'''
Loss functions for CycleGan
'''
import keras.backend as k
from keras.losses import mean_squared_error
_disc_train_thresh = 0.0
def discriminator_loss(y_true, y_pred):
loss = mean_squared_error(y_true, y_pred)
is_large = k.greater(loss, k.constant(_disc_train_thresh)) # threshold
is_large = k.cast(is_large, k.floatx())
return loss * is_large # binary threshold the loss to prevent overtraining the discriminator
def cycle_loss(y_true, y_pred):
if k.image_data_format() is 'channels_first':
x_w = 2
x_h = 3
else:
x_w = 1
x_h = 2
loss = k.abs(y_true - y_pred)
loss = k.sum(loss, axis=x_h)
loss = k.sum(loss, axis=x_w)
return loss
