def fromrgb(img, block_id, scope='fromrgb'):
with tf.variable_scope(scope):
net = img
if params.fromrgb_use_n_img_diffs:
for d in range(
1,
min(params.fromrgb_use_n_img_diffs, block_id - 2) + 1):
scale = 2**d
tf.logging.info(
'D.block_%d fromrgb uses img diff with scale %d',
block_id, scale)
img_down = img
img_down = network_utils.downscale2d(img_down, scale)
img_down = network_utils.upscale2d(img_down, scale)
net = tf.concat([net, img - img_down], axis=1)
tf.logging.info('D.block_%d fromrgb input net shape %s', block_id,
net.shape.as_list())
if params.second_conv_channels_x2:
channels = channels_at_res(block_id - 1)
else:
channels = channels_at_res(block_id)
net = condition(net)
net = network_utils.conv(net,
filters=channels,
kernel_size=1,
weight_norm=params.weight_norm)
net = tf.nn.leaky_relu(net)
return net
def _append_channels(net, infogan_vars, infogan_num_vars, block_id):
if infogan_vars is None:
return net
with tf.variable_scope('append_channels'):
if net.shape.ndims == 2:
net = tf.concat([net, infogan_vars], axis=1)
else:
infogan_channels = tf.reshape(
infogan_vars, [gpu_batch_size, infogan_num_vars, 1, 1])
infogan_num_channels = (net.shape.as_list()[1] /
params.append_channels_div)
tf.logging.info(
'Block %d, infogan_input_method append_channels, '
'infogan_num_channels %d', block_id, infogan_num_channels)
infogan_channels = network_utils.conv(
infogan_channels,
filters=infogan_num_channels,
kernel_size=1,
weight_norm=params.weight_norm)
infogan_channels = tf.nn.leaky_relu(infogan_channels)
side = net.shape.as_list()[-1]
infogan_channels = tf.tile(infogan_channels,
[1, 1, side, side])
net = tf.concat([net, infogan_channels], axis=1)
tf.logging.info(
'Block %d, infogan_input_method append_channels, '
'new net shape %s', block_id, net.shape.as_list())
return net
def torgb(net, scope='torgb', block_id=None):
with tf.variable_scope(scope):
net = condition(net, block_id)
net = network_utils.conv(net,
filters=3,
kernel_size=1,
weight_norm=params.weight_norm)
if params.torgb_tanh:
net = tf.nn.tanh(net)
return net
def CCNA(net, block_id, scope):
with tf.variable_scope(scope):
net = condition(net, block_id)
net = network_utils.conv(net,
filters=channels_at_res(block_id),
weight_norm=params.weight_norm,
kernel_size=params.kernel_size)
net = network_utils.norm(net,
axis=1,
version=params.norm,
is_training=is_training,
gpu_id=gpu_id,
per_gpu=params.norm_per_gpu)
net = tf.nn.leaky_relu(net)
if params.pn_after_act: net = network_utils.pixel_norm(net, axis=1)
return net