Coverage for hypergan/discriminators/base_discriminator.py : 51%

from hypergan.gan_component import GANComponent 

import tensorflow as tf 

class BaseDiscriminator(GANComponent): 

def create(self, net=None, x=None, g=None): 

config = self.config 

gan = self.gan 

ops = self.ops 

if net is None: 

if x is None: 

x = gan.inputs.x 

if g is None: 

g = gan.generator.sample 

x, g = self.resize(config, x, g) 

net = tf.concat(axis=0, values=[x, g]) 

net = self.layer_filter(net) 

net = self.build(net) 

self.sample = net 

return net 

def reuse(self, net=None, x=None, g=None): 

config = self.config 

gan = self.gan 

ops = self.ops 

if net is None: 

if x is None: 

x or gan.inputs.x 

if g is None: 

g or gan.generator.sample 

x, g = self.resize(config, x, g) 

net = self.combine_filter(config, x, g) 

self.ops.reuse() 

net = self.build(net) 

self.ops.stop_reuse() 

return net 

def add_noise(self, net): 

config = self.config 

if not config.noise: 

return net 

print("[discriminator] adding noise", config.noise) 

net += tf.random_normal(net.get_shape(), mean=0, stddev=config.noise, dtype=tf.float32) 

return net 

def resize(self, config, x, g): 

if(config.resize): 

# shave off layers >= resize  

def should_ignore_layer(layer, resize): 

return int(layer.get_shape()[1]) > config['resize'][0] or \ 

int(layer.get_shape()[2]) > config['resize'][1] 

xs = [px for px in xs if not should_ignore_layer(px, config['resize'])] 

gs = [pg for pg in gs if not should_ignore_layer(pg, config['resize'])] 

x = tf.image.resize_images(x,config['resize'], 1) 

g = tf.image.resize_images(g,config['resize'], 1) 

else: 

return x, g 

def layer_filter(self, net): 

config = self.config 

gan = self.gan 

ops = self.ops 

if 'layer_filter' in config and config.layer_filter is not None: 

print("[discriminator] applying layer filter", config['layer_filter']) 

stacks = ops.shape(net)[0] // gan.batch_size() 

filters = [] 

for stack in range(stacks): 

piece = tf.slice(net, [stack * gan.batch_size(), 0,0,0], [gan.batch_size(), -1, -1, -1]) 

filters.append(config.layer_filter(gan, self.config, piece)) 

layer = tf.concat(axis=0, values=filters) 

net = tf.concat(axis=3, values=[net, layer]) 

return net 

def progressive_enhancement(self, config, net, xg): 

if config.skip_connection: 

s = self.ops.shape(net) 

extra = gan.skip_connections.get(config.skip_connection, [s[0], s[1], s[2], self.gan.channels()]) 

#TODO APpend X 

tf.concat([extra, net], axis=1) 

return net