def generative_x_entropy_loss(self, sample_logits, batch_logits):
sample_loss = losses.sigmoid_cross_entropy_loss(sample_logits, 1.0)
batch_loss = losses.sigmoid_cross_entropy_loss(batch_logits, 1.0)
loss = sample_loss # + batch_loss
if H.train_batch_logits:
loss += batch_loss
wgan_loss = tr.mean(sample_logits)
return loss
def discriminative_x_entropy_loss(self, s_logits_real, s_logits_fake, b_logits_real, b_logits_fake):
s_loss_real = losses.sigmoid_cross_entropy_loss(s_logits_real, 1.0)
s_loss_fake = losses.sigmoid_cross_entropy_loss(s_logits_fake, 0.0)
s_x_entropy_loss = (s_loss_real + s_loss_fake) / 2.0
b_loss_real = losses.sigmoid_cross_entropy_loss(b_logits_real, 1.0)
b_loss_fake = losses.sigmoid_cross_entropy_loss(b_logits_fake, 0.0)
b_x_entropy_loss = (b_loss_real + b_loss_fake) / 2.0
loss = s_x_entropy_loss # + batch_x_entropy_loss
if H.train_batch_logits:
loss += b_x_entropy_loss
return loss