def fit(self,
dataloader,
nb_iter=None,
nb_epoch=None,
iter_per_epoch=None,
callbacks=[],
verbose=0):
"""Trains the underlying Keras model.
Args:
dataloader (StandardDataLoader): Manages the loading of data to
model.
nb_iter (int): The number of iterations to train the model.
nb_epoch (int): The number of epochs to train the model.
iter_per_epoch (int): Defines the number of iterations per epoch.
callbacks (list): List of Keras callbacks to run during training.
"""
nb_iter, iter_per_epoch = self._get_iterations(nb_iter, nb_epoch,
iter_per_epoch)
callbacks = CallbackList(callbacks)
callbacks._set_model(self)
callbacks.on_train_begin()
try:
epoch = 0
self.stop_training = False
for i in xrange(nb_iter):
# Begin epoch
if i % iter_per_epoch == 0:
callbacks.on_epoch_begin(epoch)
# Execution
callbacks.on_batch_begin(i)
if verbose > 0:
import time
time.sleep(0.001)
j = i % iter_per_epoch
perc = int(100 * (j + 1) / iter_per_epoch)
prog = ''.join(['='] * (perc / 2))
string = "[{:50s}] {:3d}%\r".format(prog, perc)
sys.stdout.write(string)
sys.stdout.flush()
losses = self.keras_model.train_on_batch(
*dataloader.get_training_batch())
callbacks.on_batch_end(i)
# End epoch
if (i + 1) % iter_per_epoch == 0:
callbacks.on_epoch_end(epoch, logs={'losses': losses})
epoch += 1
if self.stop_training:
break
except KeyboardInterrupt:
print "\n[BayesNet] Abort: KeyboardInterrupt"
raise
callbacks.on_train_end()
def fit(self, dataloader, nb_iter=None, nb_epoch=None, iter_per_epoch=None,
callbacks=[], verbose=0):
"""Trains the underlying Keras model.
Args:
dataloader (StandardDataLoader): Manages the loading of data to
model.
nb_iter (int): The number of iterations to train the model.
nb_epoch (int): The number of epochs to train the model.
iter_per_epoch (int): Defines the number of iterations per epoch.
callbacks (list): List of Keras callbacks to run during training.
"""
nb_iter, iter_per_epoch = self._get_iterations(
nb_iter, nb_epoch, iter_per_epoch)
callbacks = CallbackList(callbacks)
callbacks._set_model(self)
callbacks.on_train_begin()
try:
epoch = 0
self.stop_training = False
for i in xrange(nb_iter):
# Begin epoch
if i % iter_per_epoch == 0:
callbacks.on_epoch_begin(epoch)
# Execution
callbacks.on_batch_begin(i)
if verbose > 0:
import time
time.sleep(0.001)
j = i % iter_per_epoch
perc = int(100 * (j + 1) /iter_per_epoch)
prog = ''.join(['='] * (perc/2))
string = "[{:50s}] {:3d}%\r".format(prog, perc)
sys.stdout.write(string); sys.stdout.flush()
losses = self.keras_model.train_on_batch(
*dataloader.get_training_batch())
callbacks.on_batch_end(i)
# End epoch
if (i + 1) % iter_per_epoch == 0:
callbacks.on_epoch_end(epoch, logs={'losses': losses})
epoch += 1
if self.stop_training:
break
except KeyboardInterrupt:
print "\n[BayesNet] Abort: KeyboardInterrupt"
raise
callbacks.on_train_end()
callbacks += [lrs]
#mcp = ModelCheckpoint('results/experiment' + experiment_name + '_epoch{epoch}_weights.hdf5', save_best_only=True)
#callbacks += [mcp]
#lrr = INILearningRateReducer(monitor='val_acc', improve='increase', decrease_factor=0.1, patience=3, stop=3, verbose=1)
#callbacks += [lrr]
callbacks = CallbackList(callbacks)
shuffle_on_epoch_start = True
metrics = ['loss', 'acc', 'val_loss', 'val_acc',
'val_class_acc'] # show those at epoch end
do_validation = True
callbacks._set_model(model)
callbacks._set_params({
'batch_size': batch_size,
'nb_epoch': nb_epoch,
'nb_sample': nb_train_sample,
'verbose': 1,
'do_validation': do_validation,
'metrics': metrics,
})
##########################
# TRAINING
##########################
callbacks.on_train_begin()
model.stop_training = False
lrs = INILearningRateScheduler(schedule, mode='batch', logger=logger)
callbacks += [lrs]
#mcp = ModelCheckpoint('results/experiment' + experiment_name + '_epoch{epoch}_weights.hdf5', save_best_only=True)
#callbacks += [mcp]
#lrr = INILearningRateReducer(monitor='val_acc', improve='increase', decrease_factor=0.1, patience=3, stop=3, verbose=1)
#callbacks += [lrr]
callbacks = CallbackList(callbacks)
shuffle_on_epoch_start = True
metrics = ['loss', 'acc', 'val_loss', 'val_acc', 'val_class_acc'] # show those at epoch end
do_validation = True
callbacks._set_model(model)
callbacks._set_params({
'batch_size': batch_size,
'nb_epoch': nb_epoch,
'nb_sample': nb_train_sample,
'verbose': 1,
'do_validation': do_validation,
'metrics': metrics,
})
##########################
# TRAINING
##########################
callbacks.on_train_begin()
model.stop_training = False