def decoder_shared(z, is_training, **hparams):
activation = functools.partial(tf.nn.leaky_relu, alpha=0.2)
#norm = lambda x: x
norm = get_normalization_fn("instancenorm",
is_training) # TODO: Not mentioned in paper.
net = z
net = res_block(net, 512, 3, norm, activation)
return net
def encoder(x, is_training, **hparams):
activation = tf.nn.relu
norm = get_normalization_fn("instancenorm", is_training)
# TODO: norm or activation first? Paper says relu then norm, code the opposite
net = x
net = norm(activation(conv(net, 64, 7, 1)))
net = norm(activation(conv(net, 128, 4, 2)))
net = norm(activation(conv(net, 256, 4, 2)))
return net
def decoder_private(h, is_training, **hparams):
activation = functools.partial(tf.nn.leaky_relu, alpha=0.2)
#norm = lambda x: x
norm = get_normalization_fn("instancenorm",
is_training) # TODO: Not mentioned in paper.
net = h
net = res_block(net, 512, 3, norm, activation)
net = res_block(net, 512, 3, norm, activation)
net = res_block(net, 512, 3, norm, activation)
net = norm(activation(dconv(net, 256, 3, 2)))
net = norm(activation(dconv(net, 128, 3, 2)))
net = tf.nn.tanh(dconv(net, 3, 1, 1))
return net
def encoder_private(x, is_training, **hparams):
activation = functools.partial(tf.nn.leaky_relu, alpha=0.2)
#norm = lambda x: x # TODO: No normalization mentioned in paper? what about in resblock?
norm = get_normalization_fn("instancenorm",
is_training) # TODO: Not mentioned in paper.
net = x
net = activation(norm(conv(net, 64, 7, 1)))
net = activation(norm(conv(net, 128, 3, 2)))
net = activation(norm(conv(net, 256, 3, 2)))
net = res_block(net, 512, 3, norm, activation)
net = res_block(net, 512, 3, norm, activation)
net = res_block(net, 512, 3, norm, activation)
return net
def decoder(z, x_orig, n_attributes, is_training, **hparams):
activation = tf.nn.relu
norm = get_normalization_fn("instancenorm", is_training)
net = z
net = norm(activation(dconv(net, 128, 4, 2)))
def decoder_head():
h = norm(activation(dconv(net, 64, 4, 2)))
h = tf.concat((h, x_orig))
h = tf.nn.tanh(conv(h, 3, 7, 1))
return h
return tuple(decoder_head() for _ in range(n_attributes))
def bottleneck(z, is_training, **hparams):
activation = tf.nn.relu
norm = get_normalization_fn("instancenorm", is_training)
# TODO: norm or activation first in resblock?
# TODO: One or two convs in resblock?
net = z
net = res_block(net, 256, 3, norm, activation)
net = res_block(net, 256, 3, norm, activation)
net = res_block(net, 256, 3, norm, activation)
net = res_block(net, 256, 3, norm, activation)
net = res_block(net, 256, 3, norm, activation)
net = res_block(net, 256, 3, norm, activation)
return net