def model_train_validation(model_0, batch_size, patience, input_shape,
path_feature_data, indices_train, Y_train,
indices_validation, Y_validation, file_path_model,
filename_log, channel):
"""
train the model with validation early stopping
:param model_0:
:param batch_size:
:param patience:
:param input_shape:
:param path_feature_data:
:param indices_train:
:param Y_train:
:param indices_validation:
:param Y_validation:
:param file_path_model:
:param filename_log:
:return:
"""
callbacks = [
ModelCheckpoint(file_path_model,
monitor='val_loss',
verbose=0,
save_best_only=True),
EarlyStopping(monitor='val_loss', patience=patience, verbose=0),
CSVLogger(filename=filename_log, separator=';')
]
print("start training with validation...")
steps_per_epoch_train = int(np.ceil(len(indices_train) / batch_size))
steps_per_epoch_val = int(np.ceil(len(indices_validation) / batch_size))
generator_train = generator(path_feature_data=path_feature_data,
indices=indices_train,
number_of_batches=steps_per_epoch_train,
file_size=batch_size,
input_shape=input_shape,
labels=Y_train,
multi_inputs=False,
channel=channel)
generator_val = generator(path_feature_data=path_feature_data,
indices=indices_validation,
number_of_batches=steps_per_epoch_val,
file_size=batch_size,
input_shape=input_shape,
labels=Y_validation,
multi_inputs=False,
channel=channel)
model_0.fit_generator(generator=generator_train,
steps_per_epoch=steps_per_epoch_train,
epochs=500,
validation_data=generator_val,
validation_steps=steps_per_epoch_val,
callbacks=callbacks,
verbose=2)
def model_train_schluter(model_0, batch_size, input_shape, path_feature_data,
indices_all, Y_train_validation, sample_weights,
class_weights, file_path_model, filename_log,
channel):
# mmtm = MomentumScheduler(momentumIncrease)
# lrSchedule = LearningRateScheduler(lrDecrease)
# callbacks = [mmtm, lrSchedule, CSVLogger(filename=filename_log, separator=';')]
callbacks = [CSVLogger(filename=filename_log, separator=';')]
print("start training...")
# train again use all train and validation set
epochs_final = 100
steps_per_epoch_train_val = int(np.ceil(len(indices_all) / batch_size))
generator_train_val = generator(
path_feature_data=path_feature_data,
indices=indices_all,
number_of_batches=steps_per_epoch_train_val,
file_size=batch_size,
input_shape=input_shape,
labels=Y_train_validation,
sample_weights=sample_weights,
multi_inputs=False,
channel=channel)
model_0.fit_generator(
generator=generator_train_val,
steps_per_epoch=steps_per_epoch_train_val,
epochs=epochs_final,
callbacks=callbacks,
# class_weight=class_weights,
verbose=2)
model_0.save(file_path_model)
def model_train(model_0, batch_size, patience, input_shape, path_feature_data,
indices_train, Y_train, sample_weights_train,
indices_validation, Y_validation, sample_weights_validation,
indices_all, Y_train_validation, sample_weights, class_weights,
file_path_model, filename_log):
"""train the model with validation early stopping and retrain the
model with whole training dataset
"""
model_0.save_weights(basename(file_path_model))
callbacks = [
EarlyStopping(monitor='val_loss', patience=patience, verbose=0),
CSVLogger(filename=filename_log, separator=';')
]
print("start training...")
steps_per_epoch_train = int(np.ceil(len(indices_train) / batch_size))
steps_per_epoch_val = int(np.ceil(len(indices_validation) / batch_size))
generator_train = generator(path_feature_data=path_feature_data,
indices=indices_train,
number_of_batches=steps_per_epoch_train,
file_size=batch_size,
input_shape=input_shape,
labels=Y_train,
sample_weights=sample_weights_train,
multi_inputs=False)
generator_val = generator(path_feature_data=path_feature_data,
indices=indices_validation,
number_of_batches=steps_per_epoch_val,
file_size=batch_size,
input_shape=input_shape,
labels=Y_validation,
sample_weights=sample_weights_validation,
multi_inputs=False)
history = model_0.fit_generator(generator=generator_train,
steps_per_epoch=steps_per_epoch_train,
epochs=500,
validation_data=generator_val,
validation_steps=steps_per_epoch_val,
class_weight=class_weights,
callbacks=callbacks,
verbose=2)
model_0.load_weights(basename(file_path_model))
# train again use all train and validation set
epochs_final = len(history.history['val_loss'])
# epochs_final = 100
steps_per_epoch_train_val = int(np.ceil(len(indices_all) / batch_size))
generator_train_val = generator(
path_feature_data=path_feature_data,
indices=indices_all,
number_of_batches=steps_per_epoch_train_val,
file_size=batch_size,
input_shape=input_shape,
labels=Y_train_validation,
sample_weights=sample_weights,
multi_inputs=False)
model_0.fit_generator(generator=generator_train_val,
steps_per_epoch=steps_per_epoch_train_val,
epochs=epochs_final,
class_weight=class_weights,
verbose=2)
model_0.save(file_path_model)
remove(basename(file_path_model))