def trainvad():
"""Driver to train a Voice Activation Detection (VAD) classifier.
"""
from nktrainutils import SGDSVM, splitTrainEval, evalSVM
import numpy as np
import librosa
from cPickle import dump, load
if len(sys.argv) < 3:
print 'Usage: python %s <input movie names> <output classifier>' % (
sys.argv[0])
sys.exit()
posfeats, negfeats = [], []
rate = 22050
LogAmp = librosa.util.FeatureExtractor(librosa.logamplitude,
ref_power=np.max)
for i, fname in enumerate(open(sys.argv[1])):
fname = fname.rstrip('\n')
subfname = fname.rsplit('.', 2)[0] + '.srt'
try:
# read subtitles and get training segments
lines = [l.strip() for l in open(subfname) if l.strip()]
dialogs = getdialogs(lines)
pos, neg = getTrainingSegments(dialogs, 100, 600, seglen=0.5)
print fname, len(dialogs), len(pos), len(neg)
# read features
feats = load(open(fname))
def featfunc(a, b):
"""concats feats from the given times"""
a, b = int(a * 10), int(b * 10)
ret = feats[:, a:b].transpose().flatten()
return ret
curpos = [featfunc(a, b) for a, b in pos]
curneg = [featfunc(a, b) for a, b in neg]
posfeats.extend(
f for f in curpos
if f is not None and len(f) > 0 and not np.isnan(f).any())
negfeats.extend(
f for f in curneg
if f is not None and len(f) > 0 and not np.isnan(f).any())
print len(posfeats), len(negfeats)
if len(posfeats) >= 5000: break
except IOError:
continue
# split into train and eval subsets and then train and run svm
(trainpos, trainneg), (evalpos,
evalneg) = splitTrainEval(posfeats, negfeats, -20)
svm = SGDSVM()
t1 = time.time()
model, score = svm.train(trainpos + trainneg,
[1] * len(trainpos) + [-1] * len(trainneg),
ncv=0,
n_iter=50)
try:
score = evalSVM(svm, model, evalpos + evalneg,
[1] * len(evalpos) + [-1] * len(evalneg))
except Exception, e:
print set(map(lambda x: x.shape, trainpos + trainneg))
raise
def oldresyncmain():
"""Main driver for subtitle resyncing"""
from nktrainutils import SGDSVM, splitTrainEval, evalSVM
import librosa
from cPickle import dump, load
rate = 22050
if len(sys.argv) < 1:
print 'Usage: python %s <video or audio file> <subtitle file> <vad model>' % (
sys.argv[0])
sys.exit()
fname, subfname, modelfname = sys.argv[1:4]
model = SGDSVM.load(modelfname)
allfeats = getmel(fname)
def featfunc(a, b):
"""concats feats from the given times"""
a, b = int(a * 10), int(b * 10)
ret = allfeats[:, a:b].transpose().flatten()
return ret
seglen = 0.5
subels = []
A, B = 0, 300
for start in range(A, B):
feats = featfunc(start, start + seglen)
cls = SGDSVM().classify(model, [feats])[0]
subels.append(('Cls: %0.4f' % (cls), start, start + 1))
print start, cls
sys.stdout.flush()
#print subels
playpreview(fname.replace('.mel', ''), subels, A, B)
def trainvad():
"""Driver to train a Voice Activation Detection (VAD) classifier.
"""
from nktrainutils import SGDSVM, splitTrainEval, evalSVM
import numpy as np
import librosa
from cPickle import dump, load
if len(sys.argv) < 3:
print "Usage: python %s <input movie names> <output classifier>" % (sys.argv[0])
sys.exit()
posfeats, negfeats = [], []
rate = 22050
LogAmp = librosa.util.FeatureExtractor(librosa.logamplitude, ref_power=np.max)
for i, fname in enumerate(open(sys.argv[1])):
fname = fname.rstrip("\n")
subfname = fname.rsplit(".", 2)[0] + ".srt"
try:
# read subtitles and get training segments
lines = [l.strip() for l in open(subfname) if l.strip()]
dialogs = getdialogs(lines)
pos, neg = getTrainingSegments(dialogs, 100, 600, seglen=0.5)
print fname, len(dialogs), len(pos), len(neg)
# read features
feats = load(open(fname))
def featfunc(a, b):
"""concats feats from the given times"""
a, b = int(a * 10), int(b * 10)
ret = feats[:, a:b].transpose().flatten()
return ret
curpos = [featfunc(a, b) for a, b in pos]
curneg = [featfunc(a, b) for a, b in neg]
posfeats.extend(f for f in curpos if f is not None and len(f) > 0 and not np.isnan(f).any())
negfeats.extend(f for f in curneg if f is not None and len(f) > 0 and not np.isnan(f).any())
print len(posfeats), len(negfeats)
if len(posfeats) >= 5000:
break
except IOError:
continue
# split into train and eval subsets and then train and run svm
(trainpos, trainneg), (evalpos, evalneg) = splitTrainEval(posfeats, negfeats, -20)
svm = SGDSVM()
t1 = time.time()
model, score = svm.train(trainpos + trainneg, [1] * len(trainpos) + [-1] * len(trainneg), ncv=0, n_iter=50)
try:
score = evalSVM(svm, model, evalpos + evalneg, [1] * len(evalpos) + [-1] * len(evalneg))
except Exception, e:
print set(map(lambda x: x.shape, trainpos + trainneg))
raise
def oldresyncmain():
"""Main driver for subtitle resyncing"""
from nktrainutils import SGDSVM, splitTrainEval, evalSVM
import librosa
from cPickle import dump, load
rate = 22050
if len(sys.argv) < 1:
print "Usage: python %s <video or audio file> <subtitle file> <vad model>" % (sys.argv[0])
sys.exit()
fname, subfname, modelfname = sys.argv[1:4]
model = SGDSVM.load(modelfname)
allfeats = getmel(fname)
def featfunc(a, b):
"""concats feats from the given times"""
a, b = int(a * 10), int(b * 10)
ret = allfeats[:, a:b].transpose().flatten()
return ret
seglen = 0.5
subels = []
A, B = 0, 300
for start in range(A, B):
feats = featfunc(start, start + seglen)
cls = SGDSVM().classify(model, [feats])[0]
subels.append(("Cls: %0.4f" % (cls), start, start + 1))
print start, cls
sys.stdout.flush()
# print subels
playpreview(fname.replace(".mel", ""), subels, A, B)
posi += 1
elif i == -1:
ret.append(-negvals[negi])
negi += 1
elif i == 0:
ret.append(0.0)
return ret
# testing main, uses SGDSVM from nktrainutils (which uses scikit-learn)
if __name__ == '__main__':
from nktrainutils import SGDSVM
import numpy as np
import numpy.random as npr
from pprint import pprint
svm = SGDSVM()
size = (1000, 100)
posm = 0.8
negm = -0.9
# train
trainpos = npr.normal(posm, size=size)
trainneg = npr.normal(negm, size=size)
model, _ = svm.train(np.vstack((trainpos, trainneg)),
[1] * len(trainpos) + [-1] * len(trainneg),
ncv=0)
# eval
evalpos = npr.normal(posm, size=size)
evalneg = npr.normal(negm, size=size)
outs = svm.classify(model, np.vstack((evalpos, evalneg)))
# fit
params = easyfitsvm(outs)
elif i == -1:
ret.append(-negvals[negi])
negi += 1
elif i == 0:
ret.append(0.0)
return ret
# testing main, uses SGDSVM from nktrainutils (which uses scikit-learn)
if __name__ == "__main__":
from nktrainutils import SGDSVM
import numpy as np
import numpy.random as npr
from pprint import pprint
svm = SGDSVM()
size = (1000, 100)
posm = 0.8
negm = -0.9
# train
trainpos = npr.normal(posm, size=size)
trainneg = npr.normal(negm, size=size)
model, _ = svm.train(np.vstack((trainpos, trainneg)), [1] * len(trainpos) + [-1] * len(trainneg), ncv=0)
# eval
evalpos = npr.normal(posm, size=size)
evalneg = npr.normal(negm, size=size)
outs = svm.classify(model, np.vstack((evalpos, evalneg)))
# fit
params = easyfitsvm(outs)
# norm
outs = sorted(outs[::40])
