def __init__(self, wavfolder, trainingdata):
"""Initialise the classifier and train it on some WAV files.
'wavfolder' is the base folder, to be prepended to all WAV paths.
'trainingdata' is a dictionary of wavpath:label pairs."""
self.mfccMaker = melScaling(int(fs), framelen / 2, 40)
self.mfccMaker.update()
self.count = 0
allfeatures = {
wavpath: self.file_to_features(os.path.join(wavfolder, wavpath))
for wavpath in trainingdata
}
#print('ALLfeatures:',allfeatures)####checking features###
np.set_printoptions(threshold=np.nan)
foobar = open("all_features_together_dataset_13.txt", "w")
for label, aggf in allfeatures.items():
foobar.write("\"" + label + "\"")
foobar.write(str(aggf))
foobar.close()
# Determine the normalisation stats, and remember them
allconcat = np.vstack(list(allfeatures.values(
))) #stack arrays in sequence vertically(row wise))
self.means = np.mean(allconcat, 0) #compute mean
self.invstds = np.std(allconcat, 0) #compute the standard deviation
for i, val in enumerate(self.invstds):
if val == 0.0:
self.invstds[i] = 1.0
else:
self.invstds[i] = 1.0 / val
# For each label, compile a normalised concatenated list of features
aggfeatures = {}
for wavpath, features in allfeatures.items():
label = trainingdata[wavpath]
normed = self.__normalise(features)
if label not in aggfeatures:
aggfeatures[label] = normed
else:
aggfeatures[label] = np.vstack((aggfeatures[label], normed))
# For each label's aggregated features, train a GMM and remember it
self.gmms = {}
for label, aggf in aggfeatures.items():
if verbose:
print(
" Training a GMM for label %s, using data of shape %s" %
(label, str(np.shape(aggf))))
self.gmms[label] = GMM(n_components=2) # , cvtype='full')
self.gmms[label].fit(aggf)
if verbose:
print(" Trained %i classes from %i input files" %
(len(self.gmms), len(trainingdata)))
def __init__(self, wavfolder, trainingdata):
"""Initialise the classifier and train it on some WAV files.
'wavfolder' is the base folder, to be prepended to all WAV paths.
'trainingdata' is a dictionary of wavpath:label pairs."""
self.mfccMaker = melScaling(int(fs), framelen / 2, 40)
self.mfccMaker.update()
allfeatures = {
wavpath: self.file_to_features(os.path.join(wavfolder, wavpath))
for wavpath in trainingdata
}
# Determine the normalisation stats, and remember them
allconcat = np.vstack(list(allfeatures.values()))
self.means = np.mean(allconcat, 0)
self.invstds = np.std(allconcat, 0)
for i, val in enumerate(self.invstds):
if val == 0.0:
self.invstds[i] = 1.0
else:
self.invstds[i] = 1.0 / val
# For each label, compile a normalised concatenated list of features
aggfeatures = {}
for wavpath, features in allfeatures.items():
label = trainingdata[wavpath]
normed = self.__normalise(features)
if label not in aggfeatures:
aggfeatures[label] = normed
else:
aggfeatures[label] = np.vstack((aggfeatures[label], normed))
# For each label's aggregated features, train a GMM and remember it
self.GaussianMixtures = {}
for label, aggf in aggfeatures.items():
if verbose:
print(
" Training a GMM for label %s, using data of shape %s" %
(label, str(np.shape(aggf))))
self.GaussianMixtures[label] = GaussianMixture(
n_components=10) # , cvtype='full')
self.GaussianMixtures[label].fit(aggf)
if verbose:
print(" Trained %i classes from %i input files" %
(len(self.GaussianMixtures), len(trainingdata)))
def __init__(self, wavfolder, trainingdata):
"""Initialise the classifier and train it on some WAV files.
'wavfolder' is the base folder, to be prepended to all WAV paths.
'trainingdata' is a dictionary of wavpath:label pairs."""
self.mfccMaker = melScaling(int(fs), framelen/2, 40)
self.mfccMaker.update()
allfeatures = {wavpath:self.file_to_features(os.path.join(wavfolder, wavpath)) for wavpath in trainingdata}
###print('ALLfeatures:',allfeatures)####checking features###
# Determine the normalisation stats, and remember them
allconcat = np.vstack(list(allfeatures.values()))#stack arrays in sequence vertically(row wise))
self.means = np.mean(allconcat, 0)#compute mean
self.invstds = np.std(allconcat, 0)#compute the standard deviation
for i,val in enumerate(self.invstds):
if val == 0.0:
self.invstds[i] = 1.0
else:
self.invstds[i] = 1.0 / val
# For each label, compile a normalised concatenated list of features
aggfeatures = {}
for wavpath, features in allfeatures.items():
label = trainingdata[wavpath]
normed = self.__normalise(features)
if label not in aggfeatures:
aggfeatures[label] = normed
else:
aggfeatures[label] = np.vstack((aggfeatures[label], normed))
# For each label's aggregated features, train a GMM and remember it
self.gmms = {}
for label, aggf in aggfeatures.items():
if verbose: print(" Training a GMM for label %s, using data of shape %s" % (label, str(np.shape(aggf))))
self.gmms[label] = GMM(n_components=10) # , cvtype='full')
self.gmms[label].fit(aggf)
if verbose: print(" Trained %i classes from %i input files" % (len(self.gmms), len(trainingdata)))
def __init__(self):
self.mfccMaker = melScaling(int(fs), framelen / 2, 40)
self.mfccMaker.update()
def __init__(self):
self.mfccMaker = melScaling(int(fs), framelen/2, 40)
self.mfccMaker.update()
def __init__(self, wavfolder, trainingdata):
"""Initialise the classifier and train it on some WAV files.
'wavfolder' is the base folder, to be prepended to all WAV paths.
'trainingdata' is a dictionary of wavpath:label pairs."""
self.mfccMaker = melScaling(int(fs), framelen / 2, 40)
self.mfccMaker.update()
self.target = {}
self.labelarr = []
self.labelarr2 = {}
self.labelarr3 = {}
self.target2 = []
i = 0
for wavpath in trainingdata:
label = trainingdata[wavpath]
if label not in self.labelarr2.values():
print " target classes :" + label
self.labelarr.append(i)
self.labelarr2[i] = label
self.labelarr3[label] = i
i = i + 1
else:
self.labelarr.append(self.labelarr3[label])
allfeatures = {
wavpath: self.file_to_features(os.path.join(wavfolder, wavpath), trainingdata[wavpath])
for wavpath in trainingdata
}
# print("**********features :",self.file_to_features(os.path.join(wavfolder,wavpath)))
print ("allfeatures :" + str(np.shape(allfeatures.values())))
# print("target :",self.target2)
print ("target values", self.labelarr)
print ("target labels", self.labelarr2)
###print('ALLfeatures:',allfeatures)####checking features###
# Determine the normalisation stats, and remember them
allconcat = np.vstack(list(allfeatures.values())) # stack arrays in sequence vertically(row wise))
self.means = np.mean(allconcat, 0) # compute mean
self.invstds = np.std(allconcat, 0) # compute the standard deviation
for i, val in enumerate(self.invstds):
if val == 0.0:
self.invstds[i] = 1.0
else:
self.invstds[i] = 1.0 / val
# For each label, compile a normalised concatenated list of features
aggfeatures = {}
for wavpath, features in allfeatures.items():
label = trainingdata[wavpath]
normed = self.__normalise(features)
if label not in aggfeatures:
aggfeatures[label] = normed
else:
aggfeatures[label] = np.vstack((aggfeatures[label], normed))
# For each label's aggregated features, train a SVM and remember it
# self.svms = {}
self.X = []
for label, aggf in aggfeatures.items():
if verbose:
print (" Training a SVM for label %s, using data of shape %s" % (label, str(np.shape(aggf))))
# self.svms[label] = svm.SVC()
# print("aggf : ",aggf)
if not np.any(self.X):
self.X = aggf
else:
# print("X: ",np.shape(self.X))
# print("aggf: ",np.shape(aggf))
self.X = np.vstack((self.X, aggf))
# self.svms[label].fit(aggf,self.target[label])
# print("X :",self.X)
self.clf = svm.SVC()
# print("target : ",self.target2)
self.clf.fit(self.X, self.target2)
def __init__(self, wavfolder, trainingdata):
"""Initialise the classifier and train it on some WAV files.
'wavfolder' is the base folder, to be prepended to all WAV paths.
'trainingdata' is a dictionary of wavpath:label pairs."""
#open audio file as an AudioFile object
for wavpath in trainingdata:
audio_file = open((os.path.join(wavfolder, wavpath)))
file_info=self.get_info(audio_file)
#Creates a WaveReader object from the AudioFile Object
pcm_data = audio_file.to_pcm()
#Creates a FrameList object from WaveReader object. Currently reads all
#frames in file
frame_list = pcm_data.read(file_info["frames"])
#Convert samples to floats (-1.0 - +1.0)
float_frame_list = frame_list.to_float()
#eventually do some signal processing here...
#Convert back to integer FrameList
output_framelist = float_frame_list.to_int(file_info["bits"])
#now back to raw bytes
output_data = output_framelist.to_bytes(False, True)
self.mfccMaker = melScaling(int(fs), framelen/2, 40)
self.mfccMaker.update()
self.target={}
self.labelarr=[]
self.labelarr2={}
self.labelarr3={}
self.target2=[]
i=0
for wavpath in trainingdata:
label = trainingdata[wavpath]
if label not in self.labelarr2.values():
print " target classes :"+label
self.labelarr.append(i)
self.labelarr2[i]=label
self.labelarr3[label]=i
i=i+1
else:
self.labelarr.append(self.labelarr3[label])
allfeatures = {wavpath:self.file_to_features(os.path.join(wavfolder, wavpath),trainingdata[wavpath]) for wavpath in trainingdata}
#print("**********features :",self.file_to_features(os.path.join(wavfolder,wavpath)))
print("allfeatures :"+str(np.shape(allfeatures.values())))
#print("target :",self.target2)
print('target values', self.labelarr)
print('target labels', self.labelarr2)
###print('ALLfeatures:',allfeatures)####checking features###
# Determine the normalisation stats, and remember them
allconcat = np.vstack(list(allfeatures.values()))#stack arrays in sequence vertically(row wise))
self.means = np.mean(allconcat, 0)#compute mean
self.invstds = np.std(allconcat, 0)#compute the standard deviation
for i,val in enumerate(self.invstds):
if val == 0.0:
self.invstds[i] = 1.0
else:
self.invstds[i] = 1.0 / val
# For each label, compile a normalised concatenated list of features
aggfeatures = {}
for wavpath, features in allfeatures.items():
label = trainingdata[wavpath]
normed = self.__normalise(features)
if label not in aggfeatures:
aggfeatures[label] = normed
else:
aggfeatures[label] = np.vstack((aggfeatures[label], normed))
# For each label's aggregated features, train a SVM and remember it
#self.svms = {}
self.X=[]
for label, aggf in aggfeatures.items():
if verbose: print(" Training a SVM for label %s, using data of shape %s" % (label, str(np.shape(aggf))))
#self.svms[label] = svm.SVC()
#print("aggf : ",aggf)
if not np.any(self.X):
self.X=aggf
else:
#print("X: ",np.shape(self.X))
#print("aggf: ",np.shape(aggf))
self.X=np.vstack((self.X,aggf))
#self.svms[label].fit(aggf,self.target[label])
#print("X :",self.X)
self.clf=svm.SVC()
#print("target : ",self.target2)
self.clf.fit(self.X,self.target2)
def __init__(self, wavfolder, trainingdata):
"""Initialise the classifier and train it on some WAV files.
'wavfolder' is the base folder, to be prepended to all WAV paths.
'trainingdata' is a dictionary of wavpath:label pairs."""
#open audio file as an AudioFile object
for wavpath in trainingdata:
audio_file = open((os.path.join(wavfolder, wavpath)))
file_info=self.get_info(audio_file)
#Creates a WaveReader object from the AudioFile Object
pcm_data = audio_file.to_pcm()
#Creates a FrameList object from WaveReader object. Currently reads all
#frames in file
frame_list = pcm_data.read(file_info["frames"])
#Convert samples to floats (-1.0 - +1.0)
float_frame_list = frame_list.to_float()
#eventually do some signal processing here...
#Convert back to integer FrameList
output_framelist = float_frame_list.to_int(file_info["bits"])
#now back to raw bytes
output_data = output_framelist.to_bytes(False, True)
self.mfccMaker = melScaling(int(fs), framelen/2, 40)
self.mfccMaker.update()
allfeatures = {wavpath:self.file_to_features(os.path.join(wavfolder, wavpath)) for wavpath in trainingdata}
###print('ALLfeatures:',allfeatures)####checking features###
# Determine the normalisation stats, and remember them
allconcat = np.vstack(list(allfeatures.values()))#stack arrays in sequence vertically(row wise))
self.means = np.mean(allconcat, 0)#compute mean
self.invstds = np.std(allconcat, 0)#compute the standard deviation
for i,val in enumerate(self.invstds):
if val == 0.0:
self.invstds[i] = 1.0
else:
self.invstds[i] = 1.0 / val
# For each label, compile a normalised concatenated list of features
aggfeatures = {}
for wavpath, features in allfeatures.items():
label = trainingdata[wavpath]
normed = self.__normalise(features)
if label not in aggfeatures:
aggfeatures[label] = normed
else:
aggfeatures[label] = np.vstack((aggfeatures[label], normed))
# For each label's aggregated features, train a GMM and remember it
self.gmms = {}
for label, aggf in aggfeatures.items():
if verbose: print(" Training a GMM for label %s, using data of shape %s" % (label, str(np.shape(aggf))))
self.gmms[label] = GMM(n_components=10) # , cvtype='full')
self.gmms[label].fit(aggf)
if verbose: print(" Trained %i classes from %i input files" % (len(self.gmms), len(trainingdata)))