def main():
lda = LDA(solver='svd')
log.basicConfig(
level=args.debug, format='%(asctime)s %(levelname)s %(message)s', datefmt='%d/%m %H:%M:%S')
# Will be filled as dict
testreferences = None
if args.mlf:
log.info("Parsing input label file %s as mlf file" % (args.testref))
testreferences = mlffile(args.testref)
else:
log.info(
"Parsing input label file %s as test_ref file( TARGETMDL ENROLEMODEL-TESTUTT LABEL )" % (args.testref))
testreferences = test_ref(args.testref)
# Check if the given data is in marshal format or cPickle
if args.binary:
log.info("Try to read input as a binary file")
inputdata, testutts = checkBinary([args.inputdata, args.testutts])
datadim = len(inputdata.values()[0])
dvectors = np.zeros((len(inputdata.keys()), datadim))
labels = []
log.info("Getting dvectors from input data")
for i, (spk, v) in enumerate(inputdata.iteritems()):
dvectors[i] = v
labels.append(getspkmodel(spk, args.delimiter, args.indices))
log.debug("Data which was loaded in (%s) has %i labels and %i dvectors" %
(args.inputdata, len(labels), len(dvectors)))
testtofeature = testutts
else:
log.info("Given data is either a folder or a filelist. Trying to read")
inputdata = parseinputfiletomodels(
args.inputdata, args.delimiter, args.indices)
testtofeature = parsepaths(readFeats(args.testutts))
labels = []
dvectors = []
log.info("Extracting dvectors for input data")
for spk, v in inputdata.iteritems():
dvectors.extend(itertools.imap(extractmethod, v))
labels.extend([spk for i in xrange(len(v))])
spktonum = {spk: num for num, spk in enumerate(np.unique(labels))}
dvectors = np.array(dvectors)
labelsnum = np.array([spktonum[i] for i in labels])
log.debug("Overall we have %i labels" % (len(labelsnum)))
log.debug("Number of speakers: %i" % (len(spktonum.keys())))
log.debug("Dvector size is (%i,%i)" %
(dvectors.shape[0], dvectors.shape[1]))
log.info("Fitting LDA model")
lda.fit(dvectors, labelsnum)
errors = 0
log.info("Starting test")
for enrolemodel, vals in testreferences.iteritems():
if enrolemodel not in spktonum:
errors += 1
log.warn("Enrolemodel %s not found in the labels" % (enrolemodel))
continue
curspk = spktonum[enrolemodel]
for testutt, targetmdl in vals:
if testutt not in testtofeature:
log.warn("Utterance %s not found in the testset" % (testutt))
errors += 1
continue
# LDA currently needs a twodimensional imput vector
testdvector = testtofeature[testutt][np.newaxis, :]
score = lda.predict_log_proba(testdvector)[0]
# score is a list of all the scores ( not only one ). we need to
# find the current speakers score
finalscore = score[curspk]
args.scoreoutfile.write(
"{} {}-{} {:.3f}\n".format(enrolemodel, targetmdl, testutt, finalscore))
if errors > 0:
log.warn(
"Overall %i happened while processing the testutterances. The scores may not be complete" % (errors))
log.info("LDA estimation done, output file is: %s. Output file has the following structure: TARGETMODEL ENROLEMODEL-TESTUTT" %
(args.scoreoutfile.name))
##############################################################################
# TODO
##############################################################################
# Part II' : LDA
##############################################################################
from sklearn.lda import LDA
clf = LDA()
clf.fit(X, y)
display_3 = [2.5, 2.5]
values_proba_lda_1 = np.exp(clf.predict_log_proba(display_1))[0]
values_proba_lda_2 = np.exp(clf.predict_log_proba(display_2))[0]
values_proba_lda_3 = np.exp(clf.predict_log_proba(display_3))[0]
fig2 = plt.figure()
plot_2d(X, y)
frontiere(lambda xx: clf.predict(xx), X, step=resolution_param)
plt.annotate(r'' + '(%.2f' % values_proba_lda_1[0] + ', %.2f'
% values_proba_lda_1[1] + ', %.2f)' % values_proba_lda_1[2],
xy=(display_1[0], display_1[1]), xycoords='data',
color =color_text, xytext=(-150, 100), textcoords='offset points',
fontsize=12, arrowprops=dict(arrowstyle="->",
connectionstyle="arc3,rad=.2", color=color_text))
# plt.annotate(r'' + '(%.2f' % values_proba_lda_1[0] + ', %.2f'
def main():
lda = LDA(solver='svd')
log.basicConfig(level=args.debug,
format='%(asctime)s %(levelname)s %(message)s',
datefmt='%d/%m %H:%M:%S')
# Will be filled as dict
testreferences = None
if args.mlf:
log.info("Parsing input label file %s as mlf file" % (args.testref))
testreferences = mlffile(args.testref)
else:
log.info(
"Parsing input label file %s as test_ref file( TARGETMDL ENROLEMODEL-TESTUTT LABEL )"
% (args.testref))
testreferences = test_ref(args.testref)
# Check if the given data is in marshal format or cPickle
if args.binary:
log.info("Try to read input as a binary file")
inputdata, testutts = checkBinary([args.inputdata, args.testutts])
datadim = len(inputdata.values()[0])
dvectors = np.zeros((len(inputdata.keys()), datadim))
labels = []
log.info("Getting dvectors from input data")
for i, (spk, v) in enumerate(inputdata.iteritems()):
dvectors[i] = v
labels.append(getspkmodel(spk, args.delimiter, args.indices))
log.debug(
"Data which was loaded in (%s) has %i labels and %i dvectors" %
(args.inputdata, len(labels), len(dvectors)))
testtofeature = testutts
else:
log.info("Given data is either a folder or a filelist. Trying to read")
inputdata = parseinputfiletomodels(args.inputdata, args.delimiter,
args.indices)
testtofeature = parsepaths(readFeats(args.testutts))
labels = []
dvectors = []
log.info("Extracting dvectors for input data")
for spk, v in inputdata.iteritems():
dvectors.extend(itertools.imap(extractmethod, v))
labels.extend([spk for i in xrange(len(v))])
spktonum = {spk: num for num, spk in enumerate(np.unique(labels))}
dvectors = np.array(dvectors)
labelsnum = np.array([spktonum[i] for i in labels])
log.debug("Overall we have %i labels" % (len(labelsnum)))
log.debug("Number of speakers: %i" % (len(spktonum.keys())))
log.debug("Dvector size is (%i,%i)" %
(dvectors.shape[0], dvectors.shape[1]))
log.info("Fitting LDA model")
lda.fit(dvectors, labelsnum)
errors = 0
log.info("Starting test")
for enrolemodel, vals in testreferences.iteritems():
if enrolemodel not in spktonum:
errors += 1
log.warn("Enrolemodel %s not found in the labels" % (enrolemodel))
continue
curspk = spktonum[enrolemodel]
for testutt, targetmdl in vals:
if testutt not in testtofeature:
log.warn("Utterance %s not found in the testset" % (testutt))
errors += 1
continue
# LDA currently needs a twodimensional imput vector
testdvector = testtofeature[testutt][np.newaxis, :]
score = lda.predict_log_proba(testdvector)[0]
# score is a list of all the scores ( not only one ). we need to
# find the current speakers score
finalscore = score[curspk]
args.scoreoutfile.write("{} {}-{} {:.3f}\n".format(
enrolemodel, targetmdl, testutt, finalscore))
if errors > 0:
log.warn(
"Overall %i happened while processing the testutterances. The scores may not be complete"
% (errors))
log.info(
"LDA estimation done, output file is: %s. Output file has the following structure: TARGETMODEL ENROLEMODEL-TESTUTT"
% (args.scoreoutfile.name))
