|
import tensorflow as tf
import numpy as np
import hyperchamber as hc
from hypergan.generators.common import *
from .base_generator import BaseGenerator
class ResizeConvGenerator(BaseGenerator):
def required(self):
return "final_depth activation depth_increase".split()
def depths(self, initial_width=4):
gan = self.gan
ops = self.ops
config = self.config
final_depth = config.final_depth-config.depth_increase
depths = []
target_w = gan.width()
w = initial_width
#ontehuas
i = 0
depths.append(final_depth)
while w < target_w:
w*=2
i+=1
depths.append(final_depth + i*config.depth_increase)
depths = depths[1:]
depths.reverse()
return depths
def build(self, net):
gan = self.gan
ops = self.ops
config = self.config
nets = []
activation = ops.lookup(config.activation)
final_activation = ops.lookup(config.final_activation)
block = config.block or standard_block
shape = ops.shape(net)
if len(shape) == 4 and shape[2] == 1:
primes = config.initial_dimensions or [4, 4]
print("[generator] Reshaping network based on primes.", primes)
net = ops.reshape(net, [shape[0], primes[0], primes[1], -1])
print("[generator] Reshaped network layer to ", net)
if config.skip_linear:
print("[generator] Skipping linear", net)
net = self.layer_filter(net)
if config.concat_linear:
size = ops.shape(net)[1]*ops.shape(net)[2]*config.concat_linear_filters
net2 = tf.reshape(net, [ops.shape(net)[0], -1])
net2 = tf.slice(net2, [0,0], [ops.shape(net)[0], config.concat_linear])
net2 = ops.linear(net2, size)
net2 = tf.reshape(net2, [ops.shape(net)[0], ops.shape(net)[1], ops.shape(net)[2], config.concat_linear_filters])
if config.concat_linear_regularize:
net2 = self.layer_regularizer(net2)
net2 = config.activation(net2)
net = tf.concat([net, net2], axis=3)
net = ops.conv2d(net, 3, 3, 1, 1, ops.shape(net)[3]//(config.extra_layers_reduction or 1))
for i in range(config.extra_layers or 0):
net = self.layer_regularizer(net)
net = activation(net)
net = ops.conv2d(net, 3, 3, 1, 1, ops.shape(net)[3]//(config.extra_layers_reduction or 1))
else:
net = ops.reshape(net, [ops.shape(net)[0], -1])
primes = config.initial_dimensions or [4, 4]
depths = self.depths(primes[0])
initial_depth = depths[0]
new_shape = [ops.shape(net)[0], primes[0], primes[1], initial_depth]
net = ops.linear(net, initial_depth*primes[0]*primes[1])
net = ops.reshape(net, new_shape)
shape = ops.shape(net)
depths = self.depths(initial_width = shape[1])
print("[generator] Initial depth", shape)
if config.relation_layer:
net = self.layer_regularizer(net)
net = activation(net)
net = self.relation_layer(net)
print("[generator] relational layer", net)
else:
pass
depth_reduction = np.float32(config.depth_reduction)
shape = ops.shape(net)
net = self.layer_filter(net)
for i, depth in enumerate(depths[1:]):
s = ops.shape(net)
resize = [min(s[1]*2, gan.height()), min(s[2]*2, gan.width())]
net = self.layer_regularizer(net)
if config.skip_connection:
sliced = tf.slice(net, [0,0,0,0], [-1, -1, -1, gan.channels()])
sliced = final_activation(sliced)
gan.skip_connections.set(config.skip_connection, sliced, shape=ops.shape(sliced))
net = activation(net)
if block != 'deconv':
net = ops.resize_images(net, resize, config.resize_image_type or 1)
net = self.layer_filter(net)
net = block(self, net, depth, filter=3)
else:
net = ops.deconv2d(net, 5, 5, 2, 2, depth)
net = self.layer_filter(net)
size = resize[0]*resize[1]*depth
print("[generator] layer", net, size)
net = self.layer_regularizer(net)
net = activation(net)
resize = [gan.height(), gan.width()]
if block != 'deconv':
net = ops.resize_images(net, resize, config.resize_image_type or 1)
net = self.layer_filter(net)
net = block(self, net, gan.channels(), filter=config.final_filter or 3)
else:
net = ops.deconv2d(net, 5, 5, 2, 2, gan.channels())
net = self.layer_filter(net)
if final_activation:
net = self.layer_regularizer(net)
net = final_activation(net)
if config.final_layer_filter:
net = config.final_layer_filter(gan, config, net)
self.sample = net
return self.sample
|
