def model_train(file_train, file_val, model_name='u_net_model',
act_func='elu', regularizer='dropout', dropoutrate=0.1,
weighted_loss=True, class_weights=[1, 1, 1],
batch_size=8, n_epochs=50,
save_model=False, save_weights=True):
'''
Function for model training
file_train: file that contains all training (training set) instances (.npz-file)
file_val: file that contains all validation (dev set) instances (.npz-file)
model_name: name of network model
act_func: activation function (see Keras documentation for valid inputs)
regularizer: 'dropout' or 'batchnorm'
dropoutrate: dropoutrate (float between 0.0 and 1.0),
not considered when 'batchnorm' is used
weighted_loss: True or False
class_weights: if 'weighted_loss' is set to True, a list with the class weights
must be provided, the length of the list must correspond with
the number of classes (ground truth), the position of a certain
class (weighting) must correspond with the ground truth
batch_size: batch size to be used for training (must be smaller than
number of training instances)
n_epochs: number of epochs the network is trained for (in this example,
the training would be stopped after 50 epochs)
save_model: True or False, saves the whole model (including training
history etc.), may consume a lot of space on disk
save_weights: True or False, saves the weights only
'''
# Load data files
with np.load(file_train + '.npz') as data:
train_X = data['data_X']
train_Y = data['data_Y']
with np.load(file_val + '.npz') as data:
val_X = data['data_X']
val_Y = data['data_Y']
_, height, width, channels = train_X.shape
n_classes = train_Y.shape[-1]
# Definition of various input parameters
model_args = (height, width, channels, n_classes)
model_kwargs = {'act_func': act_func,
'regularizer': regularizer,
'dropoutrate': dropoutrate}
if weighted_loss == True:
loss_function = keras_custom.custom_categorical_crossentropy(class_weights)
else:
loss_function = 'categorical_crossentropy'
# Build model
model = netw_models.u_net_model(*model_args, **model_kwargs)
# Compile model
model.compile(loss=loss_function,
optimizer=optimizers.RMSprop(lr=1e-4, rho=0.9),
metrics=['acc'])
# Model training
model_fit = model.fit(train_X, train_Y, batch_size, n_epochs,
validation_data=(val_X, val_Y), shuffle=True)
if save_model == True:
model.save(model_name + '.h5')
if save_weights == True:
model.save_weights(model_name + '_weights.h5')
with open(model_name + '_init.json', 'w') as file:
json.dump(model_kwargs, file)
# Plot averaged overall accuracy and loss
keras_custom.plot_acc_loss(model_fit.history['acc'],
model_fit.history['val_acc'],
model_fit.history['loss'],
model_fit.history['val_loss'],
model_name)
del train_X, train_Y, val_X, val_Y
'pw_gauss_sdims': (2, 2),
'pw_gauss_compat': 1,
'pw_bilat_sdims': (3, 3),
'pw_bilat_schan': (1.2, ),
'pw_bilat_compat': 3,
'inf_steps': 10
}
# Initialize model
if load_entire_model == True:
model = load_model(model_name + '.h5')
else:
json_name = model_weights_name.replace('weights', 'init')
with open(json_name + '.json') as f:
model_kwargs = json.load(f)
model = netw_models.u_net_model(*model_args, **model_kwargs)
model.load_weights(model_weights_name + '.h5')
my_model = netw_test_predict.TrainedModel(model)
# Evaluate network on test data
if eval_test == True:
test_metrics = my_model.model_test(test_X,
test_Y,
verbose=1,
average=average)
print(test_metrics)
if eval_test_CRF == True:
test_metrics_CRF = my_model.model_denseCRF_test(test_X,
test_Y,
verbose=1,