def __setup_loss(self, loss):
import pro_gan_pytorch.Losses as losses
if isinstance(loss, str):
loss = loss.lower() # lowercase the string
if loss == "wgan":
loss = losses.CondWGAN_GP(self.dis, self.drift, use_gp=False)
# note if you use just wgan, you will have to use weight clipping
# in order to prevent gradient exploding
elif loss == "wgan-gp":
loss = losses.CondWGAN_GP(self.dis, self.drift, use_gp=True)
elif loss == "lsgan":
loss = losses.CondLSGAN(self.dis)
elif loss == "lsgan-with-sigmoid":
loss = losses.CondLSGAN_SIGMOID(self.dis)
elif loss == "hinge":
loss = losses.CondHingeGAN(self.dis)
elif loss == "standard-gan":
loss = losses.CondStandardGAN(self.dis)
elif loss == "relativistic-hinge":
loss = losses.CondRelativisticAverageHingeGAN(self.dis)
else:
raise ValueError("Unknown loss function requested")
elif not isinstance(loss, losses.ConditionalGANLoss):
raise ValueError("loss is neither an instance of GANLoss nor a string")
return loss
def __setup_loss(self, loss):
import pro_gan_pytorch.Losses as losses
if isinstance(loss, str):
loss = loss.lower() # lowercase the string
if loss == "wgan":
loss = losses.CondWGAN_GP(self.device,
self.dis,
self.drift,
use_gp=False)
# note if you use just wgan, you will have to use weight clipping
# in order to prevent gradient exploding
elif loss == "wgan-gp":
loss = losses.CondWGAN_GP(self.device,
self.dis,
self.drift,
use_gp=True)
else:
raise ValueError("Unknown loss function requested")
elif not isinstance(loss, losses.ConditionalGANLoss):
raise ValueError(
"loss is neither an instance of GANLoss nor a string")
return loss