def configure(self):
""" Reconfigures the neural network in case any parameter changed.
It must be used after calling the methods named as `set_*`. It can be
used after all the setters are run (not need to configure for each one)
See `set_optimizer`, `set_model`, `set_batch_size`
"""
if any(not os.path.exists(path)
for path in (self.train_path, self.validation_path,
self.test_path)):
self.logger.info('ERROR: some dataset directory do not exists!')
image_width, image_height = \
self.model.input_shape[1], self.model.input_shape[2]
# create data generator objects
data_gen_train = DataGenerator(image_width, image_height,
self.batch_size, self.train_path)
data_gen_vali = DataGenerator(image_width, image_height,
self.batch_size, self.validation_path)
data_gen_test = DataGenerator(image_width, image_height,
self.batch_size, self.test_path)
data_gen_train.configure(DataGeneratorConfig.NORM_AND_TRANSFORM)
data_gen_vali.configure(DataGeneratorConfig.NORMALISE)
data_gen_test.configure(DataGeneratorConfig.NORMALISE)
self.train_generator = \
data_gen_train.get_single(path=self.train_path)
self.validation_generator = \
data_gen_vali.get_single(path=self.validation_path)
self.test_generator = \
data_gen_test.get_single(path=self.test_path,
shuffle=False)
def function_to_optimize(bounds): # type: (ndarray) -> int
b = bounds.astype(np.int64)
batch_size, fc1_size, fc2_size = b[:, 0][0], b[:, 1][0], b[:, 2][0]
logger.info('Bounds in action {}'.format(bounds))
base_model = get_base_model()
logger.debug('Trainability of the layers:')
model = modify(base_model, fc1_size, fc2_size, dropout=False)
for layer in model.layers:
logger.debug([layer.name, layer.trainable])
data_gen = DataGenerator(img_width, img_height, batch_size,
SMALL_TRAIN_PATH)
data_gen.configure(DataGeneratorConfig.NORM_AND_TRANSFORM)
train_generator, test_generator, validation_generator = data_gen.get(
train_path=SMALL_TRAIN_PATH,
test_path=TEST_PATH,
validate_path=TEST_PATH)
init = time.time()
history = model.fit_generator(train_generator,
steps_per_epoch=(int(
400 * 1881 / 1881 // batch_size) + 1),
epochs=number_of_epoch,
validation_data=validation_generator,
validation_steps=807 // batch_size)
end = time.time()
logger.info('[Training] Done in ' + str(end - init) + ' secs.\n')
init = time.time()
scores = model.evaluate_generator(test_generator, steps=807 // 64)
end = time.time()
logger.info('[Evaluation] Done in ' + str(end - init) + ' secs.\n')
# Get ground truth
test_labels = test_generator.classes
# Predict test images
predictions_raw = model.predict_generator(test_generator)
predictions = []
for prediction in predictions_raw:
predictions.append(np.argmax(prediction))
# Evaluate results
evaluator = Evaluator(test_labels, predictions,
label_list=list([0, 1, 2, 3, 4, 5, 6, 7]))
logger.info(
'Evaluator \n'
'Acc (model)\n'
'Accuracy: {} \n'
'Precision: {} \n'
'Recall: {} \n'
'Fscore: {}'.
format(scores[1], evaluator.accuracy, evaluator.precision,
evaluator.recall, evaluator.fscore) + '\n')
cm = evaluator.confusion_matrix()
# Plot the confusion matrix on test data
logger.info('Confusion matrix:\n')
logger.info(cm)
logger.info('Final accuracy: ' + str(evaluator.accuracy) + '\n')
end = time.time()
logger.info('Done in ' + str(end - init) + ' secs.\n')
# list all data in history
if plot_history:
do_plotting(history=history, history2=None, cm=cm)
logger.info(
'Param to optimize [Accuracy] is: {}'.format(evaluator.accuracy))
return evaluator.accuracy
def main():
base_model = get_base_model()
logger.debug('Trainability of the layers:')
model = modify_model_before_block4(base_model, dropout=False)
# model = modify_model_before_block3(base_model, dropout=False)
for layer in model.layers:
logger.debug([layer.name, layer.trainable])
# Get train, validation and test dataset
# preprocessing_function=preprocess_input,
# data_generator = ImageDataGenerator(**DataGeneratorConfig.DEFAULT)
# data_generator = ImageDataGenerator(**DataGeneratorConfig.CONFIG1)
data_gen = DataGenerator(img_width, img_height, batch_size,
SMALL_TRAIN_PATH)
data_gen.configure(DataGeneratorConfig.NORMALISE)
train_generator, test_generator, validation_generator = data_gen.get(
train_path=SMALL_TRAIN_PATH,
test_path=TEST_PATH,
validate_path=TEST_PATH)
init = time.time()
history = model.fit_generator(train_generator,
steps_per_epoch=(int(
400 * 1881 / 1881 // batch_size) + 1),
epochs=number_of_epoch,
validation_data=validation_generator,
validation_steps=807 // 64)
# unlock all layers and train
model = unlock_layers(model)
history2 = model.fit_generator(train_generator,
steps_per_epoch=(int(
400 * 1881 / 1881 // batch_size) + 1),
epochs=number_of_epoch,
validation_data=validation_generator,
validation_steps=807 // 64)
end = time.time()
logger.info('[Training] Done in ' + str(end - init) + ' secs.\n')
init = time.time()
scores = model.evaluate_generator(test_generator, steps=807 // 64)
end = time.time()
logger.info('[Evaluation] Done in ' + str(end - init) + ' secs.\n')
# Get ground truth
test_labels = test_generator.classes
# Predict test images
predictions_raw = model.predict_generator(test_generator)
predictions = []
for prediction in predictions_raw:
predictions.append(np.argmax(prediction))
# Evaluate results
evaluator = Evaluator(test_labels, predictions,
label_list=list([0, 1, 2, 3, 4, 5, 6, 7]))
logger.info(
'Evaluator \n'
'Acc (model)\n'
'Accuracy: {} \n'
'Precision: {} \n'
'Recall: {} \n'
'Fscore: {}'.
format(scores[1], evaluator.accuracy, evaluator.precision,
evaluator.recall, evaluator.fscore) + '\n')
cm = evaluator.confusion_matrix()
# Plot the confusion matrix on test data
logger.info('Confusion matrix:\n')
logger.info(cm)
logger.info('Final accuracy: ' + str(evaluator.accuracy) + '\n')
end = time.time()
logger.info('Done in ' + str(end - init) + ' secs.\n')
# list all data in history
if plot_history:
do_plotting(history, history2, cm)
