Coverage for hypergan/discriminators/cramer_discriminator.py : 41%

import tensorflow as tf 

import hyperchamber as hc 

import os 

import hypergan 

from hypergan.discriminators.common import * 

from hypergan.discriminators.dcgan_discriminator import DCGANDiscriminator 

from hypergan.generators.resize_conv_generator import ResizeConvGenerator 

from hypergan.encoders.uniform_encoder import UniformEncoder 

from .base_discriminator import BaseDiscriminator 

class CramerDiscriminator(BaseDiscriminator): 

def build(self, net): 

config = self.config 

gan = self.gan 

ops = self.ops 

discriminator = DCGANDiscriminator(gan, config) 

discriminator.ops = ops 

encoder = UniformEncoder(gan, gan.config.encoder) 

# careful, this order matters 

gan.generator.reuse(encoder.create()) 

g2 = gan.generator.sample 

double = tf.concat([net] + [g2, g2], axis=0) 

original = discriminator.build(double) 

d1 = self.split_batch(original, 4) 

dg = ops.concat([d1[2], d1[3]], axis=0) # xs for baseline 

#dx is a sampling of x twice 

dx = ops.concat([d1[0], d1[0]], axis=0) # xs for baseline 

xinput = ops.concat([d1[0], d1[1]], axis=0) 

#dg is a sampling of g twice 

# net is [x,g] (stacked) 

error = self.f(xinput, dx, dg) 

return error 

# this is from the paper 

def f(self, net, dx, dg): 

# Note: this is currently not working that well. we might need a second sample of X  

return tf.norm(net - dg, axis=1) - tf.norm(dx, axis=1)