Skip to content

Commit 385ce7b

Browse files
PhonotonicAppPhonotonicApp
committed
first commit
0 parents  commit 385ce7b

File tree

10 files changed

+1183
-0
lines changed

10 files changed

+1183
-0
lines changed

‎.DS_Store‎

6 KB
Binary file not shown.

‎Kmeans.py‎

Lines changed: 100 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,100 @@
1+
#!/usr/bin/env python3
2+
# -*- coding: utf-8 -*-
3+
"""
4+
Created on Wed May 17 14:40:58 2017
5+
6+
@author: Phtc-LD
7+
"""
8+
import numpy as np
9+
from collections import Counter
10+
from acp import *
11+
from sklearn.cluster import KMeans
12+
from sklearn.metrics import confusion_matrix
13+
from sklearn import preprocessing
14+
from PFMath import *
15+
16+
def dist_nico(X, Y):
17+
18+
edist = 0.0
19+
output_dist = 0.0
20+
count = 0
21+
for i in range (1,len(X)):
22+
edist = math.pow(Y[i] - X[i], 2)
23+
if (edist <=40):
24+
count += 1
25+
output_dist += edist
26+
return output_dist
27+
28+
def MyKmeans(nb,matrice,initMatrix):
29+
if(initMatrix!=None):
30+
print("initialisation manuelle")
31+
est=KMeans(n_clusters=nb,init=initMatrix)
32+
else:
33+
print("initialisation automatique")
34+
est=KMeans(n_clusters=nb)
35+
36+
37+
est.fit(matrice)
38+
classe=est.labels_
39+
inertia=est.inertia_
40+
return classe,est.cluster_centers_,inertia
41+
42+
def CalculDistanceIntra(aa,cs,aaCs):
43+
print("Début : calcul distance intra")
44+
maDistanceI = np.zeros([len(aaCs),1])
45+
maPoids = Counter(cs)
46+
print("Répartition des clusters")
47+
print(maPoids)
48+
for n in range(0,len(aa)):
49+
maDistanceI[cs[n]] = maDistanceI[cs[n]] + DEucl(aa[n],aaCs[cs[n]])
50+
51+
for n in range(0,len(maDistanceI)):
52+
maDistanceI[n] = maDistanceI[n]/maPoids[n]
53+
54+
return maDistanceI
55+
56+
57+
58+
59+
def CalculMatriceSimilarite(maCs,aa,cs):
60+
maSimilarity = np.zeros([len(maCs),len(maCs)])
61+
maDistIntra = CalculDistanceIntra(aa,cs,maCs)
62+
63+
print("calcul matrice distance intra")
64+
print(maDistIntra)
65+
66+
for n in range(0,len(maCs)):
67+
for o in range(0,len(maCs)):
68+
#maSimilarity[n][o] = maDistIntra[n]/DEucl(maCs[n],maCs[o])
69+
#maSimilarity[n][o] = DEucl(maCs[n],maCs[o])/maDistIntra[n]
70+
maSimilarity[n][o] = DEucl(maCs[n],maCs[o])
71+
72+
print("medianne distance inter cluster")
73+
print(print(np.median(maSimilarity,axis=1)))
74+
75+
print("moyenne distance inter cluster")
76+
print(print(np.mean(maSimilarity,axis=1)))
77+
78+
[tab1,tab2,index90] = myPCA(pd.DataFrame(maCs),len(maCs[0]))
79+
80+
return maSimilarity,maDistIntra
81+
82+
def compareWithLabel(cs,lLabel):
83+
nbCls = len(Counter(cs))
84+
85+
mLabelCluster = [[]]
86+
87+
for i in range(0,nbCls):
88+
mLabelCluster.append([])
89+
90+
for i in range(0,len(cs)):
91+
mLabelCluster[cs[i]].append(lLabel[i])
92+
93+
for i in range(0,nbCls):
94+
print(Counter(mLabelCluster[i]))
95+
96+
return mLabelCluster
97+
98+
99+
100+

‎MM_analyze.py‎

Lines changed: 275 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,275 @@
1+
# MixMe: main script for structure analysis
2+
# 31-7-2017 jjburred for Phonotonic
3+
4+
import os
5+
import numpy as np
6+
import scipy.io.wavfile as wav
7+
import scipy as sp
8+
from scipy import signal
9+
import matplotlib.pyplot as plt
10+
from matplotlib import gridspec
11+
from MM_stft import stft
12+
from MM_features import *
13+
from MM_structure import *
14+
from MM_beat import *
15+
import time
16+
from shutil import copyfile
17+
from sklearn.preprocessing import StandardScaler
18+
19+
20+
# inputFolder = "/Users/jjb/Documents/research/phonotonic/db/test"
21+
inputFolder = "/Users/Phtc-LD/Desktop/Dev/Music/phonoFitTracks/GTTrack" #Apprentissage GTTrack
22+
#inputFolder = "/Users/Phtc-LD/Desktop/Dev/Music/phonoFitTracks/trackAtester/" #A tester
23+
24+
25+
inputSegmentFolder = "/Users/Phtc-LD/Desktop/Applications/MusicAndSport/mix.me_research/python/GT/Segmentation/"
26+
27+
28+
outputFolder = "/Users/Phtc-LD/Desktop/Applications/MusicAndSport/mix.me_research/python/out"
29+
30+
main_pars = {
31+
'resolution': 0.04, # analysis resolution in s
32+
'mfccWeight': 0.8, # MFCC weights for clustering (for segmentation, MFCC is always used)
33+
'chromaWeight': 0.2, # chroma weights for clustering
34+
'context': 16, # segmentation context in beats - combien il regarde autour
35+
'beatsPerSeg': 16, # minimum beats per segment
36+
'partsDetail': 0.6, # description level for parts clustering
37+
'debitsDetail': 0.7 # description level for debits clustering
38+
}
39+
40+
# =======================================================
41+
42+
# STFT parameters
43+
stft_pars = {
44+
'winLength': main_pars['resolution'], # in s. was 0.04
45+
'overlapFactor': 0.75
46+
}
47+
48+
# Beat detection parameters
49+
beat_pars = {
50+
'minBPM': 70,
51+
'maxBPM': 155
52+
}
53+
54+
# MFCC parameters
55+
mfcc_pars = {
56+
'numMFCC': 13,
57+
'numFilt': 40,
58+
'includeEnergy': 0,
59+
'maxFreq': 16000,
60+
'plot': 0
61+
}
62+
63+
# self-similarity matrix parameters
64+
struct_pars = {
65+
'dist': 'eucl', # distance: 'eucl', 'exp_cosine', 'cosine', 'corr'
66+
'cbSize': main_pars['context'], # checkerboard size
67+
'plot': 0,
68+
'partsDetail': main_pars['partsDetail'],
69+
'debitsDetail': main_pars['debitsDetail'],
70+
'beatsPerSeg': main_pars['beatsPerSeg'],
71+
'clustFeat': 'mean_std' # mean_std, stack, seq_dist
72+
}
73+
74+
# create output folder
75+
expName = time.strftime("MMout_%Y-%m-%d_%H-%M-%S")
76+
expFolder = os.path.join(outputFolder,expName)
77+
if not os.path.exists(expFolder):
78+
os.makedirs(expFolder)
79+
80+
# copy configuration file
81+
copyfile('./MM_analyze.py',os.path.join(expFolder,'MM_analyze.py'))
82+
83+
for f in os.listdir(inputFolder):
84+
print("F {}".format(os.path.splitext(f)[0]))
85+
name = os.path.splitext(f)[0]
86+
87+
str_currFileName = os.path.join(inputFolder,f)
88+
if str_currFileName.lower().endswith('.wav'):
89+
90+
#Get semgentation file for ground truth
91+
92+
93+
# create current file subfolder
94+
currFolder = os.path.join(expFolder,f)
95+
os.makedirs(currFolder)
96+
97+
# load wave file
98+
print("Loading wave file: "+str_currFileName)
99+
(fs,waveformOri) = wav.read(str_currFileName)
100+
oriLength = waveformOri.shape[0]
101+
102+
# mix to mono
103+
if len(waveformOri.shape)>1:
104+
waveform = np.sum(waveformOri,axis=1)
105+
else:
106+
waveform = waveformOri
107+
108+
# normalize to floats
109+
waveform = waveform/np.max(np.abs(waveform))
110+
111+
# compute STFT
112+
stft_pars['winSize_'] = int(2**np.ceil(np.log2(stft_pars['winLength']*fs))) # nextpow2
113+
stft_pars['hopSize_'] = int(np.round(stft_pars['winSize_']*(1-stft_pars['overlapFactor'])))
114+
stft_pars['realHopLength_'] = stft_pars['hopSize_']/fs
115+
116+
win = np.hamming(stft_pars['winSize_'])
117+
currSpec = stft(waveform,win,stft_pars['winSize_'],stft_pars['hopSize_'],stft_pars['overlapFactor'])
118+
119+
currSpec = np.squeeze(currSpec) # remove singleton dimension for mono STFTs
120+
currSpec = np.absolute(currSpec[:int(stft_pars['winSize_']/2+1),:])
121+
122+
print('spectrogram size: {} x {} (bins x frames)'.format(currSpec.shape[0],currSpec.shape[1]))
123+
124+
beatInd, beatConf = beatDetect(currSpec,stft_pars,beat_pars,currFolder)
125+
126+
# compute temporal vector
127+
stft_pars['numFrames_'] = currSpec.shape[1]
128+
tempVec = np.arange(0,stft_pars['numFrames_'])*stft_pars['realHopLength_']
129+
tempVec = tempVec[beatInd]
130+
#
131+
# # # high-resolution version
132+
# # tempVecHR = np.arange(0,stft_pars['numFrames_']*upSample)*stft_pars['realHopLength_']/upSample
133+
# # tempVecHR = tempVecHR[beatIndHR]
134+
#
135+
exportBeats(beatInd,beatConf,tempVec,currFolder)
136+
137+
# compute MFCC
138+
stft_pars['numBins_'] = currSpec.shape[0]
139+
melFB = init_mfcc(stft_pars,fs,mfcc_pars)
140+
currMfcc = mfcc(currSpec,melFB,mfcc_pars)
141+
142+
# compute chroma
143+
chromaInd = init_chroma(stft_pars,fs)
144+
currChroma = chroma(currSpec,chromaInd)
145+
146+
# compute energy
147+
currEnergy = energy(currSpec)
148+
149+
if main_pars['mfccWeight']==0:
150+
clustFeatMat = currChroma
151+
elif main_pars['chromaWeight']==0:
152+
clustFeatMat = currMfcc
153+
else:
154+
currChromaW = currChroma * main_pars['chromaWeight']
155+
currMfccW = currMfcc * main_pars['mfccWeight']
156+
clustFeatMat = np.concatenate((currMfccW,currChromaW),axis=0)
157+
158+
segFeatMat = currMfcc # feature matrix for segmentation
159+
160+
# normalize feature matrices
161+
scaler = StandardScaler()
162+
segFeatMat = scaler.fit_transform(np.transpose(segFeatMat))
163+
segFeatMat = np.transpose(segFeatMat)
164+
165+
scaler = StandardScaler()
166+
clustFeatMat = scaler.fit_transform(np.transpose(clustFeatMat))
167+
clustFeatMat = np.transpose(clustFeatMat)
168+
169+
# median filter
170+
# segFeatMat = sp.signal.medfilt(segFeatMat,5)
171+
# clustFeatMat = sp.signal.medfilt(clustFeatMat,5)
172+
173+
# quantize seg matrix to beats
174+
numBeats = len(beatInd)
175+
segFeatMatQ = np.zeros((segFeatMat.shape[0],numBeats-1))
176+
for i in range(0,numBeats-1):
177+
segFeatMatQ[:,i] = np.mean(segFeatMat[:,beatInd[i]:beatInd[i+1]],axis=1)
178+
179+
# quantize clust matrix to beats
180+
clustFeatMatQ = np.zeros((clustFeatMat.shape[0],numBeats-1))
181+
for i in range(0,numBeats-1):
182+
clustFeatMatQ[:,i] = np.mean(clustFeatMat[:,beatInd[i]:beatInd[i+1]],axis=1)
183+
184+
# quantize energy
185+
energyQ = np.zeros((numBeats-1,1))
186+
for i in range(0,numBeats-1):
187+
energyQ[i] = np.mean(currEnergy[beatInd[i]:beatInd[i+1]])
188+
189+
print('seg. feat. matr. size: {} x {} (dim x beats)'.format(segFeatMatQ.shape[0],segFeatMatQ.shape[1]))
190+
191+
# self-similarity matrix
192+
currSSM = SSM(segFeatMatQ,struct_pars,currFolder)
193+
194+
# segmentation based on SSM
195+
boundaries = SSM_segment(currSSM,struct_pars,currFolder,energyQ)
196+
197+
temp = tempVec[boundaries]
198+
199+
# export segments
200+
f = open(os.path.join(currFolder,'segments.txt'),'w')
201+
for i,b in enumerate(temp):
202+
f.write("{}\t{}\t{}\n".format(b,b,i))
203+
f.close()
204+
205+
# cluster segments
206+
#secIDs,subsecIDs,boundaries,clustVar,clusterKmeansPartie,clusterKmeansDebit,clustFeat = segCluster(clustFeatMatQ,boundaries,struct_pars,currFolder)
207+
208+
# sort debits by intensity
209+
#subsecIDs,globalIntensity = sortDebits(currSpec,beatInd,boundaries,secIDs,subsecIDs)
210+
#newSubSecIDs,globalIntensity = sortDebits(currSpec,beatInd,boundaries,clusterKmeansPartie,clusterKmeansDebit)
211+
212+
# export data
213+
#export(boundaries,secIDs,subsecIDs,tempVec,currFolder,waveformOri,fs,globalIntensity,clustVar)
214+
#export(boundaries,clusterKmeansPartie,newSubSecIDs,tempVec,currFolder,waveformOri,fs,globalIntensity,clustVar)
215+
216+
217+
218+
#TEST LAURENT
219+
220+
221+
str = "/Users/Phtc-LD/Desktop/Dev/Music/phonoFitTracks/" + name + ".png"
222+
str2 = "/Users/Phtc-LD/Desktop/Dev/Music/phonoFitTracks/" + name + "-R.png"
223+
#[clusterKmeansPartie,cCluster,inertia] = MyKmeans(5,clustFeat[1:19],None)
224+
225+
226+
fig = plt.figure()
227+
#plt.ylim([0,1])
228+
plt.stem(secIDs)
229+
#plt.savefig(str)
230+
plt.show()
231+
plt.close()
232+
233+
234+
fig = plt.figure()
235+
#plt.ylim([0,1])
236+
plt.stem(clusterKmeansPartie)
237+
#plt.savefig(str)
238+
plt.show()
239+
plt.close()
240+
241+
242+
#"""
243+
segmentPath = inputSegmentFolder + name + '.txt'
244+
maSegmentsGT = []
245+
with open(segmentPath) as fl:
246+
for line in fl:
247+
l = line.split("\t")
248+
maSegmentsGT.append(float(l[0]))
249+
250+
251+
252+
#"""
253+
254+
"""
255+
segmentPathATester = inputFolder + name + '.txt'
256+
maSegmentsGT = []
257+
with open(segmentPathATester) as fl:
258+
for line in fl:
259+
l = line.split("\t")
260+
maSegmentsGT.append(float(l[0]))
261+
"""
262+
263+
compteurGTSegment = 0
264+
for k in range(0,len(maSegmentsGT)):
265+
for m in range(0,len(temp)):
266+
if(temp[m]-0.1<maSegmentsGT[k]<temp[m]+0.1):
267+
compteurGTSegment = compteurGTSegment + 1
268+
break
269+
270+
file = open("/Users/Phtc-LD/Desktop/Dev/Music/phonoFitTracks/segmentsGT.txt","a")
271+
file.write("{}\t{}\n".format(name,100*compteurGTSegment/len(temp)))
272+
file.close()
273+
274+
275+

0 commit comments

Comments
 (0)