|
mi@0
|
1 #!/usr/bin/env python
|
|
mi@0
|
2 # encoding: utf-8
|
|
mi@0
|
3 """
|
|
mi@0
|
4 feature_combine_seg.py
|
|
mi@0
|
5
|
|
mi@0
|
6 A script to evaluation script for the segmentation results using combinations of different features.
|
|
mi@0
|
7 """
|
|
mi@0
|
8
|
|
mi@0
|
9 import matplotlib
|
|
mi@0
|
10 # matplotlib.use('Agg')
|
|
mi@0
|
11 import sys, os, optparse, csv
|
|
mi@0
|
12 from itertools import combinations
|
|
mi@0
|
13 from os.path import join, isdir, isfile, abspath, dirname, basename, split, splitext
|
|
mi@0
|
14 from copy import copy
|
|
mi@0
|
15 from mvpa2.suite import *
|
|
mi@0
|
16
|
|
mi@0
|
17 import matplotlib.pyplot as plt
|
|
mi@0
|
18 import matplotlib.gridspec as gridspec
|
|
mi@0
|
19 import numpy as np
|
|
mi@0
|
20 from scipy.signal import correlate2d, convolve2d, filtfilt, resample
|
|
mi@0
|
21 from scipy.stats import mode
|
|
mi@0
|
22 from scipy.ndimage import zoom
|
|
mi@0
|
23 from scipy.ndimage.morphology import binary_fill_holes
|
|
mi@0
|
24 from scipy.ndimage.filters import *
|
|
mi@0
|
25 from scipy.spatial.distance import squareform, pdist
|
|
mi@0
|
26 from sklearn.decomposition import PCA
|
|
mi@0
|
27 from sklearn.mixture import GMM
|
|
mi@0
|
28 from sklearn.preprocessing import normalize
|
|
mi@0
|
29 from sklearn.metrics.pairwise import pairwise_distances
|
|
mi@0
|
30 from skimage.transform import hough_line, hough_line_peaks, probabilistic_hough_line
|
|
mi@0
|
31 from skimage.filter import canny, sobel
|
|
mi@0
|
32 from skimage import data, measure, segmentation, morphology
|
|
mi@0
|
33
|
|
mi@0
|
34 from PeakPickerUtil import PeakPicker
|
|
mi@0
|
35 from gmmdist import *
|
|
mi@0
|
36 from GmmMetrics import GmmDistance
|
|
mi@0
|
37 from RankClustering import rClustering
|
|
mi@0
|
38 from kmeans import Kmeans
|
|
mi@0
|
39
|
|
mi@0
|
40 def parse_args():
|
|
mi@0
|
41 # define parser
|
|
mi@0
|
42 op = optparse.OptionParser()
|
|
mi@0
|
43 # IO options
|
|
mi@0
|
44 op.add_option('-g', '--gammatonegram-features', action="store", dest="GF", default='/Volumes/c4dm-03/people/mit/features/gammatonegram/qupujicheng/2048', type="str", help="Loading features from.." )
|
|
mi@0
|
45 op.add_option('-s', '--spectrogram-features', action="store", dest="SF", default='/Volumes/c4dm-03/people/mit/features/spectrogram/qupujicheng/2048', type="str", help="Loading features from.." )
|
|
mi@0
|
46 op.add_option('-t', '--tempogram-features', action="store", dest="TF", default='/Volumes/c4dm-03/people/mit/features/tempogram/qupujicheng/tempo_features_6s', type="str", help="Loading features from.." )
|
|
mi@0
|
47 op.add_option('-a', '--annotations', action="store", dest="GT", default='/Volumes/c4dm-03/people/mit/annotation/qupujicheng/lowercase', type="str", help="Loading annotation files from.. ")
|
|
mi@0
|
48 op.add_option('-o', '--ouput', action="store", dest="OUTPUT", default='/Volumes/c4dm-03/people/mit/segmentation/gammatone/qupujicheng', type="str", help="Write segmentation results to ")
|
|
mi@0
|
49 op.add_option('-p', '--plot-novelty', action="store_true", dest="PLOT", default=False, help="Save novelty curev plot")
|
|
mi@0
|
50 op.add_option('-e', '--test-mode', action="store_true", dest="TEST", default=False, help="Test mode")
|
|
mi@0
|
51 op.add_option('-v', '--verbose-output', action="store_true", dest="VERBOSE", default=False, help="Exported raw detections.")
|
|
mi@0
|
52
|
|
mi@0
|
53 return op.parse_args()
|
|
mi@0
|
54 options, args = parse_args()
|
|
mi@0
|
55
|
|
mi@0
|
56 class FeatureObj() :
|
|
mi@0
|
57 __slots__ = ['key', 'audio', 'timestamps', 'gammatone_features', 'tempo_features', 'timbre_features', 'harmonic_features', 'gammatone_ssm', 'tempo_ssm', 'timbre_features', 'harmonic_ssm', 'ssm_timestamps']
|
|
mi@0
|
58
|
|
mi@0
|
59 class AudioObj():
|
|
mi@0
|
60 __slots__ = ['name', 'feature_list', 'gt', 'label', 'gammatone_features', 'tempo_features', 'timbre_features', 'harmonic_features', 'combined_features',\
|
|
mi@0
|
61 'gammatone_ssm', 'tempo_ssm', 'timbre_ssm', 'harmonic_ssm', 'combined_ssm', 'ssm', 'ssm_timestamps', 'tempo_timestamps']
|
|
mi@0
|
62
|
|
mi@0
|
63 class EvalObj():
|
|
mi@0
|
64 __slots__ = ['TP', 'FP', 'FN', 'P', 'R', 'F', 'AD', 'DA']
|
|
mi@0
|
65
|
|
mi@0
|
66 class SSMseg(object):
|
|
mi@0
|
67 '''The main segmentation object'''
|
|
mi@0
|
68 def __init__(self):
|
|
mi@0
|
69 self.SampleRate = 44100
|
|
mi@0
|
70 self.NqHz = self.SampleRate/2
|
|
mi@0
|
71 self.timestamp = []
|
|
mi@0
|
72 self.previousSample = 0.0
|
|
mi@0
|
73 self.featureWindow = 6.0
|
|
mi@0
|
74 self.featureStep = 3.0
|
|
mi@0
|
75 self.kernel_size = 80 # Adjust this param according to the feature resolution.
|
|
mi@0
|
76 self.blockSize = 4094
|
|
mi@0
|
77 self.stepSize = 2048
|
|
mi@0
|
78
|
|
mi@0
|
79 '''NOTE: Match the following params with those used for feature extraction!'''
|
|
mi@0
|
80
|
|
mi@0
|
81 '''NOTE: Unlike spectrogram ones, Gammatone features are extracted without an FFT or any overlap. The windowing is done under the purpose of chunking
|
|
mi@0
|
82 the audio to facilitate the gammatone filtering. Despite of the overlap in the time domain, only the first half after the filtering is returned,
|
|
mi@0
|
83 resulting in no overlapping effect in the extracted features. To obtain features for overlapped audio input, make the gammatoneLen equal to blockSize
|
|
mi@0
|
84 and return the whole filter output.'''
|
|
mi@0
|
85 self.gammatoneLen = 2048
|
|
mi@0
|
86 self.gammatoneBandGroups = [0, 16, 34, 50, 64]
|
|
mi@0
|
87 self.nGammatoneBands = 20
|
|
mi@0
|
88 self.histRes = 40
|
|
mi@0
|
89 self.lowFreq = 100
|
|
mi@0
|
90 self.highFreq = self.SampleRate / 4
|
|
mi@0
|
91
|
|
mi@0
|
92 '''Settings for extracting tempogram features.'''
|
|
mi@0
|
93 self.tempoWindow = 6.0
|
|
mi@0
|
94 self.bpmBands = [30, 45, 60, 80, 100, 120, 180, 240, 400, 600]
|
|
mi@0
|
95
|
|
mi@0
|
96 '''Peak picking settings'''
|
|
mi@0
|
97 self.threshold = 30
|
|
mi@0
|
98 self.delta_threshold = 0.5
|
|
mi@0
|
99 self.backtracking_threshold = 2.4
|
|
mi@0
|
100 self.polyfitting_on = True
|
|
mi@0
|
101 self.medfilter_on = True
|
|
mi@0
|
102 self.LPfilter_on = True
|
|
mi@0
|
103 self.whitening_on = False
|
|
mi@0
|
104 self.aCoeffs = [1.0000, -0.5949, 0.2348]
|
|
mi@0
|
105 self.bCoeffs = [0.1600, 0.3200, 0.1600]
|
|
mi@0
|
106 self.cutoff = 0.5
|
|
mi@0
|
107 self.medianWin = 5
|
|
mi@0
|
108
|
|
mi@0
|
109 def getGaussianParams(self, length, featureRate, timeWindow):
|
|
mi@0
|
110
|
|
mi@0
|
111 win_len = round(timeWindow * featureRate)
|
|
mi@0
|
112 win_len = win_len + (win_len % 2) - 1
|
|
mi@0
|
113
|
|
mi@0
|
114 # a 50% overlap between windows
|
|
mi@0
|
115 stepsize = ceil(win_len * 0.5)
|
|
mi@0
|
116 num_win = int(floor( (length) / stepsize))
|
|
mi@0
|
117 gaussian_rate = featureRate / stepsize
|
|
mi@0
|
118
|
|
mi@0
|
119 return stepsize, num_win, win_len, gaussian_rate
|
|
mi@0
|
120
|
|
mi@0
|
121 def GaussianDistance(self, feature, featureRate, timeWindow):
|
|
mi@0
|
122
|
|
mi@0
|
123 stepsize, num_win, win_len, gr = self.getGaussianParams(feature.shape[0], featureRate, timeWindow)
|
|
mi@0
|
124 print 'stepsize, num_win, feature', stepsize, num_win, feature.shape, featureRate, timeWindow
|
|
mi@0
|
125 gaussian_list = []
|
|
mi@0
|
126 gaussian_timestamps = []
|
|
mi@0
|
127 tsi = 0
|
|
mi@0
|
128
|
|
mi@0
|
129 # f = open('/Users/mitian/Documents/experiments/features.txt','w')
|
|
mi@0
|
130 # print 'divergence computing..'
|
|
mi@0
|
131 for num in xrange(num_win):
|
|
mi@0
|
132 # print num, num * stepsize , (num * stepsize) + win_len
|
|
mi@0
|
133 gf=GaussianFeature(feature[int(num * stepsize) : int((num * stepsize) + win_len), :],2)
|
|
mi@0
|
134 # f.write("\n%s" %str(gf))
|
|
mi@0
|
135 gaussian_list.append(gf)
|
|
mi@0
|
136 tsi = int(floor( num * stepsize + 1))
|
|
mi@0
|
137 gaussian_timestamps.append(self.timestamp[tsi])
|
|
mi@0
|
138
|
|
mi@0
|
139 # f.close()
|
|
mi@0
|
140
|
|
mi@0
|
141 # print 'gaussian_list', len(gaussian_list), len(gaussian_timestamps)
|
|
mi@0
|
142 dm = np.zeros((len(gaussian_list), len(gaussian_list)))
|
|
mi@0
|
143
|
|
mi@0
|
144 for v1, v2 in combinations(gaussian_list, 2):
|
|
mi@0
|
145 i, j = gaussian_list.index(v1), gaussian_list.index(v2)
|
|
mi@0
|
146 dm[i, j] = v1.distance(v2)
|
|
mi@0
|
147 dm[j, i] = v2.distance(v1)
|
|
mi@0
|
148 # print 'dm[i,j]',dm[i,j]
|
|
mi@0
|
149 # sio.savemat("/Users/mitian/Documents/experiments/dm-from-segmenter.mat",{"dm":dm})
|
|
mi@0
|
150 return dm, gaussian_timestamps
|
|
mi@0
|
151
|
|
mi@0
|
152 def gaussian_kernel(self, size):
|
|
mi@0
|
153 '''Create a gaussian tapered 45 degrees rotated checkerboard kernel.
|
|
mi@0
|
154 TODO: Unit testing: Should produce this with kernel size 3:
|
|
mi@0
|
155 0.1353 -0.3679 0.1353
|
|
mi@0
|
156 0.3679 1.0000 0.3679
|
|
mi@0
|
157 0.1353 -0.3679 0.1353
|
|
mi@0
|
158 '''
|
|
mi@0
|
159 n = float(np.ceil(size / 2.0))
|
|
mi@0
|
160 kernel = np.zeros((size,size))
|
|
mi@0
|
161 for i in xrange(1,size+1) :
|
|
mi@0
|
162 for j in xrange(1,size+1) :
|
|
mi@0
|
163 gauss = np.exp( -4.0 * (np.square( (i-n)/n ) + np.square( (j-n)/n )) )
|
|
mi@0
|
164 # gauss = 1
|
|
mi@0
|
165 if np.logical_xor( j - n > np.floor((i-n) / 2.0), j - n > np.floor((n-i) / 2.0) ) :
|
|
mi@0
|
166 kernel[i-1,j-1] = -gauss
|
|
mi@0
|
167 else:
|
|
mi@0
|
168 kernel[i-1,j-1] = gauss
|
|
mi@0
|
169 return kernel
|
|
mi@0
|
170
|
|
mi@0
|
171 def getDiagonalSlice(self, ssm, width):
|
|
mi@0
|
172 ''' Return a diagonal slice of the ssm given its width, with 45 degrees rotation.
|
|
mi@0
|
173 Note: requres 45 degrees rotated kernel also.'''
|
|
mi@0
|
174 w = int(np.floor(width/2.0))
|
|
mi@0
|
175 length = len(np.diagonal(ssm))
|
|
mi@0
|
176 slice = np.zeros((2*w+1,length))
|
|
mi@0
|
177 # print 'diagonal', length, w, slice.shape
|
|
mi@0
|
178 for i in xrange(-w, w+1) :
|
|
mi@0
|
179 slice[w+i,:] = np.hstack(( np.zeros(int(np.floor(abs(i)/2.0))), np.diagonal(ssm,i), np.zeros(int(np.ceil(abs(i)/2.0))) ))
|
|
mi@0
|
180 return slice
|
|
mi@0
|
181
|
|
mi@0
|
182 def getNoveltyCurve(self,dm, kernel_size):
|
|
mi@0
|
183 '''Return novelty score from distance matrix.'''
|
|
mi@0
|
184
|
|
mi@0
|
185 kernel_size = int(np.floor(kernel_size/2.0)+1)
|
|
mi@0
|
186 slice = self.getDiagonalSlice(dm, kernel_size)
|
|
mi@0
|
187 kernel = self.gaussian_kernel(kernel_size)
|
|
mi@0
|
188 xc = convolve2d(slice,kernel,mode='same')
|
|
mi@0
|
189 xc[abs(xc)>1e+10]=0.00001
|
|
mi@0
|
190 # print 'xc', xc.shape, xc
|
|
mi@0
|
191 return xc[int(np.floor(xc.shape[0]/2.0)),:]
|
|
mi@0
|
192
|
|
mi@0
|
193 def mergeBlocks(self, SSM, thresh=0.9, size=5):
|
|
mi@0
|
194 '''Merge consequtive small blocks along the diagonal.'''
|
|
mi@0
|
195 # found = False
|
|
mi@0
|
196 # start = 0
|
|
mi@0
|
197 # i = 0
|
|
mi@0
|
198 # while i < len(SSM):
|
|
mi@0
|
199 # j = i + 1
|
|
mi@0
|
200 # if found: start = i
|
|
mi@0
|
201 # while(j < len(SSM) and SSM[i, j]):
|
|
mi@0
|
202 # if (j-i) > size:
|
|
mi@0
|
203 # found = True
|
|
mi@0
|
204 # i = j
|
|
mi@0
|
205 # # print 'start,end', start, i
|
|
mi@0
|
206 # start = i
|
|
mi@0
|
207 # else:
|
|
mi@0
|
208 # found = False
|
|
mi@0
|
209 # j += 1
|
|
mi@0
|
210 # if not found:
|
|
mi@0
|
211 # print 'start,end', start, i
|
|
mi@0
|
212 # SSM[start:i, start:i] = 0.9
|
|
mi@0
|
213 # i = j
|
|
mi@0
|
214 idx = 1
|
|
mi@0
|
215 while idx < len(SSM):
|
|
mi@0
|
216 i = 0
|
|
mi@0
|
217 # if ((idx-1-i) > 0 and (idx+1+i) < len(SSM)):
|
|
mi@0
|
218 while ((idx-1-i) > 0 and (idx+1+i) < len(SSM) and SSM[idx-1-i, idx] > 0 and SSM[idx+1+i, idx] > 0):
|
|
mi@0
|
219 i += 1
|
|
mi@0
|
220 if i > size/2:
|
|
mi@0
|
221 SSM[idx-1-i:min(idx+i,len(SSM)), idx-1-i:min(idx+i,len(SSM))] = 1.0
|
|
mi@0
|
222 idx += max(1, i)
|
|
mi@0
|
223 return SSM
|
|
mi@0
|
224
|
|
mi@0
|
225 def getGMMs(self, feature, segment_boundaries):
|
|
mi@0
|
226 '''Return GMMs for located segments'''
|
|
mi@0
|
227 gmm_list = []
|
|
mi@0
|
228 gmm_list.append(GmmDistance(feature[: segment_boundaries[0], :], components = 1))
|
|
mi@0
|
229 for i in xrange(1, len(segment_boundaries)):
|
|
mi@0
|
230 gmm_list.append(GmmDistance(feature[segment_boundaries[i-1] : segment_boundaries[i], :], components = 1))
|
|
mi@0
|
231 return gmm_list
|
|
mi@0
|
232
|
|
mi@0
|
233 def trackValley(self, onset_index, smoothed_df):
|
|
mi@0
|
234 '''Back track to the valley location of detected peaks'''
|
|
mi@0
|
235 prevDiff = oldDiff = 0.0
|
|
mi@0
|
236 while (onset_index > 1) :
|
|
mi@0
|
237 diff = smoothed_df[onset_index] - smoothed_df[onset_index-1]
|
|
mi@0
|
238 # if (diff < 0.0 and 0 <= prevDiff < oldDiff * self.backtracking_threshold) : break
|
|
mi@0
|
239 if (diff < 0 and prevDiff >= 0.1 * smoothed_df[onset_index]) : break
|
|
mi@0
|
240 prevDiff = diff
|
|
mi@0
|
241 oldDiff = prevDiff
|
|
mi@0
|
242 onset_index -= 1
|
|
mi@0
|
243 return onset_index
|
|
mi@0
|
244
|
|
mi@0
|
245 def normaliseFeature(self, feature_array):
|
|
mi@0
|
246
|
|
mi@0
|
247 feature_array = np.array(feature_array)
|
|
mi@0
|
248 feature_array[np.isnan(feature_array)] = 0.0
|
|
mi@0
|
249 feature_array[np.isinf(feature_array)] = 0.0
|
|
mi@0
|
250
|
|
mi@0
|
251 if len(feature_array.shape) == 1:
|
|
mi@0
|
252 feature_array = (feature_array - feature_array.min()) / (feature_array.max() - feature_array.min())
|
|
mi@0
|
253 else:
|
|
mi@0
|
254 mins = feature_array.min(axis=1)
|
|
mi@0
|
255 maxs = feature_array.max(axis=1)
|
|
mi@0
|
256 feature_array = (feature_array - mins[:, np.newaxis]) / (maxs - mins)[:, np.newaxis]
|
|
mi@0
|
257 feature_array[np.isnan(feature_array)] = 0.0
|
|
mi@0
|
258 return feature_array
|
|
mi@0
|
259
|
|
mi@0
|
260 def upSample(self, feature_array, step):
|
|
mi@0
|
261 '''Resample downsized tempogram features, tempoWindo should be in accordance with input features'''
|
|
mi@0
|
262 # print feature_array.shape
|
|
mi@0
|
263 sampleRate = 44100
|
|
mi@0
|
264 stepSize = 1024.0
|
|
mi@0
|
265 # step = np.ceil(sampleRate/stepSize/5.0)
|
|
mi@0
|
266 feature_array = zoom(feature_array, (step,1))
|
|
mi@0
|
267 # print 'resampled', feature_array.shape
|
|
mi@0
|
268 return feature_array
|
|
mi@0
|
269
|
|
mi@0
|
270 def stripeDistance(self, feature_array, feature_len, step, metric='cosine'):
|
|
mi@0
|
271 '''Return distance matrix calculated for 2d time invariant features.'''
|
|
mi@0
|
272 size = feature_array.shape[0] / feature_len
|
|
mi@0
|
273 dm = np.zeros((size, size))
|
|
mi@0
|
274
|
|
mi@0
|
275 for i in xrange(size):
|
|
mi@0
|
276 for j in xrange(i, size):
|
|
mi@0
|
277 dm[i, j] = np.sum(pairwise_distances(feature_array[i*step:(i+1)*step, :], feature_array[j*step:(j+1)*step, :], metric))
|
|
mi@0
|
278 dm[j, i] = dm[i, j]
|
|
mi@0
|
279 # print 'np.nanmax(dm)', np.nanmax(dm)
|
|
mi@0
|
280 dm[np.isnan(dm)] = np.nanmax(dm)
|
|
mi@0
|
281 ssm = 1 - (dm - dm.min()) / (dm.max() - dm.min())
|
|
mi@0
|
282 np.fill_diagonal(ssm, 1)
|
|
mi@0
|
283 return ssm
|
|
mi@0
|
284
|
|
mi@0
|
285
|
|
mi@0
|
286 def getMean(self, feature, winlen, stepsize):
|
|
mi@0
|
287 means = []
|
|
mi@0
|
288 steps = int((feature.shape[0] - winlen + stepsize) / stepsize)
|
|
mi@0
|
289 for i in xrange(steps):
|
|
mi@0
|
290 means.append(np.mean(feature[i*stepsize:(i*stepsize+winlen), :], axis=0))
|
|
mi@0
|
291 return np.array(means)
|
|
mi@0
|
292
|
|
mi@0
|
293 def getStd(self, feature, winlen, stepsize):
|
|
mi@0
|
294 std = []
|
|
mi@0
|
295 steps = int((feature.shape[0] - winlen + stepsize) / stepsize)
|
|
mi@0
|
296 for i in xrange(steps):
|
|
mi@0
|
297 std.append(np.std(feature[i*stepsize:(i*stepsize+winlen), :], axis=0))
|
|
mi@0
|
298 return np.array(std)
|
|
mi@0
|
299
|
|
mi@0
|
300 def getDelta(self, feature):
|
|
mi@0
|
301 delta_feature = np.vstack((np.zeros((1, feature.shape[1])), np.diff(feature, axis=0)))
|
|
mi@0
|
302 return delta_feature
|
|
mi@0
|
303
|
|
mi@0
|
304 def backtrack(self, onset_index, smoothed_df):
|
|
mi@0
|
305 '''Backtrack the onsets to an earlier 'perceived' location from the actually detected peak...
|
|
mi@0
|
306 This is based on the rationale that the perceived onset tends to be a few frames before the detected peak.
|
|
mi@0
|
307 This tracks the position in the detection function back to where the peak is startng to build up.
|
|
mi@0
|
308 Notice the "out of the blue" parameter: 0.9. (Ideally, this should be tested, evaluated and reported...)'''
|
|
mi@0
|
309 prevDiff = 0.0
|
|
mi@0
|
310 while (onset_index > 1) :
|
|
mi@0
|
311 diff = smoothed_df[onset_index] - smoothed_df[onset_index-1]
|
|
mi@0
|
312 if diff < prevDiff * self.backtracking_threshold : break
|
|
mi@0
|
313 prevDiff = diff
|
|
mi@0
|
314 onset_index -= 1
|
|
mi@0
|
315 return onset_index
|
|
mi@0
|
316
|
|
mi@0
|
317 def trackDF(self, onset1_index, df2):
|
|
mi@0
|
318 '''In the second round of detection, remove the known onsets from the DF by tracking from the peak given by the first round
|
|
mi@0
|
319 to a valley to deminish the recognised peaks on top of which to start new detection.'''
|
|
mi@0
|
320 for idx in xrange(len(onset1_index)) :
|
|
mi@0
|
321 remove = True
|
|
mi@0
|
322 for i in xrange(onset1_index[idx], 1, -1) :
|
|
mi@0
|
323 if remove :
|
|
mi@0
|
324 if df2[i] >= df2[i-1] :
|
|
mi@0
|
325 df2[i] == 0.0
|
|
mi@0
|
326 else:
|
|
mi@0
|
327 remove = False
|
|
mi@0
|
328 return df2
|
|
mi@0
|
329
|
|
mi@0
|
330 def getSSM(self, feature_array, metric='cosine', norm='simple'):
|
|
mi@0
|
331 '''Compute SSM given input feature array.
|
|
mi@0
|
332 args: norm: ['simple', 'remove_noise']
|
|
mi@0
|
333 '''
|
|
mi@0
|
334 dm = pairwise_distances(feature_array, metric=metric)
|
|
mi@0
|
335 dm = np.nan_to_num(dm)
|
|
mi@0
|
336 if norm == 'simple':
|
|
mi@0
|
337 ssm = 1 - (dm - np.min(dm)) / (np.max(dm) - np.min(dm))
|
|
mi@0
|
338 return ssm
|
|
mi@0
|
339
|
|
mi@0
|
340 def reduceSSM(self, ssm, maxfilter_size = 2, remove_size=50):
|
|
mi@0
|
341 ssm[ssm<0.8] = 0
|
|
mi@0
|
342 ssm = maximum_filter(ssm,size=maxfilter_size)
|
|
mi@0
|
343 ssm = morphology.remove_small_objects(ssm.astype(bool), min_size=remove_size)
|
|
mi@0
|
344 return ssm
|
|
mi@0
|
345
|
|
mi@0
|
346 def getPeakFeatures(self, peak_candidates, featureset, winlen):
|
|
mi@0
|
347 '''
|
|
mi@0
|
348 args: winlen: length of feature window before and after an investigated peak
|
|
mi@0
|
349 featureset: A list of audio features for measuring the dissimilarity.
|
|
mi@0
|
350
|
|
mi@0
|
351 return: peak_features
|
|
mi@0
|
352 A list of tuples of features for windows before and after each peak.
|
|
mi@0
|
353 '''
|
|
mi@0
|
354 prev_features = []
|
|
mi@0
|
355 post_features = []
|
|
mi@0
|
356 feature_types = len(featureset)
|
|
mi@0
|
357 # print peak_candidates[-1], winlen, featureset[0].shape
|
|
mi@0
|
358 # if peak_candidates[-1] + winlen > featureset[0].shape[0]:
|
|
mi@0
|
359 # peak_candidates = peak_candidates[:-1]
|
|
mi@0
|
360 # for x in peak_candidates:
|
|
mi@0
|
361 # prev_features.append(tuple([featureset[i][x-winlen:x, :] for i in xrange(feature_types)]))
|
|
mi@0
|
362 # post_features.append(tuple([featureset[i][x:x+winlen, :] for i in xrange(feature_types)]))
|
|
mi@0
|
363 prev_features.append(tuple([featureset[i][:peak_candidates[0], :] for i in xrange(feature_types)]))
|
|
mi@0
|
364 post_features.append(tuple([featureset[i][peak_candidates[0]:peak_candidates[1], :] for i in xrange(feature_types)]))
|
|
mi@0
|
365 for idx in xrange(1, len(peak_candidates)-1):
|
|
mi@0
|
366 prev_features.append(tuple([featureset[i][peak_candidates[idx-1]:peak_candidates[idx], :] for i in xrange(feature_types)]))
|
|
mi@0
|
367 post_features.append(tuple([featureset[i][peak_candidates[idx]:peak_candidates[idx+1], :] for i in xrange(feature_types)]))
|
|
mi@0
|
368 prev_features.append(tuple([featureset[i][peak_candidates[-2]:peak_candidates[-1], :] for i in xrange(feature_types)]))
|
|
mi@0
|
369 post_features.append(tuple([featureset[i][peak_candidates[-1]:, :] for i in xrange(feature_types)]))
|
|
mi@0
|
370 return prev_features, post_features
|
|
mi@0
|
371
|
|
mi@0
|
372 def segmentDev(self, prev_features, post_features):
|
|
mi@0
|
373 '''Deviations are measured for each given feature type.
|
|
mi@0
|
374 peak_candidates: peaks from the 1st round detection
|
|
mi@0
|
375 peak_features: Features for measuring the dissimilarity for parts before and after each peak.
|
|
mi@0
|
376 dtype: tuple.
|
|
mi@0
|
377 '''
|
|
mi@0
|
378 dev_list = []
|
|
mi@0
|
379 n_peaks = len(prev_features)
|
|
mi@0
|
380 n_features = len(prev_features[0])
|
|
mi@0
|
381 # print 'n_peaks, n_features', n_peaks, n_features
|
|
mi@0
|
382 for x in xrange(n_peaks):
|
|
mi@0
|
383 f1, f2 = prev_features[x], post_features[x]
|
|
mi@0
|
384 dev_list.append(tuple([GmmDistance(f1[i], components=1).skl_distance_full(GmmDistance(f2[i], components=1)) for i in xrange(n_features)]))
|
|
mi@0
|
385 return dev_list
|
|
mi@0
|
386
|
|
mi@0
|
387 def verifyPeaks(self, peak_canditates, dev_list):
|
|
mi@0
|
388 '''Verify peaks from the 1st round detection by applying adaptive thresholding to the deviation list.'''
|
|
mi@0
|
389
|
|
mi@0
|
390 final_peaks = copy(peak_canditates)
|
|
mi@0
|
391 dev_list = np.array([np.mean(x) for x in dev_list]) # get average of devs of different features
|
|
mi@0
|
392 med_dev = median_filter(dev_list, size=5)
|
|
mi@0
|
393 # print dev_list, np.min(dev_list), np.median(dev_list), np.mean(dev_list), np.std(dev_list)
|
|
mi@0
|
394 dev = dev_list - med_dev
|
|
mi@0
|
395 # print dev
|
|
mi@0
|
396 for i, x in enumerate(dev):
|
|
mi@0
|
397 if x < 0:
|
|
mi@0
|
398 final_peaks.remove(peak_canditates[i])
|
|
mi@0
|
399 return final_peaks
|
|
mi@0
|
400
|
|
mi@0
|
401 def pairwiseF(self, annotation, detection, tolerance=3.0, combine=1.0):
|
|
mi@0
|
402 '''Pairwise F measure evaluation of detection rates.'''
|
|
mi@0
|
403
|
|
mi@0
|
404 # print 'detection', detection
|
|
mi@0
|
405 res = EvalObj()
|
|
mi@0
|
406 res.TP = 0 # Total number of matched ground truth and experimental data points
|
|
mi@0
|
407 gt = len(annotation) # Total number of ground truth data points
|
|
mi@0
|
408 dt = len(detection) # Total number of experimental data points
|
|
mi@0
|
409 foundIdx = []
|
|
mi@0
|
410 D_AD = np.zeros(gt)
|
|
mi@0
|
411 D_DA = np.zeros(dt)
|
|
mi@0
|
412
|
|
mi@0
|
413 for dtIdx in xrange(dt):
|
|
mi@0
|
414 # print detection[dtIdx], abs(detection[dtIdx] - annotation)
|
|
mi@0
|
415 D_DA[dtIdx] = np.min(abs(detection[dtIdx] - annotation))
|
|
mi@0
|
416 # D_DA[dtIdx] = min([abs(annot - detection[dtIdx]) for annot in annotation])
|
|
mi@0
|
417 for gtIdx in xrange(gt):
|
|
mi@0
|
418 D_AD[gtIdx] = np.min(abs(annotation[gtIdx] - detection))
|
|
mi@0
|
419 # D_AD[gtIdx] = min([abs(det - annotation[gtIdx]) for det in detection])
|
|
mi@0
|
420 for dtIdx in xrange(dt):
|
|
mi@0
|
421 if (annotation[gtIdx] >= detection[dtIdx] - tolerance/2.0) and (annotation[gtIdx] <= detection[dtIdx] + tolerance/2.0):
|
|
mi@0
|
422 res.TP = res.TP + 1.0
|
|
mi@0
|
423 foundIdx.append(gtIdx)
|
|
mi@0
|
424 foundIdx = list(set(foundIdx))
|
|
mi@0
|
425 res.TP = len(foundIdx)
|
|
mi@0
|
426 res.FP = dt - res.TP
|
|
mi@0
|
427 res.FN = gt - res.TP
|
|
mi@0
|
428
|
|
mi@0
|
429 res.AD = np.mean(D_AD)
|
|
mi@0
|
430 res.DA = np.mean(D_DA)
|
|
mi@0
|
431
|
|
mi@0
|
432 res.P, res.R, res.F = 0.0, 0.0, 0.0
|
|
mi@0
|
433
|
|
mi@0
|
434 if res.TP == 0:
|
|
mi@0
|
435 return res
|
|
mi@0
|
436
|
|
mi@0
|
437 res.P = res.TP / float(dt)
|
|
mi@0
|
438 res.R = res.TP / float(gt)
|
|
mi@0
|
439 # res.F = 2 * res.P * res.R / (res.P + res.F)
|
|
mi@0
|
440 res.F = 2.0 / (1.0/res.P + 1.0/res.R)
|
|
mi@0
|
441 # return TP3, FP3, FN3, pairwisePrecision3, pairwiseRecall3, pairwiseFValue3, TP05, FP05, FN05, pairwisePrecision05, pairwiseRecall05, pairwiseFValue05
|
|
mi@0
|
442 return res
|
|
mi@0
|
443
|
|
mi@0
|
444 def plotDetection(self, ssm, novelty, smoothed_novelty, gt, det, filename):
|
|
mi@0
|
445 '''Plot performance curve.
|
|
mi@0
|
446 x axis: distance threshold for feature selection; y axis: f measure'''
|
|
mi@0
|
447
|
|
mi@0
|
448 plt.figure(figsize=(10,16))
|
|
mi@0
|
449 gt_plot = gt / gt[-1] * len(novelty)
|
|
mi@0
|
450 det_plot = det / gt[-1] * len(novelty)
|
|
mi@0
|
451
|
|
mi@0
|
452 gs = gridspec.GridSpec(2, 1, height_ratios=[3,1])
|
|
mi@0
|
453 ax0 = plt.subplot(gs[0])
|
|
mi@0
|
454 ax1 = plt.subplot(gs[1], sharex=ax0)
|
|
mi@0
|
455
|
|
mi@0
|
456 ax0.imshow(ssm)
|
|
mi@0
|
457 ax0.vlines(gt_plot, 0, len(ssm), colors ='w', linestyles='solid')
|
|
mi@0
|
458 ax0.vlines(det_plot, 0, len(ssm), colors='k', linestyles='dashed')
|
|
mi@0
|
459 ax1.plot(np.linspace(0, len(novelty)-1, len(novelty)), novelty, 'g', np.linspace(0, len(novelty)-1, len(novelty)), smoothed_novelty,'b')
|
|
mi@0
|
460 y_min, y_max = min([min(novelty), min(smoothed_novelty)]), max([max(novelty), max(smoothed_novelty)])
|
|
mi@0
|
461 ax1.vlines(gt_plot, y_min, y_max, colors ='r', linestyles='solid')
|
|
mi@0
|
462 ax1.vlines(det_plot, y_min, y_max, colors='k', linestyles='dashed')
|
|
mi@0
|
463
|
|
mi@0
|
464 # f, ax = plt.subplots(2, sharex=True)
|
|
mi@0
|
465 # ax[0].imshow(ssm)
|
|
mi@0
|
466 # ax[1].plot(np.linspace(0, len(novelty)-1, len(novelty)), novelty)
|
|
mi@0
|
467 # ax[1].vlines(gt_plot, 0, len(novelty), colors ='r', linestyles='solid')
|
|
mi@0
|
468 # ax[1].vlines(det_plot, 0, len(novelty), colors='b', linestyles='dashed')
|
|
mi@0
|
469 #
|
|
mi@0
|
470 # plt.show()
|
|
mi@0
|
471 plt.savefig(filename)
|
|
mi@0
|
472
|
|
mi@0
|
473 return None
|
|
mi@0
|
474
|
|
mi@0
|
475 def process(self):
|
|
mi@0
|
476 '''For the aggregated input features, discard a propertion each time as the pairwise distances within the feature space descending.
|
|
mi@0
|
477 In the meanwhile evaluate the segmentation result and track the trend of perfomance changing by measuring the feature selection
|
|
mi@0
|
478 threshold - segmentation f measure curve.
|
|
mi@0
|
479 '''
|
|
mi@0
|
480 ssom = SimpleSOMMapper((30,30), 800, learning_rate=0.001)
|
|
mi@0
|
481
|
|
mi@0
|
482 peak_picker = PeakPicker()
|
|
mi@0
|
483 peak_picker.params.alpha = 9.0 # Alpha norm
|
|
mi@0
|
484 peak_picker.params.delta = self.delta_threshold # Adaptive thresholding delta
|
|
mi@0
|
485 peak_picker.params.QuadThresh_a = (100 - self.threshold) / 1000.0
|
|
mi@0
|
486 peak_picker.params.QuadThresh_b = 0.0
|
|
mi@0
|
487 peak_picker.params.QuadThresh_c = (100 - self.threshold) / 1500.0
|
|
mi@0
|
488 peak_picker.params.rawSensitivity = 20
|
|
mi@0
|
489 peak_picker.params.aCoeffs = self.aCoeffs
|
|
mi@0
|
490 peak_picker.params.bCoeffs = self.bCoeffs
|
|
mi@0
|
491 peak_picker.params.preWin = self.medianWin
|
|
mi@0
|
492 peak_picker.params.postWin = self.medianWin + 1
|
|
mi@0
|
493 peak_picker.params.LP_on = self.LPfilter_on
|
|
mi@0
|
494 peak_picker.params.Medfilt_on = self.medfilter_on
|
|
mi@0
|
495 peak_picker.params.Polyfit_on = self.polyfitting_on
|
|
mi@0
|
496 peak_picker.params.isMedianPositive = False
|
|
mi@0
|
497
|
|
mi@0
|
498 # Settings used for feature extraction
|
|
mi@0
|
499 feature_window_frame = int(self.SampleRate / self.gammatoneLen * self.featureWindow)
|
|
mi@0
|
500 feature_step_frame = int(0.5 * self.SampleRate / self.gammatoneLen * self.featureStep)
|
|
mi@0
|
501 aggregation_window, aggregation_step = 100, 50
|
|
mi@0
|
502 featureRate = float(self.SampleRate) / self.stepSize
|
|
mi@0
|
503
|
|
mi@0
|
504 audio_files = [x for x in os.listdir(options.GT) if not x.startswith(".") ]
|
|
mi@0
|
505 # audio_files = audio_files[:2]
|
|
mi@0
|
506 audio_files.sort()
|
|
mi@0
|
507 audio_list = []
|
|
mi@0
|
508
|
|
mi@0
|
509 gammatone_feature_list = [i for i in os.listdir(options.GF) if not i.startswith('.')]
|
|
mi@0
|
510 gammatone_feature_list = ['rolloff', 'contrast']
|
|
mi@0
|
511 tempo_feature_list = [i for i in os.listdir(options.TF) if not i.startswith('.')]
|
|
mi@0
|
512 # tempo_feature_list = ['intensity_bpm_renamed', 'loudness_bpm_renamed']
|
|
mi@0
|
513 timbre_feature_list = ['mfcc']
|
|
mi@0
|
514 harmonic_feature_list = ['nnls']
|
|
mi@0
|
515
|
|
mi@0
|
516 gammatone_feature_list = [join(options.GF, f) for f in gammatone_feature_list]
|
|
mi@0
|
517 timbre_feature_list = [join(options.SF, f) for f in timbre_feature_list]
|
|
mi@0
|
518 tempo_feature_list = [join(options.TF, f) for f in tempo_feature_list]
|
|
mi@0
|
519 harmonic_feature_list = [join(options.SF, f) for f in harmonic_feature_list]
|
|
mi@0
|
520
|
|
mi@0
|
521 fobj_list = []
|
|
mi@0
|
522
|
|
mi@0
|
523 # For each audio file, load specific features
|
|
mi@0
|
524 for audio in audio_files:
|
|
mi@0
|
525 ao = AudioObj()
|
|
mi@0
|
526 ao.name = splitext(audio)[0]
|
|
mi@0
|
527 # print 'audio:', ao.name
|
|
mi@0
|
528 # annotation_file = join(options.GT, ao.name+'.txt') # iso, salami
|
|
mi@0
|
529 # ao.gt = np.genfromtxt(annotation_file, usecols=0)
|
|
mi@0
|
530 # ao.label = np.genfromtxt(annotation_file, usecols=1, dtype=str)
|
|
mi@0
|
531 annotation_file = join(options.GT, ao.name+'.csv') # qupujicheng
|
|
mi@0
|
532 ao.gt = np.genfromtxt(annotation_file, usecols=0, delimiter=',')
|
|
mi@0
|
533 ao.label = np.genfromtxt(annotation_file, usecols=1, delimiter=',', dtype=str)
|
|
mi@0
|
534
|
|
mi@0
|
535 gammatone_featureset, timbre_featureset, tempo_featureset, harmonic_featureset = [], [], [], []
|
|
mi@0
|
536 for feature in gammatone_feature_list:
|
|
mi@0
|
537 for f in os.listdir(feature):
|
|
mi@0
|
538 if f[:f.find('_vamp')]==ao.name:
|
|
mi@0
|
539 gammatone_featureset.append(np.genfromtxt(join(feature, f), delimiter=',',filling_values=0.0)[:,1:])
|
|
mi@0
|
540 break
|
|
mi@0
|
541 if len(gammatone_feature_list) > 1:
|
|
mi@0
|
542 n_frame = np.min([x.shape[0] for x in gammatone_featureset])
|
|
mi@0
|
543 gammatone_featureset = [x[:n_frame,:] for x in gammatone_featureset]
|
|
mi@0
|
544 ao.gammatone_features = np.hstack((gammatone_featureset))
|
|
mi@0
|
545 else:
|
|
mi@0
|
546 ao.gammatone_features = gammatone_featureset[0]
|
|
mi@0
|
547
|
|
mi@0
|
548 for feature in timbre_feature_list:
|
|
mi@0
|
549 for f in os.listdir(feature):
|
|
mi@0
|
550 if f[:f.find('_vamp')]==ao.name:
|
|
mi@0
|
551 timbre_featureset.append(np.genfromtxt(join(feature, f), delimiter=',',filling_values=0.0)[:,1:])
|
|
mi@0
|
552 break
|
|
mi@0
|
553 if len(timbre_feature_list) > 1:
|
|
mi@0
|
554 n_frame = np.min([x.shape[0] for x in timbre_featureset])
|
|
mi@0
|
555 timbre_featureset = [x[:n_frame,:] for x in timbre_featureset]
|
|
mi@0
|
556 ao.timbre_features = np.hstack((timbre_featureset))
|
|
mi@0
|
557 else:
|
|
mi@0
|
558 ao.timbre_features = timbre_featureset[0]
|
|
mi@0
|
559 for feature in tempo_feature_list:
|
|
mi@0
|
560 for f in os.listdir(feature):
|
|
mi@0
|
561 if f[:f.find('_vamp')]==ao.name:
|
|
mi@0
|
562 tempo_featureset.append(np.genfromtxt(join(feature, f), delimiter=',',filling_values=0.0)[1:,1:])
|
|
mi@0
|
563 ao.tempo_timestamps = np.genfromtxt(join(feature, f), delimiter=',',filling_values=0.0)[1:,0]
|
|
mi@0
|
564 break
|
|
mi@0
|
565 if len(tempo_feature_list) > 1:
|
|
mi@0
|
566 n_frame = np.min([x.shape[0] for x in tempo_featureset])
|
|
mi@0
|
567 tempo_featureset = [x[:n_frame,:] for x in tempo_featureset]
|
|
mi@0
|
568 ao.tempo_features = np.hstack((tempo_featureset))
|
|
mi@0
|
569 else:
|
|
mi@0
|
570 ao.tempo_features = tempo_featureset[0]
|
|
mi@0
|
571 for feature in harmonic_feature_list:
|
|
mi@0
|
572 for f in os.listdir(feature):
|
|
mi@0
|
573 if f[:f.find('_vamp')]==ao.name:
|
|
mi@0
|
574 harmonic_featureset.append(np.genfromtxt(join(feature, f), delimiter=',',filling_values=0.0)[:,1:])
|
|
mi@0
|
575 break
|
|
mi@0
|
576 if len(harmonic_feature_list) > 1:
|
|
mi@0
|
577 n_frame = np.min([x.shape[0] for x in harmonic_featureset])
|
|
mi@0
|
578 harmonic_featureset = [x[:n_frame,:] for x in harmonic_featureset]
|
|
mi@0
|
579 ao.harmonic_features = np.hstack((harmonic_featureset))
|
|
mi@0
|
580 else:
|
|
mi@0
|
581 ao.harmonic_features = harmonic_featureset[0]
|
|
mi@0
|
582
|
|
mi@0
|
583 # # Reshaping features to keep identical dimension
|
|
mi@0
|
584 # n_frames = np.array([ao.gammatone_features.shape[0], ao.harmonic_features.shape[0], ao.timbre_features.shape[0]]).min()
|
|
mi@0
|
585 # step = n_frames / float(ao.tempo_features.shape[0])
|
|
mi@0
|
586 # # ao.tempo_features = self.upSample(ao.tempo_features, step)
|
|
mi@0
|
587 # ao.gammatone_features = ao.gammatone_features[:n_frames, :]
|
|
mi@0
|
588 # ao.harmonic_features = ao.harmonic_features[:n_frames, :]
|
|
mi@0
|
589 # ao.timbre_features = ao.timbre_features[:n_frames, :]
|
|
mi@0
|
590 # print ao.gammatone_features.shape, ao.harmonic_features.shape, ao.tempo_features.shape, ao.timbre_features.shape
|
|
mi@0
|
591
|
|
mi@0
|
592 # Reshape features (downsample) to match tempogram ones
|
|
mi@0
|
593 step = ao.tempo_features.shape[0]
|
|
mi@0
|
594 # aggregation_step = (n_frames / (step+1.0))
|
|
mi@0
|
595 # Get aggregated features for computing ssm
|
|
mi@0
|
596 aggregation_window, aggregation_step = 1,1
|
|
mi@0
|
597 featureRate = float(self.SampleRate) /self.stepSize
|
|
mi@0
|
598 pca = PCA(n_components=5)
|
|
mi@0
|
599
|
|
mi@0
|
600 ao.gammatone_features = resample(ao.gammatone_features, step)
|
|
mi@0
|
601 ao.gammatone_features = (ao.gammatone_features - np.min(ao.gammatone_features, axis=-1)[:,np.newaxis]) / (np.max(ao.gammatone_features, axis=-1) - np.min(ao.gammatone_features, axis=-1))[:,np.newaxis]
|
|
mi@0
|
602 ao.gammatone_features[np.isnan(ao.gammatone_features)] = 0.0
|
|
mi@0
|
603 ao.gammatone_features[np.isinf(ao.gammatone_features)] = 0.0
|
|
mi@0
|
604 ao.timbre_features = resample(ao.timbre_features, step)
|
|
mi@0
|
605 ao.timbre_features = (ao.timbre_features - np.min(ao.timbre_features, axis=-1)[:,np.newaxis]) / (np.max(ao.timbre_features, axis=-1) - np.min(ao.timbre_features, axis=-1))[:,np.newaxis]
|
|
mi@0
|
606 ao.timbre_features[np.isnan(ao.timbre_features)] = 0.0
|
|
mi@0
|
607 ao.timbre_features[np.isinf(ao.timbre_features)] = 0.0
|
|
mi@0
|
608 ao.harmonic_features = resample(ao.harmonic_features, step)
|
|
mi@0
|
609 ao.harmonic_features = (ao.harmonic_features - np.min(ao.harmonic_features, axis=-1)[:,np.newaxis]) / (np.max(ao.harmonic_features, axis=-1) - np.min(ao.harmonic_features, axis=-1))[:,np.newaxis]
|
|
mi@0
|
610 ao.harmonic_features[np.isnan(ao.harmonic_features)] = 0.0
|
|
mi@0
|
611 ao.harmonic_features[np.isinf(ao.harmonic_features)] = 0.0
|
|
mi@0
|
612 ao.tempo_features = (ao.tempo_features - np.min(ao.tempo_features, axis=-1)[:,np.newaxis]) / (np.max(ao.tempo_features, axis=-1) - np.min(ao.tempo_features, axis=-1))[:,np.newaxis]
|
|
mi@0
|
613 ao.tempo_features[np.isnan(ao.tempo_features)] = 0.0
|
|
mi@0
|
614 ao.tempo_features[np.isinf(ao.tempo_features)] = 0.0
|
|
mi@0
|
615 # print 'resampled', ao.gammatone_features.shape, ao.timbre_features.shape, ao.harmonic_features.shape
|
|
mi@0
|
616 # gt_feature_matrix = (ao.gammatone_features - np.min(ao.gammatone_features, axis=-1)[:,np.newaxis]) / (np.max(ao.gammatone_features, axis=-1) - np.min(ao.gammatone_features, axis=-1))[:,np.newaxis]
|
|
mi@0
|
617 # gt_feature_matrix[np.isnan(gt_feature_matrix)] = 0.0
|
|
mi@0
|
618 # mean_gt_feature = self.getMean(gt_feature_matrix, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
619 # std_gt_feature = self.getStd(gt_feature_matrix, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
620 # delta_gt_feature = self.getDelta(gt_feature_matrix)
|
|
mi@0
|
621 # mean_dgt_feature = self.getMean(delta_gt_feature, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
622 # std_dgt_feature = self.getStd(delta_gt_feature, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
623 # aggregated_gt_feature = np.hstack((mean_gt_feature, std_gt_feature))
|
|
mi@0
|
624 # aggregated_gt_feature = np.hstack((mean_gt_feature, std_gt_feature, mean_dgt_feature, std_dgt_feature))
|
|
mi@0
|
625 # aggregated_gt_feature = ao.gammatone_features
|
|
mi@0
|
626 aggregated_gt_feature = self.getMean(ao.gammatone_features, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
627 pca.fit(aggregated_gt_feature)
|
|
mi@0
|
628 aggregated_gt_feature = pca.transform(aggregated_gt_feature)
|
|
mi@0
|
629 distance_gt_matrix = pairwise_distances(aggregated_gt_feature, metric = 'cosine')
|
|
mi@0
|
630 distance_gt_matrix = np.nan_to_num(distance_gt_matrix)
|
|
mi@0
|
631 ao.gammatone_ssm = 1 - (distance_gt_matrix - distance_gt_matrix.min()) / (distance_gt_matrix.max() - distance_gt_matrix.min())
|
|
mi@0
|
632
|
|
mi@0
|
633 # tempo_feature_matrix = (ao.tempo_features - np.min(ao.tempo_features, axis=-1)[:,np.newaxis]) / (np.max(ao.tempo_features, axis=-1) - np.min(ao.tempo_features, axis=-1))[:,np.newaxis]
|
|
mi@0
|
634 # tempo_feature_matrix[np.isnan(tempo_feature_matrix)] = 0.0
|
|
mi@0
|
635 # mean_tempo_feature = self.getMean(tempo_feature_matrix, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
636 # std_tempo_feature = self.getStd(tempo_feature_matrix, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
637 # delta_tempo_feature = self.getDelta(tempo_feature_matrix)
|
|
mi@0
|
638 # mean_dtempo_feature = self.getMean(delta_tempo_feature, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
639 # std_dtempo_feature = self.getStd(delta_tempo_feature, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
640 # aggregated_tempo_feature = np.hstack((mean_tempo_feature, std_tempo_feature))
|
|
mi@0
|
641 # aggregated_tempo_feature = np.hstack((mean_tempo_feature, std_tempo_feature, mean_dtempo_feature, std_dtempo_feature))
|
|
mi@0
|
642 # aggregated_tempo_feature = ao.tempo_features
|
|
mi@0
|
643 aggregated_tempo_feature = self.getMean(ao.tempo_features, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
644 pca.fit(aggregated_tempo_feature)
|
|
mi@0
|
645 aggregated_tempo_feature = pca.transform(aggregated_tempo_feature)
|
|
mi@0
|
646 distance_tempo_matrix = pairwise_distances(aggregated_tempo_feature, metric = 'cosine')
|
|
mi@0
|
647 distance_tempo_matrix = np.nan_to_num(distance_tempo_matrix)
|
|
mi@0
|
648 ao.tempo_ssm = 1 - (distance_tempo_matrix - distance_tempo_matrix.min()) / (distance_tempo_matrix.max() - distance_tempo_matrix.min())
|
|
mi@0
|
649
|
|
mi@0
|
650 # timbre_feature_matrix = (ao.timbre_features - np.min(ao.timbre_features, axis=-1)[:,np.newaxis]) / (np.max(ao.timbre_features, axis=-1) - np.min(ao.timbre_features, axis=-1))[:,np.newaxis]
|
|
mi@0
|
651 # timbre_feature_matrix[np.isnan(timbre_feature_matrix)] = 0.0
|
|
mi@0
|
652 # mean_timbre_feature = self.getMean(timbre_feature_matrix, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
653 # std_timbre_feature = self.getStd(timbre_feature_matrix, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
654 # delta_timbre_feature = self.getDelta(timbre_feature_matrix)
|
|
mi@0
|
655 # mean_dtimbre_feature = self.getMean(delta_timbre_feature, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
656 # std_dtimbre_feature = self.getStd(delta_timbre_feature, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
657 # aggregated_timbre_feature = np.hstack((mean_timbre_feature, std_timbre_feature)
|
|
mi@0
|
658 # aggregated_timbre_feature = np.hstack((mean_timbre_feature, std_timbre_feature, mean_dtimbre_feature, std_dtimbre_feature))
|
|
mi@0
|
659 # aggregated_timbre_feature = ao.timbre_features
|
|
mi@0
|
660 aggregated_timbre_feature = self.getMean(ao.timbre_features, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
661 pca.fit(aggregated_timbre_feature)
|
|
mi@0
|
662 aggregated_timbre_feature = pca.transform(aggregated_timbre_feature)
|
|
mi@0
|
663 distance_timbre_matrix = pairwise_distances(aggregated_timbre_feature, metric = 'cosine')
|
|
mi@0
|
664 distance_timbre_matrix = np.nan_to_num(distance_timbre_matrix)
|
|
mi@0
|
665 ao.timbre_ssm = 1 - (distance_timbre_matrix - distance_timbre_matrix.min()) / (distance_timbre_matrix.max() - distance_timbre_matrix.min())
|
|
mi@0
|
666
|
|
mi@0
|
667 # harmonic_feature_matrix = (ao.harmonic_features - np.min(ao.harmonic_features, axis=-1)[:,np.newaxis]) / (np.max(ao.harmonic_features, axis=-1) - np.min(ao.harmonic_features, axis=-1))[:,np.newaxis]
|
|
mi@0
|
668 # harmonic_feature_matrix[np.isnan(harmonic_feature_matrix)] = 0.0
|
|
mi@0
|
669 # mean_harmonic_feature = self.getMean(harmonic_feature_matrix, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
670 # std_harmonic_feature = self.getStd(harmonic_feature_matrix, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
671 # delta_harmonic_feature = self.getDelta(harmonic_feature_matrix)
|
|
mi@0
|
672 # mean_dharmonic_feature = self.getMean(delta_harmonic_feature, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
673 # std_dharmonic_feature = self.getStd(delta_harmonic_feature, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
674 # aggregated_harmonic_feature = np.hstack((mean_harmonic_feature, std_harmonic_feature))
|
|
mi@0
|
675 # aggregated_harmonic_feature = np.hstack((mean_harmonic_feature, std_harmonic_feature, mean_dharmonic_feature, std_dharmonic_feature))
|
|
mi@0
|
676 aggregated_harmonic_feature = ao.harmonic_features
|
|
mi@0
|
677 aggregated_harmonic_feature = self.getMean(ao.harmonic_features, winlen=aggregation_window, stepsize=aggregation_step)
|
|
mi@0
|
678 pca.fit(aggregated_harmonic_feature)
|
|
mi@0
|
679 aggregated_harmonic_feature = pca.transform(aggregated_harmonic_feature)
|
|
mi@0
|
680 distance_harmonic_matrix = pairwise_distances(aggregated_harmonic_feature, metric = 'cosine')
|
|
mi@0
|
681 distance_harmonic_matrix = np.nan_to_num(distance_harmonic_matrix)
|
|
mi@0
|
682 ao.harmonic_ssm = 1 - (distance_harmonic_matrix - distance_harmonic_matrix.min()) / (distance_harmonic_matrix.max() - distance_harmonic_matrix.min())
|
|
mi@0
|
683
|
|
mi@0
|
684 ao.combined_features = np.hstack((aggregated_gt_feature, aggregated_harmonic_feature, aggregated_timbre_feature, aggregated_tempo_feature))
|
|
mi@0
|
685 pca.fit(ao.combined_features)
|
|
mi@0
|
686 ao.combined_features = pca.transform(ao.combined_features)
|
|
mi@0
|
687 distance_combined_matrix = pairwise_distances(ao.combined_features, metric = 'cosine')
|
|
mi@0
|
688 distance_combined_matrix = np.nan_to_num(distance_combined_matrix)
|
|
mi@0
|
689 ao.combined_ssm = 1 - (distance_combined_matrix - distance_combined_matrix.min()) / (distance_combined_matrix.max() - distance_combined_matrix.min())
|
|
mi@0
|
690
|
|
mi@0
|
691 # Resample timestamps
|
|
mi@0
|
692 # ao.ssm_timestamps = np.array(map(lambda step: step * aggregation_step / featureRate, np.arange(0.0, aggregated_gt_feature.shape[0])))
|
|
mi@0
|
693 ao.ssm_timestamps = np.array(map(lambda x: ao.tempo_timestamps[aggregation_step*x], np.arange(0, ao.gammatone_ssm.shape[0])))
|
|
mi@0
|
694 # print ao.gammatone_ssm.shape, ao.tempo_ssm.shape, ao.timbre_ssm.shape, ao.harmonic_ssm.shape, len(ao.ssm_timestamps)
|
|
mi@0
|
695
|
|
mi@0
|
696 # # Save SSMs.
|
|
mi@0
|
697 # gammatone_ssm = copy(ao.gammatone_ssm)
|
|
mi@0
|
698 # gammatone_ssm[gammatone_ssm<0.8]=0.0
|
|
mi@0
|
699 # plt.figure(figsize=(10, 10))
|
|
mi@0
|
700 # plt.vlines(ao.gt / ao.gt[-1] * gammatone_ssm.shape[0], 0, gammatone_ssm.shape[0], colors='r')
|
|
mi@0
|
701 # plt.imshow(gammatone_ssm, cmap='Greys')
|
|
mi@0
|
702 # plt.savefig(join(options.OUTPUT, 'ssm', ao.name+'-gammatone.pdf'),format='pdf')
|
|
mi@0
|
703 #
|
|
mi@0
|
704 # tempo_ssm = copy(ao.tempo_ssm)
|
|
mi@0
|
705 # tempo_ssm[tempo_ssm<0.8]=0.0
|
|
mi@0
|
706 # plt.figure(figsize=(10, 10))
|
|
mi@0
|
707 # plt.vlines(ao.gt / ao.gt[-1] * tempo_ssm.shape[0], 0, tempo_ssm.shape[0], colors='r')
|
|
mi@0
|
708 # plt.imshow(tempo_ssm, cmap='Greys')
|
|
mi@0
|
709 # plt.savefig(join(options.OUTPUT, 'ssm', ao.name+'-tempo.pdf'),format='pdf')
|
|
mi@0
|
710 #
|
|
mi@0
|
711 # timbre_ssm = copy(ao.timbre_ssm)
|
|
mi@0
|
712 # timbre_ssm[timbre_ssm<0.8]=0.0
|
|
mi@0
|
713 # plt.figure(figsize=(10, 10))
|
|
mi@0
|
714 # plt.vlines(ao.gt / ao.gt[-1] * timbre_ssm.shape[0], 0, timbre_ssm.shape[0], colors='r')
|
|
mi@0
|
715 # plt.imshow(timbre_ssm, cmap='Greys')
|
|
mi@0
|
716 # plt.savefig(join(options.OUTPUT, 'ssm', ao.name+'-timbre.pdf'),format='pdf')
|
|
mi@0
|
717 #
|
|
mi@0
|
718 # harmonic_ssm = copy(ao.harmonic_ssm)
|
|
mi@0
|
719 # harmonic_ssm[harmonic_ssm<0.8]=0.0
|
|
mi@0
|
720 # plt.figure(figsize=(10, 10))
|
|
mi@0
|
721 # plt.vlines(ao.gt / ao.gt[-1] * harmonic_ssm.shape[0], 0, harmonic_ssm.shape[0], colors='r')
|
|
mi@0
|
722 # plt.imshow(harmonic_ssm, cmap='Greys')
|
|
mi@0
|
723 # plt.savefig(join(options.OUTPUT, 'ssm', ao.name+'-harmonic.pdf'),format='pdf')
|
|
mi@0
|
724 #
|
|
mi@0
|
725 # ssm_cleaned = copy(ao.combined_ssm)
|
|
mi@0
|
726 # ssm_cleaned[ssm_cleaned<0.8] = 0
|
|
mi@0
|
727 # plt.figure(figsize=(10, 10))
|
|
mi@0
|
728 # plt.vlines(ao.gt / ao.gt[-1] * ssm_cleaned.shape[0], 0, ssm_cleaned.shape[0], colors='r')
|
|
mi@0
|
729 # plt.imshow(ssm_cleaned, cmap='Greys')
|
|
mi@0
|
730 # plt.savefig(join(options.OUTPUT, 'ssm', ao.name+'-combined.pdf'),format='pdf')
|
|
mi@0
|
731
|
|
mi@0
|
732 audio_list.append(ao)
|
|
mi@0
|
733
|
|
mi@0
|
734 # Evaluate individual segmentation results.
|
|
mi@0
|
735 outfile1 = join(options.OUTPUT, 'individualSOM.csv')
|
|
mi@0
|
736 with open(outfile1, 'a') as f:
|
|
mi@0
|
737 csvwriter = csv.writer(f, delimiter=',')
|
|
mi@0
|
738 csvwriter.writerow(['audio', 'gammatone_tp_05', 'gammatone_fp_05', 'gammatone_fn_05', 'gammatone_P_05', 'gammatone_R_05', 'gammatone_F_05', 'gammatone_AD_05', 'gammatone_DA_05', 'gammatone_tp_3', \
|
|
mi@0
|
739 'gammatone_fp_3', 'gammatone_fn_3', 'gammatone_P_3', 'gammatone_R_3', 'gammatone_F_3', 'gammatone_AD_3', 'gammatone_DA_3', 'harmonic_tp_05', 'harmonic_fp_05', 'harmonic_fn_05', 'harmonic_P_05', \
|
|
mi@0
|
740 'harmonic_R_05', 'harmonic_F_05', 'harmonic_AD_05', 'harmonic_DA_05', 'harmonic_tp_3', 'harmonic_fp_3', 'harmonic_fn_3', 'harmonic_P_3', 'harmonic_R_3', 'harmonic_F_3', 'harmonic_AD_3', 'harmonic_DA_3', \
|
|
mi@0
|
741 'timbre_tp_05', 'timbre_fp_05', 'timbre_fn_05', 'timbre_P_05', 'timbre_R_05', 'timbre_F_05', 'timbre_AD_05', 'timbre_DA_05', 'timbre_tp_3', 'timbre_fp_3', 'timbre_fn_3', 'timbre_P_3', 'timbre_R_3', \
|
|
mi@0
|
742 'timbre_F_3', 'timbre_AD_3', 'timbre_DA_3', 'tempo_tp_05', 'tempo_fp_05', 'tempo_fn_05', 'tempo_P_05', 'tempo_R_05', 'tempo_F_05', 'tempo_AD_05', 'tempo_DA_05', \
|
|
mi@0
|
743 'tempo_tp_3', 'tempo_fp_3', 'tempo_fn_3', 'tempo_P_3', 'tempo_R_3', 'tempo_F_3', 'tempo_AD_3', 'tempo_DA_3'])
|
|
mi@0
|
744
|
|
mi@0
|
745 # outfile4 = join(options.OUTPUT, 'individualResDF.csv')
|
|
mi@0
|
746 # with open(outfile4, 'a') as f:
|
|
mi@0
|
747 # csvwriter = csv.writer(f, delimiter=',')
|
|
mi@0
|
748 # csvwriter.writerow(['audio', 'gammatone_tp_05', 'gammatone_fp_05', 'gammatone_fn_05', 'gammatone_P_05', 'gammatone_R_05', 'gammatone_F_05', 'gammatone_AD_05', 'gammatone_DA_05', 'gammatone_tp_3', \
|
|
mi@0
|
749 # 'gammatone_fp_3', 'gammatone_fn_3', 'gammatone_P_3', 'gammatone_R_3', 'gammatone_F_3', 'gammatone_AD_3', 'gammatone_DA_3', 'harmonic_tp_05', 'harmonic_fp_05', 'harmonic_fn_05', 'harmonic_P_05', \
|
|
mi@0
|
750 # 'harmonic_R_05', 'harmonic_F_05', 'harmonic_AD_05', 'harmonic_DA_05', 'harmonic_tp_3', 'harmonic_fp_3', 'harmonic_fn_3', 'harmonic_P_3', 'harmonic_R_3', 'harmonic_F_3', 'harmonic_AD_3', 'harmonic_DA_3', \
|
|
mi@0
|
751 # 'timbre_tp_05', 'timbre_fp_05', 'timbre_fn_05', 'timbre_P_05', 'timbre_R_05', 'timbre_F_05', 'timbre_AD_05', 'timbre_DA_05', 'timbre_tp_3', 'timbre_fp_3', 'timbre_fn_3', 'timbre_P_3', 'timbre_R_3', \
|
|
mi@0
|
752 # 'timbre_F_3', 'timbre_AD_3', 'timbre_DA_3', 'tempo_tp_05', 'tempo_fp_05', 'tempo_fn_05', 'tempo_P_05', 'tempo_R_05', 'tempo_F_05', 'tempo_AD_05', 'tempo_DA_05', \
|
|
mi@0
|
753 # 'tempo_tp_3', 'tempo_fp_3', 'tempo_fn_3', 'tempo_P_3', 'tempo_R_3', 'tempo_F_3', 'tempo_AD_3', 'tempo_DA_3'])
|
|
mi@0
|
754
|
|
mi@0
|
755 # Fuse novelty curves from individual segmentation results.
|
|
mi@0
|
756 outfile2 = join(options.OUTPUT, 'individualFuseSOM.csv')
|
|
mi@0
|
757 with open(outfile2, 'a') as f:
|
|
mi@0
|
758 csvwriter = csv.writer(f, delimiter=',')
|
|
mi@0
|
759 csvwriter.writerow(['audio', 'gt_tb_P_0.5', 'gt_tb_R_0.5', 'gt_tb_F_0.5', 'gt_tb_P_3', 'gt_tb_R_3', 'gt_tb_F_3', 'gt_tp_P_0.5', 'gt_tp_R_0.5', 'gt_tp_F_0.5', 'gt_tp_P_3', 'gt_tp_R_3', 'gt_tp_F_3',\
|
|
mi@0
|
760 'gt_hm_P_0.5', 'gt_hm_R_0.5', 'gt_hm_F_0.5', 'gt_hm_P_3', 'gt_hm_R_3', 'gt_hm_F_3', 'tb_tp_P_0.5', 'tb_tp_R_0.5', 'tb_tp_F_0.5', 'tb_tp_P_3', 'tb_tp_R_3', 'tb_tp_F_3', 'tb_hm_P_0.5', 'tb_hm_R_0.5', 'tb_hm_F_0.5', \
|
|
mi@0
|
761 'tb_hm_P_3', 'tb_hm_R_3', 'tb_hm_F_3', 'tp_hm_P_0.5', 'tp_hm_R_0.5', 'tp_hm_F_0.5', 'tp_hm_P_3', 'tp_hm_R_3', 'tp_hm_F_3', 'gt_tb_tp_P_0.5', 'gt_tb_tp_R_0.5', 'gt_tb_tp_F_0.5', 'gt_tb_tp_P_3', 'gt_tb_tp_R_3', \
|
|
mi@0
|
762 'gt_tb_tp_F_3', 'gt_tb_hm_P_0.5', 'gt_tb_hm_R_0.5', 'gt_tb_hm_F_0.5', 'gt_tb_hm_P_3', 'gt_tb_hm_R_3', 'gt_tb_hm_F_3', 'gt_tp_hm_P_0.5', 'gt_tp_hm_R_0.5', 'gt_tp_hm_F_0.5', 'gt_tp_hm_P_3', 'gt_tp_hm_R_3', 'gt_tp_hm_F_3', \
|
|
mi@0
|
763 'tb_tp_hm_P_0.5', 'tb_tp_hm_R_0.5', 'tb_tp_hm_F_0.5', 'tb_tp_hm_P_3', 'tb_tp_hm_R_3', 'tb_tp_hm_F_3', 'gt_tb_tp_hm_P_0.5', 'gt_tb_tp_hm_R_0.5', 'gt_tb_tp_hm_F_0.5', 'gt_tb_tp_hm_P_3', 'gt_tb_tp_hm_R_3', 'gt_tb_tp_hm_F_3'])
|
|
mi@0
|
764
|
|
mi@0
|
765
|
|
mi@0
|
766 for i,ao in enumerate(audio_list):
|
|
mi@0
|
767
|
|
mi@0
|
768 print 'processing self organizing maps for %s' %ao.name
|
|
mi@0
|
769
|
|
mi@0
|
770 # 1.Novelty based segmentation.
|
|
mi@0
|
771 # Correlate an Gaussian on the diagonal to contruct the novelty curve
|
|
mi@0
|
772 # print 'ssm', ao.gammatone_ssm.shape, ao.timbre_ssm.shape, ao.tempo_ssm.shape, ao.harmonic_ssm.shape
|
|
mi@0
|
773 ssom.train(ao.gammatone_features)
|
|
mi@0
|
774 gammatone_som = ssom(ao.gammatone_features)
|
|
mi@0
|
775 ssom.train(ao.timbre_features)
|
|
mi@0
|
776 timbre_som = ssom(ao.timbre_features)
|
|
mi@0
|
777 ssom.train(ao.tempo_features)
|
|
mi@0
|
778 tempo_som = ssom(ao.tempo_features)
|
|
mi@0
|
779 ssom.train(ao.harmonic_features)
|
|
mi@0
|
780 harmonic_som = ssom(ao.harmonic_features)
|
|
mi@0
|
781
|
|
mi@0
|
782 gammatone_harmonic_features = np.hstack((ao.gammatone_features, ao.harmonic_features))
|
|
mi@0
|
783 gammatone_timbre_features = np.hstack((ao.gammatone_features, ao.timbre_features))
|
|
mi@0
|
784 gammatone_tempo_features = np.hstack((ao.gammatone_features, ao.tempo_features))
|
|
mi@0
|
785 harmonic_timbre_features = np.hstack((ao.harmonic_features, ao.timbre_features))
|
|
mi@0
|
786 harmonic_tempo_features = np.hstack((ao.harmonic_features, ao.tempo_features))
|
|
mi@0
|
787 timbre_tempo_features = np.hstack((ao.timbre_features, ao.tempo_features))
|
|
mi@0
|
788
|
|
mi@0
|
789 gammatone_harmonic_timbre_features = np.hstack((ao.gammatone_features, ao.harmonic_features, ao.timbre_features))
|
|
mi@0
|
790 gammatone_harmonic_tempo_features = np.hstack((ao.gammatone_features, ao.harmonic_features, ao.tempo_features))
|
|
mi@0
|
791 gammatone_timbre_tempo_features = np.hstack((ao.gammatone_features, ao.timbre_features, ao.tempo_features))
|
|
mi@0
|
792 harmonic_timbre_tempo_features = np.hstack((ao.harmonic_features, ao.timbre_features, ao.tempo_features))
|
|
mi@0
|
793
|
|
mi@0
|
794 gammatone_harmonic_timbre_tempo_features = np.hstack((ao.gammatone_features, ao.harmonic_features, ao.timbre_features, ao.tempo_features))
|
|
mi@0
|
795
|
|
mi@0
|
796 ssom.train(gammatone_harmonic_features)
|
|
mi@0
|
797 gammatone_harmonic_som = ssom(gammatone_harmonic_features)
|
|
mi@0
|
798 ssom.train(gammatone_timbre_features)
|
|
mi@0
|
799 gammatone_timbre_som = ssom(gammatone_timbre_features)
|
|
mi@0
|
800 ssom.train(gammatone_tempo_features)
|
|
mi@0
|
801 gammatone_tempo_som = ssom(gammatone_tempo_features)
|
|
mi@0
|
802 ssom.train(harmonic_timbre_features)
|
|
mi@0
|
803 harmonic_timbre_som = ssom(harmonic_timbre_features)
|
|
mi@0
|
804 ssom.train(harmonic_timbre_features)
|
|
mi@0
|
805 harmonic_timbre_som = ssom(harmonic_timbre_features)
|
|
mi@0
|
806 ssom.train(harmonic_tempo_features)
|
|
mi@0
|
807 harmonic_tempo_som = ssom(harmonic_tempo_features)
|
|
mi@0
|
808 ssom.train(timbre_tempo_features)
|
|
mi@0
|
809 timbre_tempo_som = ssom(timbre_tempo_features)
|
|
mi@0
|
810
|
|
mi@0
|
811 ssom.train(gammatone_harmonic_timbre_features)
|
|
mi@0
|
812 gammatone_harmonic_timbre_som = ssom(gammatone_harmonic_timbre_features)
|
|
mi@0
|
813 ssom.train(gammatone_harmonic_tempo_features)
|
|
mi@0
|
814 gammatone_harmonic_tempo_som = ssom(gammatone_harmonic_tempo_features)
|
|
mi@0
|
815 ssom.train(gammatone_timbre_tempo_features)
|
|
mi@0
|
816 gammatone_timbre_tempo_som = ssom(gammatone_timbre_tempo_features)
|
|
mi@0
|
817 ssom.train(harmonic_timbre_tempo_features)
|
|
mi@0
|
818 harmonic_timbre_tempo_som = ssom(harmonic_timbre_tempo_features)
|
|
mi@0
|
819
|
|
mi@0
|
820 ssom.train(gammatone_harmonic_timbre_tempo_features)
|
|
mi@0
|
821 gammatone_harmonic_timbre_tempo_som = ssom(gammatone_harmonic_timbre_tempo_features)
|
|
mi@0
|
822
|
|
mi@0
|
823 gammatone_ssm = self.getSSM(gammatone_som)
|
|
mi@0
|
824 harmonic_ssm = self.getSSM(harmonic_som)
|
|
mi@0
|
825 timbre_ssm = self.getSSM(timbre_som)
|
|
mi@0
|
826 tempo_ssm = self.getSSM(tempo_som)
|
|
mi@0
|
827 gammatone_harmonic_ssm = self.getSSM(gammatone_harmonic_som)
|
|
mi@0
|
828 gammatone_timbre_ssm = self.getSSM(gammatone_timbre_som)
|
|
mi@0
|
829 gammatone_tempo_ssm = self.getSSM(gammatone_tempo_som)
|
|
mi@0
|
830 harmonic_timbre_ssm = self.getSSM(harmonic_timbre_som)
|
|
mi@0
|
831 harmonic_tempo_ssm = self.getSSM(harmonic_tempo_som)
|
|
mi@0
|
832 timbre_tempo_ssm = self.getSSM(timbre_tempo_som)
|
|
mi@0
|
833 gammatone_harmonic_timbre_ssm = self.getSSM(gammatone_harmonic_timbre_som)
|
|
mi@0
|
834 gammatone_harmonic_tempo_ssm = self.getSSM(gammatone_harmonic_tempo_som)
|
|
mi@0
|
835 gammatone_timbre_tempo_ssm = self.getSSM(gammatone_timbre_tempo_som)
|
|
mi@0
|
836 harmonic_timbre_tempo_ssm = self.getSSM(harmonic_timbre_tempo_som)
|
|
mi@0
|
837 gammatone_harmonic_timbre_tempo_ssm = self.getSSM(gammatone_harmonic_timbre_tempo_som)
|
|
mi@0
|
838
|
|
mi@0
|
839
|
|
mi@0
|
840 # Noise removal in ssm
|
|
mi@0
|
841 reduced_gammatone_ssm = self.reduceSSM(gammatone_ssm)
|
|
mi@0
|
842 reduced_timbre_ssm = self.reduceSSM(timbre_ssm)
|
|
mi@0
|
843 reduced_tempo_ssm = self.reduceSSM(ao.tempo_ssm)
|
|
mi@0
|
844 reduced_harmonic_ssm = self.reduceSSM(ao.harmonic_ssm)
|
|
mi@0
|
845 reduced_gammatone_harmonic_ssm = self.reduceSSM(gammatone_harmonic_ssm)
|
|
mi@0
|
846 reduced_gammatone_timbre_ssm = self.reduceSSM(gammatone_timbre_ssm)
|
|
mi@0
|
847 reduced_gammatone_tempo_ssm = self.reduceSSM(gammatone_tempo_ssm)
|
|
mi@0
|
848 reduced_harmonic_timbre_ssm = self.reduceSSM(harmonic_timbre_ssm)
|
|
mi@0
|
849 reduced_harmonic_tempo_ssm = self.reduceSSM(harmonic_tempo_ssm)
|
|
mi@0
|
850 reduced_timbre_tempo_ssm = self.reduceSSM(timbre_tempo_ssm)
|
|
mi@0
|
851 reduced_gammatone_harmonic_timbre_ssm = self.reduceSSM(gammatone_harmonic_timbre_ssm)
|
|
mi@0
|
852 reduced_gammatone_harmonic_tempo_ssm = self.reduceSSM(gammatone_harmonic_tempo_ssm)
|
|
mi@0
|
853 reduced_gammatone_timbre_tempo_ssm = self.reduceSSM(gammatone_timbre_tempo_ssm)
|
|
mi@0
|
854 reduced_harmonic_timbre_tempo_ssm = self.reduceSSM(harmonic_timbre_tempo_ssm)
|
|
mi@0
|
855 reduced_gammatone_harmonic_timbre_tempo_ssm = self.reduceSSM(gammatone_harmonic_timbre_tempo_ssm)
|
|
mi@0
|
856
|
|
mi@0
|
857
|
|
mi@0
|
858 gammatone_novelty = self.getNoveltyCurve(reduced_gammatone_ssm, self.kernel_size)
|
|
mi@0
|
859 gammatone_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gammatone_novelty]
|
|
mi@0
|
860 timbre_novelty = self.getNoveltyCurve(reduced_timbre_ssm, self.kernel_size)
|
|
mi@0
|
861 timbre_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in timbre_novelty]
|
|
mi@0
|
862 tempo_novelty = self.getNoveltyCurve(reduced_tempo_ssm, self.kernel_size)
|
|
mi@0
|
863 tempo_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in tempo_novelty]
|
|
mi@0
|
864 harmonic_novelty = self.getNoveltyCurve(reduced_harmonic_ssm, self.kernel_size)
|
|
mi@0
|
865 harmonic_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in harmonic_novelty]
|
|
mi@0
|
866
|
|
mi@0
|
867 # Peak picking from the novelty curve
|
|
mi@0
|
868 smoothed_gammatone_novelty, gammatone_novelty_peaks = peak_picker.process(gammatone_novelty)
|
|
mi@0
|
869 gammatone_detection = [ao.ssm_timestamps[int(i)] for i in gammatone_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
870 smoothed_timbre_novelty, timbre_novelty_peaks = peak_picker.process(timbre_novelty)
|
|
mi@0
|
871 timbre_detection = [ao.ssm_timestamps[int(i)] for i in timbre_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
872 smoothed_harmonic_novelty, harmonic_novelty_peaks = peak_picker.process(harmonic_novelty)
|
|
mi@0
|
873 harmonic_detection = [ao.ssm_timestamps[int(i)] for i in harmonic_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
874 smoothed_tempo_novelty, tempo_novelty_peaks = peak_picker.process(tempo_novelty)
|
|
mi@0
|
875 tempo_detection = [ao.ssm_timestamps[int(i)] for i in tempo_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
876
|
|
mi@0
|
877 gt_res_05 = self.pairwiseF(ao.gt, gammatone_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
878 gt_res_3 = self.pairwiseF(ao.gt, gammatone_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
879 harmonic_res_05 = self.pairwiseF(ao.gt, harmonic_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
880 harmonic_res_3 = self.pairwiseF(ao.gt, harmonic_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
881 tempo_res_05 = self.pairwiseF(ao.gt, tempo_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
882 tempo_res_3 = self.pairwiseF(ao.gt, tempo_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
883 timbre_res_05 = self.pairwiseF(ao.gt, timbre_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
884 timbre_res_3 = self.pairwiseF(ao.gt, timbre_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
885
|
|
mi@0
|
886 with open(outfile1, 'a') as f:
|
|
mi@0
|
887 csvwriter = csv.writer(f, delimiter=',')
|
|
mi@0
|
888 csvwriter.writerow([ao.name, gt_res_05.TP, gt_res_05.FP, gt_res_05.FN, gt_res_05.P, gt_res_05.R, gt_res_05.F, gt_res_05.AD, gt_res_05.DA, gt_res_3.TP, gt_res_3.FP, gt_res_3.FN, gt_res_3.P, \
|
|
mi@0
|
889 gt_res_3.R, gt_res_3.F, gt_res_3.AD, gt_res_3.DA, harmonic_res_05.TP, harmonic_res_05.FP, harmonic_res_05.FN, harmonic_res_05.P, harmonic_res_05.R, harmonic_res_05.F, harmonic_res_05.AD, harmonic_res_05.DA, \
|
|
mi@0
|
890 harmonic_res_3.TP, harmonic_res_3.FP, harmonic_res_3.FN, harmonic_res_3.P, harmonic_res_3.R, harmonic_res_3.F, harmonic_res_3.AD, harmonic_res_3.DA, timbre_res_05.TP, timbre_res_05.FP, \
|
|
mi@0
|
891 timbre_res_05.FN, timbre_res_05.P, timbre_res_05.R, timbre_res_05.F, timbre_res_05.AD, timbre_res_05.DA, timbre_res_3.TP, timbre_res_3.FP, timbre_res_3.FN, timbre_res_3.P, timbre_res_3.R, timbre_res_3.F, \
|
|
mi@0
|
892 timbre_res_3.AD, timbre_res_3.DA, tempo_res_05.TP, tempo_res_05.FP, tempo_res_05.FN, tempo_res_05.P, tempo_res_05.R, tempo_res_05.F, tempo_res_05.AD, tempo_res_05.DA, tempo_res_3.TP, tempo_res_3.FP, \
|
|
mi@0
|
893 tempo_res_3.FN, tempo_res_3.P, tempo_res_3.R, tempo_res_3.F, tempo_res_3.AD, tempo_res_3.DA])
|
|
mi@0
|
894
|
|
mi@0
|
895 gt_hm_novelty = self.getNoveltyCurve(reduced_gammatone_harmonic_ssm, self.kernel_size)
|
|
mi@0
|
896 gt_hm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_hm_novelty]
|
|
mi@0
|
897 gt_tb_novelty = self.getNoveltyCurve(reduced_gammatone_timbre_ssm, self.kernel_size)
|
|
mi@0
|
898 gt_tb_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_tb_novelty]
|
|
mi@0
|
899 gt_tp_novelty = self.getNoveltyCurve(reduced_gammatone_tempo_ssm, self.kernel_size)
|
|
mi@0
|
900 gt_tp_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_tp_novelty]
|
|
mi@0
|
901 hm_tb_novelty = self.getNoveltyCurve(reduced_harmonic_timbre_ssm, self.kernel_size)
|
|
mi@0
|
902 hm_tb_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in hm_tb_novelty]
|
|
mi@0
|
903 hm_tp_novelty = self.getNoveltyCurve(reduced_harmonic_tempo_ssm, self.kernel_size)
|
|
mi@0
|
904 hm_tp_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in hm_tp_novelty]
|
|
mi@0
|
905 tb_tp_novelty = self.getNoveltyCurve(reduced_timbre_tempo_ssm, self.kernel_size)
|
|
mi@0
|
906 tb_tp_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in tb_tp_novelty]
|
|
mi@0
|
907
|
|
mi@0
|
908 smoothed_gt_tb_novelty, gt_tb_novelty_peaks = peak_picker.process(gt_tb_novelty)
|
|
mi@0
|
909 gt_tb_detection = [ao.ssm_timestamps[int(i)] for i in gt_tb_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
910 smoothed_gt_tp_novelty, gt_tp_novelty_peaks = peak_picker.process(gt_tp_novelty)
|
|
mi@0
|
911 gt_tp_detection = [ao.ssm_timestamps[int(i)] for i in gt_tp_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
912 smoothed_gt_hm_novelty, gt_hm_novelty_peaks = peak_picker.process(gt_hm_novelty)
|
|
mi@0
|
913 gt_hm_detection = [ao.ssm_timestamps[int(i)] for i in gt_hm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
914 smoothed_tb_tp_novelty, tb_tp_novelty_peaks = peak_picker.process(tb_tp_novelty)
|
|
mi@0
|
915 tb_tp_detection = [ao.ssm_timestamps[int(i)] for i in tb_tp_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
916 smoothed_tb_hm_novelty, tb_hm_novelty_peaks = peak_picker.process(tb_hm_novelty)
|
|
mi@0
|
917 tb_hm_detection = [ao.ssm_timestamps[int(i)] for i in tb_hm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
918 smoothed_tp_hm_novelty, tp_hm_novelty_peaks = peak_picker.process(tp_hm_novelty)
|
|
mi@0
|
919 tp_hm_detection = [ao.ssm_timestamps[int(i)] for i in tp_hm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
920
|
|
mi@0
|
921 gt_tb_tp_novelty = self.getNoveltyCurve(reduced_gammatone_timbre_tempo_ssm, self.kernel_size)
|
|
mi@0
|
922 gt_tb_tp_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_tb_tp_novelty]
|
|
mi@0
|
923 gt_tb_hm_novelty = self.getNoveltyCurve(reduced_gammatone_harmonic_timbre_ssm, self.kernel_size)
|
|
mi@0
|
924 gt_tb_hm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_tb_hm_novelty]
|
|
mi@0
|
925 gt_tp_hm_novelty = self.getNoveltyCurve(reduced_gammatone_harmonic_tempo_ssm, self.kernel_size)
|
|
mi@0
|
926 gt_tp_hm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_tp_hm_novelty]
|
|
mi@0
|
927 tb_tp_hm_novelty = self.getNoveltyCurve(reduced_harmonic_timbre_tempo_ssm, self.kernel_size)
|
|
mi@0
|
928 tb_tp_hm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in tb_tp_hm_novelty]
|
|
mi@0
|
929 gt_tb_tp_hm_novelty = self.getNoveltyCurve(reduced_gammatone_harmonic_timbre_tempo_ssm, self.kernel_size)
|
|
mi@0
|
930 gt_tb_tp_hm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_tb_tp_hm_novelty]
|
|
mi@0
|
931
|
|
mi@0
|
932 smoothed_gt_tb_tp_novelty, gt_tb_tp_novelty_peaks = peak_picker.process(gt_tb_tp_novelty)
|
|
mi@0
|
933 gt_tb_tp_detection = [ao.ssm_timestamps[int(i)] for i in gt_tb_tp_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
934 smoothed_gt_tb_hm_novelty, gt_tb_hm_novelty_peaks = peak_picker.process(gt_tb_hm_novelty)
|
|
mi@0
|
935 gt_tb_hm_detection = [ao.ssm_timestamps[int(i)] for i in gt_tb_hm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
936 smoothed_gt_tp_hm_novelty, gt_tp_hm_novelty_peaks = peak_picker.process(gt_tp_hm_novelty)
|
|
mi@0
|
937 gt_tp_hm_detection = [ao.ssm_timestamps[int(i)] for i in gt_tp_hm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
938 smoothed_tb_tp_hm_novelty, tb_tp_hm_novelty_peaks = peak_picker.process(tb_tp_hm_novelty)
|
|
mi@0
|
939 tb_tp_hm_detection = [ao.ssm_timestamps[int(i)] for i in tb_tp_hm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
940 smoothed_gt_tb_tp_hm_novelty, gt_tb_tp_hm_novelty_peaks = peak_picker.process(gt_tb_tp_hm_novelty)
|
|
mi@0
|
941 gt_tb_tp_hm_detection = [ao.ssm_timestamps[int(i)] for i in gt_tb_tp_hm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
942
|
|
mi@0
|
943 # novelty_peaks = gt_tb_tp_hm_novelty_peaks
|
|
mi@0
|
944 # novelty_detection = [ao.ssm_timestamps[int(i)] for i in novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
945
|
|
mi@0
|
946 if options.PLOT:
|
|
mi@0
|
947 self.plotDetection(ao.ssm, novelty, smoothed_novelty, ao.gt, detection, filename=join(options.OUTPUT+ ao.name)+'.pdf')
|
|
mi@0
|
948
|
|
mi@0
|
949 gt_tb_res_05 = self.pairwiseF(ao.gt, gt_tb_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
950 gt_tb_res_3 = self.pairwiseF(ao.gt, gt_tb_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
951 gt_tp_res_05 = self.pairwiseF(ao.gt, gt_tp_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
952 gt_tp_res_3 = self.pairwiseF(ao.gt, gt_tp_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
953 gt_hm_res_05 = self.pairwiseF(ao.gt, gt_hm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
954 gt_hm_res_3 = self.pairwiseF(ao.gt, gt_hm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
955 tb_tp_res_05 = self.pairwiseF(ao.gt, tb_tp_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
956 tb_tp_res_3 = self.pairwiseF(ao.gt, tb_tp_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
957 tb_hm_res_05 = self.pairwiseF(ao.gt, tb_hm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
958 tb_hm_res_3 = self.pairwiseF(ao.gt, tb_hm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
959 tp_hm_res_05 = self.pairwiseF(ao.gt, tp_hm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
960 tp_hm_res_3 = self.pairwiseF(ao.gt, tp_hm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
961
|
|
mi@0
|
962 gt_tb_tp_res_05 = self.pairwiseF(ao.gt, gt_tb_tp_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
963 gt_tb_tp_res_3 = self.pairwiseF(ao.gt, gt_tb_tp_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
964 gt_tb_hm_res_05 = self.pairwiseF(ao.gt, gt_tb_hm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
965 gt_tb_hm_res_3 = self.pairwiseF(ao.gt, gt_tb_hm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
966 gt_tp_hm_res_05 = self.pairwiseF(ao.gt, gt_tp_hm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
967 gt_tp_hm_res_3 = self.pairwiseF(ao.gt, gt_tp_hm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
968 tb_tp_hm_res_05 = self.pairwiseF(ao.gt, tb_tp_hm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
969 tb_tp_hm_res_3 = self.pairwiseF(ao.gt, tb_tp_hm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
970
|
|
mi@0
|
971 gt_tb_tp_hm_res_05 = self.pairwiseF(ao.gt, gt_tb_tp_hm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
972 gt_tb_tp_hm_res_3 = self.pairwiseF(ao.gt, gt_tb_tp_hm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
973
|
|
mi@0
|
974
|
|
mi@0
|
975 # Output detected segment locations.
|
|
mi@0
|
976 if options.VERBOSE:
|
|
mi@0
|
977 outdir = join(options.OUTPUT, 'detection', ao.name)
|
|
mi@0
|
978 if not isdir(outdir):
|
|
mi@0
|
979 os.mkdir(outdir)
|
|
mi@0
|
980 np.savetxt(join(outdir, 'gammatone.csv'), gammatone_detection)
|
|
mi@0
|
981 np.savetxt(join(outdir, 'timbre.csv'), timbre_detection)
|
|
mi@0
|
982 np.savetxt(join(outdir, 'tempo.csv'), tempo_detection)
|
|
mi@0
|
983 np.savetxt(join(outdir, 'harmonic.csv'), harmonic_detection)
|
|
mi@0
|
984
|
|
mi@0
|
985 np.savetxt(join(outdir, 'gammatone_timbre_novelty.csv'), gt_tb_detection)
|
|
mi@0
|
986 np.savetxt(join(outdir, 'gammatone_tempo_novelty.csv'), gt_tp_detection)
|
|
mi@0
|
987 np.savetxt(join(outdir, 'gammatone_harmonic_novelty.csv'), gt_hm_detection)
|
|
mi@0
|
988 np.savetxt(join(outdir, 'timbre_tempo_novelty.csv'), tb_tp_detection)
|
|
mi@0
|
989 np.savetxt(join(outdir, 'timbre_harmonic_novelty.csv'), tb_hm_detection)
|
|
mi@0
|
990 np.savetxt(join(outdir, 'tempo_harmonic_novelty.csv'), tp_hm_detection)
|
|
mi@0
|
991
|
|
mi@0
|
992 np.savetxt(join(outdir, 'gammatone_timbre_tempo_novelty.csv'), gt_tb_tp_detection)
|
|
mi@0
|
993 np.savetxt(join(outdir, 'gammatone_timbre_harmonic_novelty.csv'), gt_tb_hm_detection)
|
|
mi@0
|
994 np.savetxt(join(outdir, 'gammatone_tempo_harmonic_novelty.csv'), gt_tp_hm_detection)
|
|
mi@0
|
995 np.savetxt(join(outdir, 'timbre_tempo_harmonic_novelty.csv'), tb_tp_hm_detection)
|
|
mi@0
|
996 np.savetxt(join(outdir, 'gammatone_timbre_tempo_harmonic_novelty.csv'), gt_tb_tp_hm_detection)
|
|
mi@0
|
997
|
|
mi@0
|
998 # with open(outfile4, 'a') as f:
|
|
mi@0
|
999 # csvwriter = csv.writer(f, delimiter=',')
|
|
mi@0
|
1000 # csvwriter.writerow([ao.name, gt_df_05.TP, gt_df_05.FP, gt_df_05.FN, gt_df_05.P, gt_df_05.R, gt_df_05.F, gt_df_05.AD, gt_df_05.DA, gt_df_3.TP, gt_df_3.FP, gt_df_3.FN, gt_df_3.P, \
|
|
mi@0
|
1001 # gt_df_3.R, gt_df_3.F, gt_df_3.AD, gt_df_3.DA, harmonic_df_05.TP, harmonic_df_05.FP, harmonic_df_05.FN, harmonic_df_05.P, harmonic_df_05.R, harmonic_df_05.F, harmonic_df_05.AD, harmonic_df_05.DA, \
|
|
mi@0
|
1002 # harmonic_df_3.TP, harmonic_df_3.FP, harmonic_df_3.FN, harmonic_df_3.P, harmonic_df_3.R, harmonic_df_3.F, harmonic_df_3.AD, harmonic_df_3.DA, timbre_df_05.TP, timbre_df_05.FP, \
|
|
mi@0
|
1003 # timbre_df_05.FN, timbre_df_05.P, timbre_df_05.R, timbre_df_05.F, timbre_df_05.AD, timbre_df_05.DA, timbre_df_3.TP, timbre_df_3.FP, timbre_df_3.FN, timbre_df_3.P, timbre_df_3.R, timbre_df_3.F, \
|
|
mi@0
|
1004 # timbre_df_3.AD, timbre_df_3.DA, tempo_df_05.TP, tempo_df_05.FP, tempo_df_05.FN, tempo_df_05.P, tempo_df_05.R, tempo_df_05.F, tempo_df_05.AD, tempo_df_05.DA, tempo_df_3.TP, tempo_df_3.FP, \
|
|
mi@0
|
1005 # tempo_df_3.FN, tempo_df_3.P, tempo_df_3.R, tempo_df_3.F, tempo_df_3.AD, tempo_df_3.DA])
|
|
mi@0
|
1006
|
|
mi@0
|
1007 with open(outfile2, 'a') as f:
|
|
mi@0
|
1008 csvwriter = csv.writer(f, delimiter=',')
|
|
mi@0
|
1009 csvwriter.writerow([ao.name, gt_tb_res_05.P, gt_tb_res_05.R, gt_tb_res_05.F, gt_tb_res_3.P, gt_tb_res_3.R, gt_tb_res_3.F, gt_tp_res_05.P, gt_tp_res_05.R, gt_tp_res_05.F, gt_tp_res_3.P, gt_tp_res_3.R, gt_tp_res_3.F, \
|
|
mi@0
|
1010 gt_hm_res_05.P, gt_hm_res_05.R, gt_hm_res_05.F, gt_hm_res_3.P, gt_hm_res_3.R, gt_hm_res_3.F, tb_tp_res_05.P, tb_tp_res_05.R, tb_tp_res_05.F, tb_tp_res_3.P, tb_tp_res_3.R, tb_tp_res_3.F, \
|
|
mi@0
|
1011 tb_hm_res_05.P, tb_hm_res_05.R, tb_hm_res_05.F, tb_hm_res_3.P, tb_hm_res_3.R, tb_hm_res_3.F, tp_hm_res_05.P, tp_hm_res_05.R, tp_hm_res_05.F, tp_hm_res_3.P, tp_hm_res_3.R, tp_hm_res_3.F, \
|
|
mi@0
|
1012 gt_tb_tp_res_05.P, gt_tb_tp_res_05.R, gt_tb_tp_res_05.F, gt_tb_tp_res_3.P, gt_tb_tp_res_3.R, gt_tb_tp_res_3.F, gt_tb_hm_res_05.P, gt_tb_hm_res_05.R, gt_tb_hm_res_05.F, gt_tb_hm_res_3.P, gt_tb_hm_res_3.R, gt_tb_hm_res_3.F, \
|
|
mi@0
|
1013 gt_tp_hm_res_05.P, gt_tp_hm_res_05.R, gt_tp_hm_res_05.F, gt_tp_hm_res_3.P, gt_tp_hm_res_3.R, gt_tp_hm_res_3.F, tb_tp_hm_res_05.P, tb_tp_hm_res_05.R, tb_tp_hm_res_05.F, tb_tp_hm_res_3.P, tb_tp_hm_res_3.R, tb_tp_hm_res_3.F, \
|
|
mi@0
|
1014 gt_tb_tp_hm_res_05.P, gt_tb_tp_hm_res_05.R, gt_tb_tp_hm_res_05.F, gt_tb_tp_hm_res_3.P, gt_tb_tp_hm_res_3.R, gt_tb_tp_hm_res_3.F])
|
|
mi@0
|
1015
|
|
mi@0
|
1016
|
|
mi@0
|
1017 # Verification of detected boundaries by novelty fusion from the first round
|
|
mi@0
|
1018 # ao_featureset = [ao.gammatone_features, ao.harmonic_features, ao.timbre_features, ao.tempo_features]
|
|
mi@0
|
1019 # winlen = 1.5 * self.SampleRate / self.stepSize
|
|
mi@0
|
1020 # prev_features, post_features = self.getPeakFeatures(gt_tb_tp_hm_novelty_peaks, ao_featureset, winlen=10)
|
|
mi@0
|
1021 # dev_list = self.segmentDev(prev_features, post_features)
|
|
mi@0
|
1022 # gt_tb_tp_hm_novelty_peaks = gt_tb_tp_hm_novelty_peaks[:len(dev_list)]
|
|
mi@0
|
1023 # # print 'len(dev_list)', len(dev_list), len(gt_tb_tp_hm_novelty_peaks)
|
|
mi@0
|
1024 # # print gt_tb_tp_hm_novelty_peaks, dev_list
|
|
mi@0
|
1025 # dev_mean = [np.mean(x) for x in dev_list]
|
|
mi@0
|
1026 # np.savetxt(join(options.OUTPUT, 'dev', ao.name+'.csv'), np.vstack((gt_tb_tp_hm_detection[:len(dev_list)], dev_mean)).T, delimiter=',')
|
|
mi@0
|
1027 # peak_verified = self.verifyPeaks(gt_tb_tp_hm_novelty_peaks, dev_list)
|
|
mi@0
|
1028 #
|
|
mi@0
|
1029 # verified_detection = [ao.ssm_timestamps[int(i)] for i in peak_verified] + [ao.gt[-1]]
|
|
mi@0
|
1030 # verified_detection_05 = self.pairwiseF(ao.gt, verified_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1031 # verified_detection_3 = self.pairwiseF(ao.gt, verified_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1032 #
|
|
mi@0
|
1033 # print gt_tb_tp_hm_res_05.TP, gt_tb_tp_hm_res_05.FP, gt_tb_tp_hm_res_05.FN, gt_tb_tp_hm_res_05.P, gt_tb_tp_hm_res_05.R, gt_tb_tp_hm_res_05.F
|
|
mi@0
|
1034 # print gt_tb_tp_hm_res_3.TP, gt_tb_tp_hm_res_3.FP, gt_tb_tp_hm_res_3.FN, gt_tb_tp_hm_res_3.P, gt_tb_tp_hm_res_3.R, gt_tb_tp_hm_res_3.F
|
|
mi@0
|
1035 #
|
|
mi@0
|
1036 # print verified_detection_05.TP, verified_detection_05.FP, verified_detection_05.FN, verified_detection_05.P, verified_detection_05.R, verified_detection_05.F
|
|
mi@0
|
1037 # print verified_detection_3.TP, verified_detection_3.FP, verified_detection_3.FN, verified_detection_3.P, verified_detection_3.R, verified_detection_3.F
|
|
mi@0
|
1038
|
|
mi@0
|
1039 # if len(novelty_peaks):
|
|
mi@0
|
1040 # ao.gammatone_gmm = self.getGMMs(ao.gammatone_features, novelty_peaks)
|
|
mi@0
|
1041 # ao.harmonic_gmm = self.getGMMs(ao.harmonic_features, novelty_peaks)
|
|
mi@0
|
1042 # ao.tempo_gmm = self.getGMMs(ao.tempo_features, novelty_peaks)
|
|
mi@0
|
1043 # ao.timbre_gmm = self.getGMMs(ao.timbre_features, novelty_peaks)
|
|
mi@0
|
1044 #
|
|
mi@0
|
1045 # rc = rClustering(eps=1., k=8, rank='max_neighbors')
|
|
mi@0
|
1046 # rc.fit(ao.gammatone_gmm)
|
|
mi@0
|
1047 # gammatone_clf = rc.classification
|
|
mi@0
|
1048 # gammatone_neighborhood_size, gammatone_average_div, gammatone_node_rank = rc.getNodeRank()
|
|
mi@0
|
1049 # np.savetxt(join(options.OUTPUT, 'classification', ao.name+'-gammatone.csv'), np.vstack((novelty_detection[:-1], gammatone_clf)).T, delimiter=',')
|
|
mi@0
|
1050 # np.savetxt(join(options.OUTPUT, 'neighborhood_size', ao.name+'-gammatone.csv'), np.vstack((novelty_detection[:-1], gammatone_neighborhood_size)).T, delimiter=',')
|
|
mi@0
|
1051 # np.savetxt(join(options.OUTPUT, 'node_rank', ao.name+'-gammatone.csv'), np.vstack((novelty_detection[:-1], gammatone_average_div)).T, delimiter=',')
|
|
mi@0
|
1052 # np.savetxt(join(options.OUTPUT, 'average_div', ao.name+'-gammatone.csv'), np.vstack((novelty_detection[:-1], gammatone_node_rank)).T, delimiter=',')
|
|
mi@0
|
1053 #
|
|
mi@0
|
1054 # rc = rClustering(eps=1., k=8, rank='max_neighbors')
|
|
mi@0
|
1055 # rc.fit(ao.harmonic_gmm)
|
|
mi@0
|
1056 # harmonic_clf = rc.classification
|
|
mi@0
|
1057 # harmonic_neighborhood_size, harmonic_average_div, harmonic_node_rank = rc.getNodeRank()
|
|
mi@0
|
1058 # np.savetxt(join(options.OUTPUT, 'classification', ao.name+'-harmonic.csv'), np.vstack((novelty_detection[:-1], harmonic_clf)).T, delimiter=',')
|
|
mi@0
|
1059 # np.savetxt(join(options.OUTPUT, 'neighborhood_size', ao.name+'-harmonic.csv'), np.vstack((novelty_detection[:-1], harmonic_neighborhood_size)).T, delimiter=',')
|
|
mi@0
|
1060 # np.savetxt(join(options.OUTPUT, 'node_rank', ao.name+'-harmonic.csv'), np.vstack((novelty_detection[:-1], harmonic_average_div)).T, delimiter=',')
|
|
mi@0
|
1061 # np.savetxt(join(options.OUTPUT, 'average_div', ao.name+'-harmonic.csv'), np.vstack((novelty_detection[:-1], harmonic_node_rank)).T, delimiter=',')
|
|
mi@0
|
1062 #
|
|
mi@0
|
1063 # rc = rClustering(eps=1., k=8, rank='max_neighbors')
|
|
mi@0
|
1064 # rc.fit(ao.tempo_gmm)
|
|
mi@0
|
1065 # tempo_clf = rc.classification
|
|
mi@0
|
1066 # tempo_neighborhood_size, tempo_average_div, tempo_node_rank = rc.getNodeRank()
|
|
mi@0
|
1067 # np.savetxt(join(options.OUTPUT, 'classification', ao.name+'-tempo.csv'), np.vstack((novelty_detection[:-1], tempo_clf)).T, delimiter=',')
|
|
mi@0
|
1068 # np.savetxt(join(options.OUTPUT, 'neighborhood_size', ao.name+'-tempo.csv'), np.vstack((novelty_detection[:-1], tempo_neighborhood_size)).T, delimiter=',')
|
|
mi@0
|
1069 # np.savetxt(join(options.OUTPUT, 'node_rank', ao.name+'-tempo.csv'), np.vstack((novelty_detection[:-1], tempo_average_div)).T, delimiter=',')
|
|
mi@0
|
1070 # np.savetxt(join(options.OUTPUT, 'average_div', ao.name+'-tempo.csv'), np.vstack((novelty_detection[:-1], tempo_node_rank)).T, delimiter=',')
|
|
mi@0
|
1071 #
|
|
mi@0
|
1072 # rc = rClustering(eps=1., k=8, rank='max_neighbors')
|
|
mi@0
|
1073 # rc.fit(ao.timbre_gmm)
|
|
mi@0
|
1074 # timbre_clf = rc.classification
|
|
mi@0
|
1075 # timbre_neighborhood_size, timbre_average_div, timbre_node_rank = rc.getNodeRank()
|
|
mi@0
|
1076 # np.savetxt(join(options.OUTPUT, 'classification', ao.name+'-timbre.csv'), np.vstack((novelty_detection[:-1], timbre_clf)).T, delimiter=',')
|
|
mi@0
|
1077 # np.savetxt(join(options.OUTPUT, 'neighborhood_size', ao.name+'-timbre.csv'), np.vstack((novelty_detection[:-1], timbre_neighborhood_size)).T, delimiter=',')
|
|
mi@0
|
1078 # np.savetxt(join(options.OUTPUT, 'node_rank', ao.name+'-timbre.csv'), np.vstack((novelty_detection[:-1], timbre_average_div)).T, delimiter=',')
|
|
mi@0
|
1079 # np.savetxt(join(options.OUTPUT, 'average_div', ao.name+'-timbre.csv'), np.vstack((novelty_detection[:-1], timbre_node_rank)).T, delimiter=',')
|
|
mi@0
|
1080
|
|
mi@0
|
1081
|
|
mi@0
|
1082 # # Evaluate segmentation results using combined SSMs.
|
|
mi@0
|
1083 # outfile3 = join(options.OUTPUT, 'combinedSSMRes.csv')
|
|
mi@0
|
1084 # with open(outfile3, 'a') as f:
|
|
mi@0
|
1085 # csvwriter = csv.writer(f, delimiter=',')
|
|
mi@0
|
1086 # csvwriter.writerow(['audio', 'gt_tb_P_0.5', 'gt_tb_R_0.5', 'gt_tb_F_0.5', 'gt_tb_P_3', 'gt_tb_R_3', 'gt_tb_F_3', 'gt_tp_P_0.5', 'gt_tp_R_0.5', 'gt_tp_F_0.5', 'gt_tp_P_3', 'gt_tp_R_3', 'gt_tp_F_3',\
|
|
mi@0
|
1087 # 'gt_hm_P_0.5', 'gt_hm_R_0.5', 'gt_hm_F_0.5', 'gt_hm_P_3', 'gt_hm_R_3', 'gt_hm_F_3', 'tb_tp_P_0.5', 'tb_tp_R_0.5', 'tb_tp_F_0.5', 'tb_tp_P_3', 'tb_tp_R_3', 'tb_tp_F_3', 'tb_hm_P_0.5', 'tb_hm_R_0.5', 'tb_hm_F_0.5', \
|
|
mi@0
|
1088 # 'tb_hm_P_3', 'tb_hm_R_3', 'tb_hm_F_3', 'tp_hm_P_0.5', 'tp_hm_R_0.5', 'tp_hm_F_0.5', 'tp_hm_P_3', 'tp_hm_R_3', 'tp_hm_F_3', 'gt_tb_tp_P_0.5', 'gt_tb_tp_R_0.5', 'gt_tb_tp_F_0.5', 'gt_tb_tp_P_3', 'gt_tb_tp_R_3', \
|
|
mi@0
|
1089 # 'gt_tb_tp_F_3', 'gt_tb_hm_P_0.5', 'gt_tb_hm_R_0.5', 'gt_tb_hm_F_0.5', 'gt_tb_hm_P_3', 'gt_tb_hm_R_3', 'gt_tb_hm_F_3', 'gt_tp_hm_P_0.5', 'gt_tp_hm_R_0.5', 'gt_tp_hm_F_0.5', 'gt_tp_hm_P_3', 'gt_tp_hm_R_3', 'gt_tp_hm_F_3', \
|
|
mi@0
|
1090 # 'tb_tp_hm_P_0.5', 'tb_tp_hm_R_0.5', 'tb_tp_hm_F_0.5', 'tb_tp_hm_P_3', 'tb_tp_hm_R_3', 'tb_tp_hm_F_3', 'gt_tb_tp_hm_P_0.5', 'gt_tb_tp_hm_R_0.5', 'gt_tb_tp_hm_F_0.5', 'gt_tb_tp_hm_P_3', 'gt_tb_tp_hm_R_3', 'gt_tb_tp_hm_F_3'])
|
|
mi@0
|
1091 #
|
|
mi@0
|
1092 # for i,ao in enumerate(audio_list):
|
|
mi@0
|
1093 # # Combine SSMs computed from different features
|
|
mi@0
|
1094 # gt_hm_ssm = np.multiply(ao.gammatone_ssm, ao.harmonic_ssm)
|
|
mi@0
|
1095 # gt_tb_ssm = np.multiply(ao.gammatone_ssm, ao.timbre_ssm)
|
|
mi@0
|
1096 # gt_tp_ssm = np.multiply(ao.gammatone_ssm, ao.tempo_ssm)
|
|
mi@0
|
1097 # tb_tp_ssm = np.multiply(ao.timbre_ssm, ao.tempo_ssm)
|
|
mi@0
|
1098 # tb_hm_ssm = np.multiply(ao.timbre_ssm, ao.harmonic_ssm)
|
|
mi@0
|
1099 # tp_hm_ssm = np.multiply(ao.tempo_ssm, ao.harmonic_ssm)
|
|
mi@0
|
1100 #
|
|
mi@0
|
1101 # gt_hm_tb_ssm = np.multiply(ao.gammatone_ssm, ao.harmonic_ssm, ao.timbre_ssm)
|
|
mi@0
|
1102 # gt_hm_tp_ssm = np.multiply(ao.gammatone_ssm, ao.harmonic_ssm, ao.tempo_ssm)
|
|
mi@0
|
1103 # gt_tb_tp_ssm = np.multiply(ao.gammatone_ssm, ao.timbre_ssm, ao.tempo_ssm)
|
|
mi@0
|
1104 # hm_tb_tp_ssm = np.multiply(ao.harmonic_ssm, ao.timbre_ssm, ao.tempo_ssm)
|
|
mi@0
|
1105 #
|
|
mi@0
|
1106 # gt_hm_tb_tp_ssm = np.multiply(np.multiply(ao.gammatone_ssm, ao.harmonic_ssm), np.multiply(ao.timbre_ssm, ao.tempo_ssm))
|
|
mi@0
|
1107 #
|
|
mi@0
|
1108 # gt_hm_ssm_novelty = self.getNoveltyCurve(gt_hm_ssm, self.kernel_size)
|
|
mi@0
|
1109 # gt_hm_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_hm_ssm_novelty]
|
|
mi@0
|
1110 # gt_tb_ssm_novelty = self.getNoveltyCurve(gt_tb_ssm, self.kernel_size)
|
|
mi@0
|
1111 # gt_tb_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_tb_ssm_novelty]
|
|
mi@0
|
1112 # gt_tp_ssm_novelty = self.getNoveltyCurve(gt_hm_ssm, self.kernel_size)
|
|
mi@0
|
1113 # gt_tp_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_tp_ssm_novelty]
|
|
mi@0
|
1114 # tb_tp_ssm_novelty = self.getNoveltyCurve(tb_tp_ssm, self.kernel_size)
|
|
mi@0
|
1115 # tb_tp_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in tb_tp_ssm_novelty]
|
|
mi@0
|
1116 # tb_hm_ssm_novelty = self.getNoveltyCurve(tb_hm_ssm, self.kernel_size)
|
|
mi@0
|
1117 # tb_hm_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in tb_hm_ssm_novelty]
|
|
mi@0
|
1118 # tp_hm_ssm_novelty = self.getNoveltyCurve(tp_hm_ssm, self.kernel_size)
|
|
mi@0
|
1119 # tp_hm_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in tp_hm_ssm_novelty]
|
|
mi@0
|
1120 #
|
|
mi@0
|
1121 # gt_hm_tb_ssm_novelty = self.getNoveltyCurve(gt_hm_tb_ssm, self.kernel_size)
|
|
mi@0
|
1122 # gt_hm_tb_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_hm_tb_ssm_novelty]
|
|
mi@0
|
1123 # gt_hm_tp_ssm_novelty = self.getNoveltyCurve(gt_hm_tp_ssm, self.kernel_size)
|
|
mi@0
|
1124 # gt_hm_tp_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_hm_tp_ssm_novelty]
|
|
mi@0
|
1125 # gt_tb_tp_ssm_novelty = self.getNoveltyCurve(gt_tb_tp_ssm, self.kernel_size)
|
|
mi@0
|
1126 # gt_tb_tp_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_tb_tp_ssm_novelty]
|
|
mi@0
|
1127 # hm_tb_tp_ssm_novelty = self.getNoveltyCurve(hm_tb_tp_ssm, self.kernel_size)
|
|
mi@0
|
1128 # hm_tb_tp_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in hm_tb_tp_ssm_novelty]
|
|
mi@0
|
1129 #
|
|
mi@0
|
1130 # gt_hm_tb_tp_ssm_novelty = self.getNoveltyCurve(gt_hm_tb_tp_ssm, self.kernel_size)
|
|
mi@0
|
1131 # gt_hm_tb_tp_ssm_novelty = [0.0 if (np.isnan(x) or np.isinf(x) or x > 1e+100) else x for x in gt_hm_tb_tp_ssm_novelty]
|
|
mi@0
|
1132 #
|
|
mi@0
|
1133 # smoothed_gt_hm_ssm_novelty, gt_hm_ssm_novelty_peaks = peak_picker.process(gt_hm_ssm_novelty)
|
|
mi@0
|
1134 # gt_hm_ssm_detection = [ao.ssm_timestamps[int(i)] for i in gt_hm_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1135 # smoothed_gt_tb_ssm_novelty, gt_tb_ssm_novelty_peaks = peak_picker.process(gt_tb_ssm_novelty)
|
|
mi@0
|
1136 # gt_tb_ssm_detection = [ao.ssm_timestamps[int(i)] for i in gt_tb_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1137 # smoothed_gt_tp_ssm_novelty, gt_tp_ssm_novelty_peaks = peak_picker.process(gt_tp_ssm_novelty)
|
|
mi@0
|
1138 # gt_tp_ssm_detection = [ao.ssm_timestamps[int(i)] for i in gt_tp_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1139 # smoothed_tb_tp_ssm_novelty, tb_tp_ssm_novelty_peaks = peak_picker.process(tb_tp_ssm_novelty)
|
|
mi@0
|
1140 # tb_tp_ssm_detection = [ao.ssm_timestamps[int(i)] for i in tb_tp_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1141 # smoothed_tb_hm_ssm_novelty, tb_hm_ssm_novelty_peaks = peak_picker.process(tb_hm_ssm_novelty)
|
|
mi@0
|
1142 # tb_hm_ssm_detection = [ao.ssm_timestamps[int(i)] for i in tb_hm_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1143 # smoothed_tp_hm_ssm_novelty, tp_hm_ssm_novelty_peaks = peak_picker.process(tp_hm_ssm_novelty)
|
|
mi@0
|
1144 # tp_hm_ssm_detection = [ao.ssm_timestamps[int(i)] for i in tp_hm_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1145 #
|
|
mi@0
|
1146 # smoothed_gt_hm_tb_ssm_novelty, gt_hm_tb_ssm_novelty_peaks = peak_picker.process(gt_hm_tb_ssm_novelty)
|
|
mi@0
|
1147 # gt_hm_tb_ssm_detection = [ao.ssm_timestamps[int(i)] for i in gt_hm_tb_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1148 # smoothed_gt_hm_tp_ssm_novelty, gt_hm_tp_ssm_novelty_peaks = peak_picker.process(gt_hm_tp_ssm_novelty)
|
|
mi@0
|
1149 # gt_hm_tp_ssm_detection = [ao.ssm_timestamps[int(i)] for i in gt_hm_tp_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1150 # smoothed_gt_tb_tp_ssm_novelty, gt_tb_tp_ssm_novelty_peaks = peak_picker.process(gt_tb_tp_ssm_novelty)
|
|
mi@0
|
1151 # gt_tb_tp_ssm_detection = [ao.ssm_timestamps[int(i)] for i in gt_tb_tp_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1152 # smoothed_hm_tb_tp_ssm_novelty, hm_tb_tp_ssm_novelty_peaks = peak_picker.process(hm_tb_tp_ssm_novelty)
|
|
mi@0
|
1153 # hm_tb_tp_ssm_detection = [ao.ssm_timestamps[int(i)] for i in hm_tb_tp_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1154 #
|
|
mi@0
|
1155 # smoothed_gt_hm_tb_tp_ssm_novelty, gt_hm_tb_tp_ssm_novelty_peaks = peak_picker.process(gt_hm_tb_tp_ssm_novelty)
|
|
mi@0
|
1156 # gt_hm_tb_tp_ssm_detection = [ao.ssm_timestamps[int(i)] for i in gt_hm_tb_tp_ssm_novelty_peaks] + [ao.gt[-1]]
|
|
mi@0
|
1157 #
|
|
mi@0
|
1158 # # Output detected segment locations.
|
|
mi@0
|
1159 # if options.VERBOSE:
|
|
mi@0
|
1160 # outdir = join(options.OUTPUT, 'detection', ao.name)
|
|
mi@0
|
1161 # if not isdir(outdir):
|
|
mi@0
|
1162 # os.mkdir(outdir)
|
|
mi@0
|
1163 #
|
|
mi@0
|
1164 # np.savetxt(join(outdir, 'gammatone_timbre_ssm.csv'), gt_tb_ssm_detection)
|
|
mi@0
|
1165 # np.savetxt(join(outdir, 'gammatone_tempo_ssm.csv'), gt_tp_ssm_detection)
|
|
mi@0
|
1166 # np.savetxt(join(outdir, 'gammatone_harmonic_ssm.csv'), gt_hm_ssm_detection)
|
|
mi@0
|
1167 # np.savetxt(join(outdir, 'timbre_tempo_ssm.csv'), tb_tp_ssm_detection)
|
|
mi@0
|
1168 # np.savetxt(join(outdir, 'timbre_harmonic_ssm.csv'), tb_hm_ssm_detection)
|
|
mi@0
|
1169 # np.savetxt(join(outdir, 'tempo_harmonic_ssm.csv'), tp_hm_ssm_detection)
|
|
mi@0
|
1170 #
|
|
mi@0
|
1171 # np.savetxt(join(outdir, 'gammatone_timbre_tempo_ssm.csv'), gt_tb_tp_ssm_detection)
|
|
mi@0
|
1172 # np.savetxt(join(outdir, 'gammatone_timbre_harmonic_ssm.csv'), gt_hm_tb_ssm_detection)
|
|
mi@0
|
1173 # np.savetxt(join(outdir, 'gammatone_tempo_harmonic_ssm.csv'), gt_hm_tp_ssm_detection)
|
|
mi@0
|
1174 # np.savetxt(join(outdir, 'timbre_tempo_harmonic_ssm.csv'), hm_tb_tp_ssm_detection)
|
|
mi@0
|
1175 # np.savetxt(join(outdir, 'gammatone_timbre_tempo_harmonic_ssm.csv'), gt_hm_tb_tp_ssm_detection)
|
|
mi@0
|
1176 #
|
|
mi@0
|
1177 # gt_hm_ssm_res_05 = self.pairwiseF(ao.gt, gt_hm_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1178 # gt_hm_ssm_res_3 = self.pairwiseF(ao.gt, gt_hm_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1179 # gt_tb_ssm_res_05 = self.pairwiseF(ao.gt, gt_tb_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1180 # gt_tb_ssm_res_3 = self.pairwiseF(ao.gt, gt_tb_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1181 # gt_tp_ssm_res_05 = self.pairwiseF(ao.gt, gt_tp_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1182 # gt_tp_ssm_res_3 = self.pairwiseF(ao.gt, gt_tp_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1183 # tb_tp_ssm_res_05 = self.pairwiseF(ao.gt, tb_tp_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1184 # tb_tp_ssm_res_3 = self.pairwiseF(ao.gt, tb_tp_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1185 # tb_hm_ssm_res_05 = self.pairwiseF(ao.gt, tb_hm_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1186 # tb_hm_ssm_res_3 = self.pairwiseF(ao.gt, tb_hm_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1187 # tp_hm_ssm_res_05 = self.pairwiseF(ao.gt, tp_hm_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1188 # tp_hm_ssm_res_3 = self.pairwiseF(ao.gt, tp_hm_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1189 #
|
|
mi@0
|
1190 # gt_hm_tb_ssm_res_05 = self.pairwiseF(ao.gt, gt_hm_tb_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1191 # gt_hm_tb_ssm_res_3 = self.pairwiseF(ao.gt, gt_hm_tb_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1192 # gt_hm_tp_ssm_res_05 = self.pairwiseF(ao.gt, gt_hm_tp_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1193 # gt_hm_tp_ssm_res_3 = self.pairwiseF(ao.gt, gt_hm_tp_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1194 # gt_tb_tp_ssm_res_05 = self.pairwiseF(ao.gt, gt_tb_tp_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1195 # gt_tb_tp_ssm_res_3 = self.pairwiseF(ao.gt, gt_tb_tp_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1196 # hm_tb_tp_ssm_res_05 = self.pairwiseF(ao.gt, hm_tb_tp_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1197 # hm_tb_tp_ssm_res_3 = self.pairwiseF(ao.gt, hm_tb_tp_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1198 #
|
|
mi@0
|
1199 # gt_hm_tb_tp_ssm_res_05 = self.pairwiseF(ao.gt, gt_hm_tb_tp_ssm_detection, tolerance=0.5, combine=1.0)
|
|
mi@0
|
1200 # gt_hm_tb_tp_ssm_res_3 = self.pairwiseF(ao.gt, gt_hm_tb_tp_ssm_detection, tolerance=3, combine=1.0)
|
|
mi@0
|
1201 #
|
|
mi@0
|
1202 # with open(outfile3, 'a') as f:
|
|
mi@0
|
1203 # csvwriter = csv.writer(f, delimiter=',')
|
|
mi@0
|
1204 # csvwriter.writerow([ao.name, gt_tb_ssm_res_05.P, gt_tb_ssm_res_05.R, gt_tb_ssm_res_05.F, gt_tb_ssm_res_3.P, gt_tb_ssm_res_3.R, gt_tb_ssm_res_3.F, gt_tp_ssm_res_05.P, gt_tp_ssm_res_05.R, gt_tp_ssm_res_05.F, \
|
|
mi@0
|
1205 # gt_tp_ssm_res_3.P, gt_tp_ssm_res_3.R, gt_tp_ssm_res_3.F, gt_hm_ssm_res_05.P, gt_hm_ssm_res_05.R, gt_hm_ssm_res_05.F, gt_hm_ssm_res_3.P, gt_hm_ssm_res_3.R, gt_hm_ssm_res_3.F, \
|
|
mi@0
|
1206 # tb_tp_ssm_res_05.P, tb_tp_ssm_res_05.R, tb_tp_ssm_res_05.F, tb_tp_ssm_res_3.P, tb_tp_ssm_res_3.R, tb_tp_ssm_res_3.F, tb_hm_ssm_res_05.P, tb_hm_ssm_res_05.R, tb_hm_ssm_res_05.F, \
|
|
mi@0
|
1207 # tb_hm_ssm_res_3.P, tb_hm_ssm_res_3.R, tb_hm_ssm_res_3.F, tp_hm_ssm_res_05.P, tp_hm_ssm_res_05.R, tp_hm_ssm_res_05.F, tp_hm_ssm_res_3.P, tp_hm_ssm_res_3.R, tp_hm_ssm_res_3.F, \
|
|
mi@0
|
1208 # gt_tb_tp_ssm_res_05.P, gt_tb_tp_ssm_res_05.R, gt_tb_tp_ssm_res_05.F, gt_tb_tp_ssm_res_3.P, gt_tb_tp_ssm_res_3.R, gt_tb_tp_ssm_res_3.F, gt_hm_tb_ssm_res_05.P, gt_hm_tb_ssm_res_05.R, gt_hm_tb_ssm_res_05.F, \
|
|
mi@0
|
1209 # gt_hm_tb_ssm_res_3.P, gt_hm_tb_ssm_res_3.R, gt_hm_tb_ssm_res_3.F, gt_hm_tp_ssm_res_05.P, gt_hm_tp_ssm_res_05.R, gt_hm_tp_ssm_res_05.F, gt_hm_tp_ssm_res_3.P, gt_hm_tp_ssm_res_3.R, gt_hm_tp_ssm_res_3.F, \
|
|
mi@0
|
1210 # hm_tb_tp_ssm_res_05.P, hm_tb_tp_ssm_res_05.R, hm_tb_tp_ssm_res_05.F, hm_tb_tp_ssm_res_3.P, hm_tb_tp_ssm_res_3.R, hm_tb_tp_ssm_res_3.F, gt_hm_tb_tp_ssm_res_05.P, gt_hm_tb_tp_ssm_res_05.R, gt_hm_tb_tp_ssm_res_05.F, \
|
|
mi@0
|
1211 # gt_hm_tb_tp_ssm_res_3.P, gt_hm_tb_tp_ssm_res_3.R, gt_hm_tb_tp_ssm_res_3.F])
|
|
mi@0
|
1212
|
|
mi@0
|
1213
|
|
mi@0
|
1214 def main():
|
|
mi@0
|
1215 segmenter = SSMseg()
|
|
mi@0
|
1216 segmenter.process()
|
|
mi@0
|
1217
|
|
mi@0
|
1218
|
|
mi@0
|
1219 if __name__ == '__main__':
|
|
mi@0
|
1220 main()
|
|
mi@0
|
1221
|
