def __init__(self,
weights=opt.half_weight_dir,
num_classes=opt.num_classes,
im_size=opt.im_size,
channels=opt.channels):
"""
Wrapper for easier predictions and debugging of a HalfDenseNetFCN model.
:param weights: Path to a file containing the weights of a pre-trained HalfDenseNetFCN.
:param num_classes: An integer representing the number of classes.
:param im_size: An integer representing the size of a single dimension each image.
:param channels: How many channels does the image have (e.g. 1 -> grayscale, 3 -> RGB)
"""
print('Instantiating a Half Model:')
print('weights: {}'.format(weights))
print('classes: {}'.format(num_classes))
image_dims = (im_size, im_size)
input_shape = image_dims + (channels, )
print('input shape: {}'.format(input_shape))
# Main model
self.model = custom_models.HalfDenseNetFCN(input_shape, num_classes,
weights)
# Ak model
self.Ak_model = Model(inputs=self.model.input,
outputs=self.model.layers[-3].output)
# Wk array
self.Wk = self.model.layers[-1].get_weights()[0]
order=[0, 1], # use nearest neighbour or bilinear interpolation (fast)
cval=0,
mode='constant'
)),
iaa.Fliplr(0.5)
])
# Initialize Image Generators for batching
tg = ImageDataGenerator(preprocessing_function=seq.augment_image)
vg = ImageDataGenerator()
train_gen = tg.flow_from_directory(train_path, target_size=img_dims, batch_size=batch_size)
test_gen = vg.flow_from_directory(test_path, target_size=img_dims, batch_size=batch_size)
# Model
model = custom_models.HalfDenseNetFCN(input_shape, num_classes)
# Load previously trained model
if half_weight_dir:
previous_weights = half_weight_dir
model.load_weights(previous_weights)
# Define keras callbacks
lr_reducer = ReduceLROnPlateau(monitor='loss', factor=np.sqrt(0.1), patience=10, cooldown=0, min_lr=1e-5, verbose=1)
tb = TensorBoard(log_dir=log_dir, batch_size=batch_size)
model_chp = ModelCheckpoint(os.path.join(results_dir, 'best_weights.h5'), monitor='val_acc',
save_best_only=True, save_weights_only=True)
# Train the model
model.fit_generator(train_gen, epochs=epochs, callbacks=[tb, model_chp, lr_reducer], validation_data=test_gen)
factor=np.sqrt(0.1),
patience=10,
cooldown=0,
min_lr=1e-5)
model_chp = ModelCheckpoint(os.path.join(results_dir, 'best_weights.h5'),
monitor='loss',
save_best_only=True,
save_weights_only=True)
tb = TensorBoard(log_dir=log_dir, batch_size=batch_size)
# Instantiate full and half models
model = custom_models.AuxiliaryDenseNetFCN(input_shape,
num_classes,
weights=None)
pretrained = custom_models.HalfDenseNetFCN(input_shape,
num_classes,
weights=weight_dir)
# Transfer weights from half to full and freeze them
transfer_weights(pretrained, model)
del pretrained
for layer in model.layers[:121] + [
model.layers[-5], model.layers[-3], model.layers[-1]
]:
layer.trainable = False
# Re-compile the model
model.compile(optimizer='adadelta',
loss={
'main_output': 'mean_squared_error',