32b8ced6fec636347b89848a98e5aef2aac3fb78,hypergan/losses/lsgan_loss.py,,create,#Any#Any#,20

Before Change

        d_real = gan.graph.d_reals[config.discriminator]
        d_fake = gan.graph.d_fakes[config.discriminator]

    with tf.variable_scope("d_linear", reuse=False):
        d_real = config.reduce(d_real, axis=1)
    with tf.variable_scope("d_linear", reuse=True):
        d_fake = config.reduce(d_fake, axis=1)

    a,b,c = config.labels
    d_loss = tf.square(d_real - b)+tf.square(d_fake - a)
    g_loss = tf.square(d_fake - c)

    d_fake_loss = -d_fake
    d_real_loss = d_real

    gan.graph.d_fake_loss=tf.reduce_mean(d_fake_loss)
    gan.graph.d_real_loss=tf.reduce_mean(d_real_loss)

    return [d_loss, g_loss]

After Change

        d_real = gan.graph.d_reals[config.discriminator]
        d_fake = gan.graph.d_fakes[config.discriminator]

    net = tf.concat([d_real, d_fake], 0)
    net = config.reduce(net, axis=1)
    s = [int(x) for x in net.get_shape()]
    d_real = tf.slice(net, [0,0], [s[0]//2,-1])
    d_fake = tf.slice(net, [s[0]//2,0], [s[0]//2,-1])

    a,b,c = config.labels
    d_loss = tf.square(d_real - b)+tf.square(d_fake - a)
    g_loss = tf.square(d_fake - c)

    d_fake_loss = -d_fake
    d_real_loss = d_real

    gan.graph.d_fake_loss=tf.reduce_mean(d_fake_loss)
    gan.graph.d_real_loss=tf.reduce_mean(d_real_loss)

    return [d_loss, g_loss]

Instances