12, len(timeSteps)))
midi_data = pretty_midi.pretty_midi.PrettyMIDI(midi)
for instrument in midi_data.instruments:
if instrument.is_drum:
continue
notes = []
for n in instrument.notes:
notes.append([n.start, n.end, n.pitch])
CCs = instrument.control_changes
pitch, chroma = score_to_matrices(notes, timeSteps, bottomPitch, topPitch, controlChanges=CCs)
pitchMatrix += pitch
chromaMatrix += chroma
return (pitchMatrix>0)*1.0, (chromaMatrix>0)*1.0
# overlay cqt and pitch activation plot
def validate_alignments(cqt, pitch, hopLength, bottomNote='C1', fs=22050, binsPerKey=1, centerNoteToBins=True, compress=0, limTime=None, limPitch=None, ticksPerOctave=3, figsize=(15,10)):
if compress:
cqt = np.log10(1+compress*cqt)
time = librosa.times_like(cqt, sr=fs, hop_length=hopLength)
nBins = cqt.shape[0]
nKeys = nBins/binsPerKey
startBin = 0
if centerNoteToBins:
startBin += int((binsPerKey-1)/2)
ind_bins = np.arange(0, nKeys, ticksPerOctave)
ind_midi = librosa.midi_to_note(ind_bins+librosa.note_to_midi(bottomNote))
pitch_alpha = np.zeros((cqt.shape[0], pitch.shape[1], 4))
for i in range(pitch.shape[0]-1):
cqtLine = i*binsPerKey + startBin
pitch_alpha[cqtLine,:,0] = 1
pitch_alpha[cqtLine,:,3] = pitch[i,:]*0.5
plt.figure(figsize=figsize)
plt.imshow(cqt[::-1,:], extent=(time[0], time[-1],-startBin/binsPerKey-1/(2*binsPerKey),nKeys-1+startBin/binsPerKey+1/(2*binsPerKey)), aspect='auto', cmap='gray_r')
plt.imshow(pitch_alpha[::-1,:,:], extent=(time[0], time[-1], -startBin/binsPerKey-1/(2*binsPerKey),nKeys-1+startBin/binsPerKey+1/(2*binsPerKey)), aspect='auto')
plt.yticks(ind_bins, ind_midi)
if limTime is not None:
plt.xlim([limTime[0], limTime[1]])
if limPitch is not None:
plt.ylim([librosa.note_to_midi(limPitch[0])-librosa.note_to_midi(bottomNote), librosa.note_to_midi(limPitch[1])-librosa.note_to_midi(bottomNote)])
plt.grid()
plt.xlabel('Time [s]')
plt.show()
# investigate chromagram statistics for a specific dataset (i.e. percentage of active frames...)
def chromaStatistics(datasetPath, fileList=None, plot=False, frameRate=10):
eps = 1e-15
if fileList is None:
fileList = os.listdir(datasetPath)
numFrames = 0
noChromaFrame = 0
oneChromaFrame = 0
chromas = np.zeros(12)
for f in fileList:
ch = np.load(os.path.join(datasetPath, f))
numFrames += ch.shape[1]
noChromaFrame += np.sum(np.sum(ch, axis=0) < eps)
oneChromaFrame += np.sum(np.abs(np.sum(ch, axis=0) - 1) < eps)
chromas += np.sum(ch, axis=1)
duration = numFrames/frameRate
print('Statistics for ',datasetPath)
print('Number of frames: %i, approximately %.2f hours'%(numFrames, duration/3600))
print('Number of frames without active chroma: %i = %.2f%%'%(noChromaFrame, noChromaFrame/numFrames*100))
print('Number of frames with ONE active chroma: %i = %.2f%%'%(oneChromaFrame, oneChromaFrame/numFrames*100))
if plot:
plt.figure()
plt.stem(chromas/numFrames, use_line_collection=True)
xlabels = ['C', 'C#', 'D', 'D#', 'E', 'F', 'F#', 'G', 'G#', 'A', 'A#', 'B']
plt.xticks(np.arange(12), xlabels)
plt.xlabel('Chroma')
plt.ylabel('Frequency')
plt.title('Frequency of Chroma Values in '+datasetPath)
plt.show()
print('------------------------------------------\n')
return {'dataset': datasetPath,
'numFrames': numFrames,
'noChromaFrame': noChromaFrame,
'oneChromaFrame': oneChromaFrame,
'chromas':chromas}