def _train(self):
xor_data_generator = get_xor_generator(self.batch_size)
validation_size = 4
xor_validation_data = get_xor_data(validation_size)
self.history = self.model.fit(xor_data_generator,
steps_per_epoch=1,
epochs=self.epochs,
validation_data=xor_validation_data,
verbose=0,
callbacks=self._callbacks().values())
return self._get_result()
def memory_efficiency(batch_size_continuous,
lr_exp,
momentum,
layer_size_continuous,
layer_count_continuous,
tune_report_call_back=None):
# parameter initialize
batch_size = int(batch_size_continuous / 4) * 4
lr = 10**lr_exp # float(sys.argv[3])
momentum = momentum
layer_size = int(layer_size_continuous)
layer_count = int(layer_count_continuous)
op = tf.keras.optimizers.SGD(lr=lr, momentum=momentum, nesterov=True)
model = xor_net(layer_size, layer_count, verbose=False)
loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)
model.compile(loss=loss, optimizer=op, metrics=['acc'])
stop_if_goal_reached = GoalReachedStopper()
time_callback = TimeHistory()
validation_size = 4
xor_data_generator = get_xor_generator(batch_size)
xor_validation_data = get_xor_data(validation_size)
if tune_report_call_back == None:
callbacks = [stop_if_goal_reached, time_callback]
else:
callbacks = [
stop_if_goal_reached, time_callback, tune_report_call_back
]
history = model.fit_generator(xor_data_generator,
steps_per_epoch=1,
epochs=5,
validation_data=xor_validation_data,
verbose=0,
callbacks=callbacks)
memory_usage = get_model_memory_usage(batch_size, model)
epoch_usage = len(history.history['val_acc'])
final_acc = history.history['val_acc'][-1]
print("memory_usage:", memory_usage, "epoch size:", epoch_usage, "acc:",
final_acc)
memory_efficiency = 1. / memory_usage
del model
gc.collect()
tf.keras.backend.clear_session()
if final_acc < 1.0 or epoch_usage > 3:
return 0. - epoch_usage + final_acc
else:
return memory_efficiency
def initial_slop(batch_size_continuous,
lr_exp,
momentum,
layer_size_continuous,
layer_count_continuous,
tune_report_call_back=None):
# parameter initialize
batch_size = int(batch_size_continuous / 4) * 4
lr = 10**lr_exp # float(sys.argv[3])
momentum = momentum
layer_size = int(layer_size_continuous)
layer_count = int(layer_count_continuous)
op = tf.keras.optimizers.SGD(lr=lr, momentum=momentum, nesterov=True)
model = xor_net(layer_size, layer_count, verbose=False)
loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)
model.compile(loss=loss, optimizer=op, metrics=['acc'])
time_callback = TimeHistory()
validation_size = 4
xor_data_generator = get_xor_generator(batch_size)
xor_validation_data = get_xor_data(validation_size)
if tune_report_call_back == None:
callbacks = [time_callback]
else:
callbacks = [time_callback, tune_report_call_back]
history = model.fit_generator(xor_data_generator,
steps_per_epoch=1,
epochs=10,
validation_data=xor_validation_data,
verbose=0,
callbacks=callbacks)
computation_time = np.sum(time_callback.times)
slop = (history.history['val_loss'][0] -
history.history['val_loss'][-1]) / computation_time
del model
gc.collect()
tf.keras.backend.clear_session()
return slop
def initial_acc(batch_size_continuous,
lr_exp,
momentum,
layer_size_continuous,
layer_count_continuous,
tune_report_call_back=None):
# parameter initialize
batch_size = int(batch_size_continuous / 4) * 4
lr = 10**lr_exp
momentum = momentum
layer_size = int(layer_size_continuous)
layer_count = int(layer_count_continuous)
op = tf.keras.optimizers.SGD(lr=lr, momentum=momentum, nesterov=True)
model = xor_net(layer_size, layer_count, verbose=False)
loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)
model.compile(loss=loss, optimizer=op, metrics=['acc'])
one_second_stop_callback = OneSecondStopper()
validation_size = 4
xor_data_generator = get_xor_generator(batch_size)
xor_validation_data = get_xor_data(validation_size)
if tune_report_call_back == None:
callbacks = [one_second_stop_callback]
else:
callbacks = [one_second_stop_callback, tune_report_call_back]
history = model.fit_generator(
xor_data_generator,
steps_per_epoch=1, # this is a virtual parameter
epochs=10,
validation_data=xor_validation_data,
verbose=0,
callbacks=callbacks)
final_acc = history.history['val_acc'][-1]
del model
gc.collect()
tf.keras.backend.clear_session()
return final_acc
def model_contruction_time(batch_size_continuous,
lr_exp,
momentum,
layer_size_continuous,
layer_count_continuous,
tune_report_call_back=None):
construct_start_time = time.time()
batch_size = int(batch_size_continuous / 4) * 4
lr = 10**lr_exp
momentum = momentum
layer_size = int(layer_size_continuous)
layer_count = int(layer_count_continuous)
op = tf.keras.optimizers.SGD(lr=lr, momentum=momentum, nesterov=True)
model = xor_net(layer_size, layer_count, verbose=False)
loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)
model.compile(loss=loss, optimizer=op, metrics=['acc'])
validation_size = 4
xor_data_generator = get_xor_generator(batch_size)
xor_validation_data = get_xor_data(validation_size)
if tune_report_call_back == None:
callbacks = []
else:
callbacks = [tune_report_call_back]
model.fit_generator(
xor_data_generator,
steps_per_epoch=1, # this is a virtual parameter
epochs=1,
validation_data=xor_validation_data,
callbacks=callbacks,
verbose=0)
del model
gc.collect()
tf.keras.backend.clear_session()
construction_time = time.time() - construct_start_time
return 1. / construction_time