## You should have received a copy of the GNU General Public License
## along with this program. If not, see <http://www.gnu.org/licenses/>.
import sys
import numpy
import keras
import kaldiIO
from signal import signal, SIGPIPE, SIG_DFL
if __name__ == '__main__':
model = sys.argv[1]
if not model.endswith('.h5'):
raise TypeError(
'Unsupported model type. Please use h5 format. Update Keras if needed'
)
## Load model
m = keras.models.load_model(model)
arkIn = sys.stdin.buffer
arkOut = sys.stdout.buffer
encoding = sys.stdout.encoding
signal(SIGPIPE, SIG_DFL)
uttId, featMat = kaldiIO.readUtterance(arkIn)
while uttId:
logProbMat = numpy.log(m.predict(featMat))
kaldiIO.writeUtterance(uttId, logProbMat, arkOut, encoding)
uttId, featMat = kaldiIO.readUtterance(arkIn)
## You should have received a copy of the GNU General Public License
## along with this program. If not, see <http://www.gnu.org/licenses/>.
import sys
import numpy
import keras
import kaldiIO
from signal import signal, SIGPIPE, SIG_DFL
if __name__ == '__main__':
model = sys.argv[1]
if not model.endswith('.h5'):
raise TypeError ('Unsupported model type. Please use h5 format. Update Keras if needed')
## Load model
m = keras.models.load_model (model)
arkIn = sys.stdin.buffer
arkOut = sys.stdout.buffer
encoding = sys.stdout.encoding
signal (SIGPIPE, SIG_DFL)
uttId, featMat = kaldiIO.readUtterance(arkIn)
while uttId:
logProbMat = numpy.log (m.predict (featMat))
kaldiIO.writeUtterance(uttId, logProbMat, arkOut, encoding)
uttId, featMat = kaldiIO.readUtterance(arkIn)
print('Loading models ....')
model_list = [
keras.models.load_model('pavans_1024x4/' + m) for m in model_list
]
print('Loading models, DONE')
# Splice features
p1 = Popen([
'splice-feats', '--print-args=false', '--left-context=5',
'--right-context=5', 'scp:' + data + '/feats.scp', 'ark:-'
],
stdout=PIPE)
f = open(location + '/temp.ark', 'wb')
st = time.time()
while True:
uid, featMat = readUtterance(p1.stdout)
if uid == None:
print('Reached the end of feats.scp')
break
print('Processing utt id: ' + uid)
#log_avg_prediction = np.log(arithmetic_mean(featMat, model_list, weights))
log_avg_prediction = np.log(
geometric_mean(featMat, model_list, weights))
#writeUtteranceText(uid, log_avg_prediction, f)
kaldiIO.writeUtterance(uid, log_avg_prediction, f, encoding)
et = time.time()
print('Time taken: ', et - st)
f.close()
## Save teacher predictions on disk
st = time.time()
outfile = data_tr + '/teacher_predictions.ark'
print('Writing teacher predictions...', outfile)
p1 = Popen([
'splice-feats', '--print-args=false', '--left-context=5',
'--right-context=5', 'scp:' + data_tr + '/feats.scp', 'ark:-'
],
stdout=PIPE)
with open('temp_teacher.ark', 'wb') as f:
while True:
uid, featMat = kaldiIO.readUtterance(p1.stdout)
if uid == None:
break
avg_prediction = geometric_mean(featMat, model_list, weights)
kaldiIO.writeUtterance(uid, avg_prediction, f, sys.stdout.encoding)
p1.stdout.close()
et = time.time()
os.rename('temp_teacher.ark', outfile)
print('Done writing teacher predictions for ', data_tr, '. Time taken: ',
et - st)
## Initialize data generator
trGen = dataGenerator_student_tr(data_tr, ali_tr, sgmm,
learning['batchSize'])
cvGen = dataGenerator_student_cv(data_cv, ali_cv, sgmm,
learning['batchSize'])
## Define DNN architecture and initialize weights
m = keras.models.Sequential([
keras.layers.Dense(1024, activation='relu', input_dim=429),