def adversary(embedded, inputs, hparams, name, reuse=False):
with tf.variable_scope(name, reuse=reuse):
h0, i0 = common_layers.pad_to_same_length(embedded,
inputs,
final_length_divisible_by=16)
h0 = tf.concat([h0, tf.expand_dims(i0, axis=2)], axis=-1)
h0 = tf.layers.dense(h0, hparams.hidden_size, name="io")
h1 = transformer_vae.compress(h0, None, False, hparams, "compress1")
h2 = transformer_vae.compress(h1, None, False, hparams, "compress2")
res_dense = tf.reduce_mean(h2, axis=[1, 2])
res_single = tf.squeeze(tf.layers.dense(res_dense, 1), axis=-1)
return tf.nn.sigmoid(res_single)
def discriminator(x, compress, hparams, name, reuse=None):
with tf.variable_scope(name, reuse=reuse):
x = tf.stop_gradient(2 * x) - x # Reverse gradient.
if compress:
x = transformer_vae.compress(x, None, hparams, "compress")
else:
x = transformer_vae.residual_conv(x, 1, hparams, "compress_rc")
y = tf.reduce_mean(x, axis=1)
return tf.tanh(tf.layers.dense(y, 1, name="reduce"))
def discriminator(x, compress, hparams, name, reuse=None):
with tf.variable_scope(name, reuse=reuse):
x = tf.stop_gradient(2 * x) - x # Reverse gradient.
if compress:
x = transformer_vae.compress(x, None, hparams, "compress")
else:
x = transformer_vae.residual_conv(x, 1, hparams, "compress_rc")
y = tf.reduce_mean(x, axis=1)
return tf.tanh(tf.layers.dense(y, 1, name="reduce"))
def encode(self, inputs, target_space, hparams):
"""Add two layers strided convolutions ontop of encode."""
hparams.num_compress_steps = 2
compressed_inputs = transformer_vae.compress(inputs, c=None, is_2d=True,
hparams=hparams,
name="convolutions")
return super(TransformerSketch, self).encode(
compressed_inputs, target_space, hparams)
def encode(self, inputs, target_space, hparams):
"""Add two layers strided convolutions ontop of encode."""
inputs = common_layers.conv_block(
inputs,
hparams.hidden_size, [((1, 1), (3, 3))],
first_relu=False,
padding="SAME",
force2d=True,
name="small_image_conv")
hparams.num_compress_steps = 2
compressed_inputs = transformer_vae.compress(inputs, is_2d=True,
hparams=hparams,
name="convolutions")
return super(TransformerSketch, self).encode(
compressed_inputs, target_space, hparams)
def encode(self, inputs, target_space, hparams):
"""Add two layers strided convolutions ontop of encode."""
inputs = common_layers.conv_block(inputs,
hparams.hidden_size,
[((1, 1), (3, 3))],
first_relu=False,
padding="SAME",
force2d=True,
name="small_image_conv")
hparams.num_compress_steps = 2
compressed_inputs = transformer_vae.compress(inputs,
is_2d=True,
hparams=hparams,
name="convolutions")
return super(TransformerSketch, self).encode(compressed_inputs,
target_space, hparams)