def gradient_penalty_loss(self, y_true, y_pred, phi):
"""
Computes gradient penalty on phi to ensure smoothness
"""
# when ytrue = 0 but the discriminator give ypred =1 then it should be a small loss for the generator case
#if y_true == 0:
#lr = -K.log(K.maximum(y_pred, 1e-15) ) #ensure numerical stability avoid log 0 # negative sign because the loss should be a positive value
#else:
# lr = 0 # no loss in the other case because the y_true in all the generation case should be 0
#return lr
#
lr = K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
# compute the numerical gradient of phi
gradients = numerical_gradient_3D(phi)
# #if self.DEBUG: gradients = K.print_tensor(gradients, message='gradients are:')
#
# compute the euclidean norm by squaring ...
gradients_sqr = K.square(gradients)
# ... summing over the rows ...
gradients_sqr_sum = K.sum(gradients_sqr,
axis=np.arange(1, len(gradients_sqr.shape)))
# # ... and sqrt
# #gradient_l2_norm = K.sqrt(gradients_sqr_sum)
# # compute lambda * (1 - ||grad||)^2 still for each single sample
# #gradient_penalty = K.square(1 - gradient_l2_norm)
# # return the mean as loss over all the batch samples
# #return K.mean(gradient_l2_norm) + lr
return gradients_sqr_sum + lr
def gradient_penalty_loss(self, y_true, y_pred, phi):
"""
Computes gradient penalty on phi to ensure smoothness
"""
lr = K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
# compute the numerical gradient of phi
gradients = numerical_gradient_3D(phi)
# #if self.DEBUG: gradients = K.print_tensor(gradients, message='gradients are:')
#
# compute the euclidean norm by squaring ...
gradients_sqr = K.square(gradients)
# ... summing over the rows ...
gradients_sqr_sum = K.sum(gradients_sqr, axis=np.arange(1, len(gradients_sqr.shape)))
# # ... and sqrt
gradient_l2_norm = K.sqrt(gradients_sqr_sum)
# # compute lambda * (1 - ||grad||)^2 still for each single sample
# #gradient_penalty = K.square(1 - gradient_l2_norm)
# # return the mean as loss over all the batch samples
return K.mean(gradient_l2_norm) + lr
def smoothness_loss(self, y_true, y_pred, phi):
# mean square error loss
mse_loss = K.mean(K.square(y_pred - y_true), axis=-1)
# compute the numerical gradient of phi
gradients = numerical_gradient_3D(phi)
# #if self.DEBUG: gradients = K.print_tensor(gradients, message='gradients are:')
#
# compute the euclidean norm by squaring ...
gradients_sqr = K.square(gradients)
# ... summing over the rows ...
gradients_sqr_sum = K.sum(gradients_sqr, axis=np.arange(1, len(gradients_sqr.shape)))
# # ... and sqrt
#gradient_l2_norm = K.sqrt(gradients_sqr_sum)
# # compute lambda * (1 - ||grad||)^2 still for each single sample
#gradient_penalty = K.square(1 - gradient_l2_norm)
# # return the mean as loss over all the batch samples
#return K.mean(gradient_l2_norm) + mse_loss
return K.mean(gradients_sqr_sum) + mse_loss