Coverage for hypergan/losses/standard_loss.py : 90%

import tensorflow as tf 

import hyperchamber as hc 

from hypergan.losses.base_loss import BaseLoss 

class StandardLoss(BaseLoss): 

def required(self): 

return "reduce".split() 

def _create(self, d_real, d_fake): 

ops = self.ops 

config = self.config 

gan = self.gan 

generator_target_probability = config.generator_target_probability or 0.8 

label_smooth = config.label_smooth or 0.2 

zeros = tf.zeros_like(d_fake) 

ones = tf.ones_like(d_fake) 

if config.improved: 

g_loss = self.sigmoid_kl_with_logits(d_fake, generator_target_probability) 

d_loss = self.sigmoid_kl_with_logits(d_real, 1.-label_smooth) + \ 

tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake, labels=zeros) 

else: 

g_loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake, labels=zeros) 

d_loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=d_real, labels=zeros) + \ 

tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake, labels=ones) 

return [d_loss, g_loss]