cyclegan_loss.py

    def loss_calc(self):

        # Cycle Consistency Loss
        cyc_loss = tf.reduce_mean(tf.abs(self.x-self.cyc_x)) + tf.reduce_mean(tf.abs(self.y-self.cyc_y))

        # Adversarial Loss
        #  Discriminatorのloss(偽物を偽物と見分ける)
        disc_loss_x = tf.reduce_mean(tf.squared_difference(self.fake_rec_x,1))
        disc_loss_y = tf.reduce_mean(tf.squared_difference(self.fake_rec_y,1))

        # Generatorのloss
        g_loss_x = cyc_loss*10 + disc_loss_y
        g_loss_y = cyc_loss*10 + disc_loss_x

        # Discriminatorのloss(本物を本物と見分ける)
        d_loss_x = (tf.reduce_mean(tf.square(self.fake_pool_rec_x)) + tf.reduce_mean(tf.squared_difference(self.rec_x,1)))/2.0
        d_loss_y = (tf.reduce_mean(tf.square(self.fake_pool_rec_y)) + tf.reduce_mean(tf.squared_difference(self.rec_y,1)))/2.0

        optimizer = tf.train.AdamOptimizer(self.lr, beta1=0.5)

        self.model_vars = tf.trainable_variables()

        d_x_vars = [var for var in self.model_vars if 'd_x' in var.name]
        g_x_vars = [var for var in self.model_vars if 'g_x' in var.name]
        d_y_vars = [var for var in self.model_vars if 'd_y' in var.name]
        g_y_vars = [var for var in self.model_vars if 'g_y' in var.name]

        self.d_x_trainer = optimizer.minimize(d_loss_x, var_list=d_x_vars)
        self.d_y_trainer = optimizer.minimize(d_loss_y, var_list=d_y_vars)
        self.g_x_trainer = optimizer.minimize(g_loss_x, var_list=g_x_vars)
        self.g_y_trainer = optimizer.minimize(g_loss_y, var_list=g_y_vars)

        self.g_x_loss_summ = tf.summary.scalar("g_x_loss", g_loss_x)
        self.g_y_loss_summ = tf.summary.scalar("g_y_loss", g_loss_y)
        self.d_x_loss_summ = tf.summary.scalar("d_x_loss", d_loss_x)
        self.d_y_loss_summ = tf.summary.scalar("d_y_loss", d_loss_y)