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()]
net = tf.reshape(net, [s[0], -1])
d_real = tf.slice(net, [0,0], [s[0]//2,-1])
d_fake = tf.slice(net, [s[0]//2,0], [s[0]//2,-1])
After Change
net = config.reduce(net, axis=1)
shape = ops.shape(net)
net = tf.reshape(net, [shape[0], -1])
//TODO can we generalize this based on `gan.inputs`?
// split in half
