def generator_loss_calculation(self, fake_examples, unlabeled_examples):
"""Calculates the generator's loss."""
_, predicted_class_logits = self.D(fake_examples)
fake_binary_logits = logsumexp(predicted_class_logits, dim=1)
zeros = torch.zeros_like(fake_binary_logits)
generator_loss = self.gan_criterion(fake_binary_logits, zeros).neg()
return generator_loss
def fake_loss_calculation(self, fake_examples):
"""Calculates the fake loss."""
_, predicted_class_logits = self.D(fake_examples.detach())
fake_binary_logits = logsumexp(predicted_class_logits, dim=1)
zeros = torch.zeros_like(fake_binary_logits)
fake_loss = self.gan_criterion(fake_binary_logits, zeros)
fake_loss *= self.settings.unlabeled_loss_multiplier
return fake_loss
def interpolate_loss_calculation(self, interpolates):
"""Calculates the interpolate loss for use in the gradient penalty."""
_, predicted_class_logits = self.D(interpolates)
interpolate_binary_logits = logsumexp(predicted_class_logits, dim=1)
zeros = torch.zeros_like(interpolate_binary_logits)
interpolates_loss = self.gan_criterion(interpolate_binary_logits, zeros)
interpolates_loss *= self.settings.gradient_penalty_multiplier
return interpolates_loss
def unlabeled_loss_calculation(self, unlabeled_examples):
"""Calculates the unlabeled loss."""
_, predicted_class_logits = self.D(unlabeled_examples)
unlabeled_binary_logits = logsumexp(predicted_class_logits, dim=1)
ones = torch.ones_like(unlabeled_binary_logits)
unlabeled_loss = self.gan_criterion(unlabeled_binary_logits, ones)
unlabeled_loss *= self.settings.unlabeled_loss_multiplier
return unlabeled_loss
