def do_train_sgd(self, data, label, num_data, batch_size,
learning_rate_array):
return sgd_train(model=self,
data=data,
label=label,
num_data=num_data,
batch_size=batch_size,
learning_rate_array=learning_rate_array)
def do_log_layer_training_only(self, data, label, num_data, batch_size, learning_rate_array, num_epochs=1):
self.cost = self.logLayer.cost
for i in range(num_epochs):
print('epoch %d' % i)
monitor_results = sgd_train(
model=self, data=data, label=label,
num_data=num_data, batch_size=batch_size,
learning_rate_array=learning_rate_array,
params=self.logLayer.params)
for monitor_index, monitor in enumerate(self.monitors):
print("\t%10s: %f" % (monitor.name, monitor_results[monitor_index]))
def do_fine_tuning(self, data, label, num_data, batch_size, learning_rate_array, num_epochs=1):
monitor_results = None
self.cost = self.logLayer.cost
for i in range(num_epochs):
self.logger.info('epoch %d' % i)
monitor_results = sgd_train(
model=self, data=data, label=label,
num_data=num_data, batch_size=batch_size,
learning_rate_array=learning_rate_array)
for monitor_index, monitor in enumerate(self.monitors):
self.logger.info("\t%10s: %f" % (monitor.name, monitor_results[monitor_index]))
return monitor_results
def do_pretraining(self, data, num_data, batch_size, learning_rate_array, num_epochs=1):
for i in range(len(self.dA_layers)):
# Training auto-encoder layer by layer
self.logger.info("Pre-training layer %d" % i)
self.cost = self.dA_layers[i].cost
for j in range(num_epochs):
self.logger.info("Epoch %d" % j)
monitor_results = sgd_train(
model=self, data=data, label=None,
num_data=num_data, batch_size=batch_size,
learning_rate_array=learning_rate_array,
params=self.dA_layers[i].params,
monitors=self.pretrain_monitor_array[i])
for monitor_index, monitor in enumerate(self.monitors):
self.logger.info("\t%10s: %f" % (monitor.name, monitor_results[monitor_index]))
return
def do_training_sgd(self,
data,
label,
num_data,
batch_size,
learning_rate_array,
num_epochs=1,
prevalence=None):
"""
Use Stochastic Gradient Descent to train the model
:param data: Training Data Tensor
:param label: Corresponding Class Tensor
:param num_data: The total amount of data
:param batch_size: The size per batch
:param learning_rate_array: Learning Rate for each parameter
:param num_epochs: Epochs
:param prevalence: inverse of Percentage of each class in the training samples
:return: None
"""
if prevalence is not None:
self.logRegressionLayer.prevalence = prevalence
for i in range(num_epochs):
self.logger.info("Epoch %d" % i)
monitor_results = sgd_train(
model=self,
data=data,
label=label,
num_data=num_data,
batch_size=batch_size,
learning_rate_array=learning_rate_array)
for monitor_index, monitor in enumerate(self.monitors):
self.logger.info(
"\t%10s: %f" %
(monitor.name, monitor_results[monitor_index]))
self.logRegressionLayer.prevalence = np.ones(
(self.logRegressionLayer.W.get_value().shape[1], ),
dtype=theano.config.floatX)