def model(his_style, fut_style, use_prior, params, is_training):
quantity = his_style.get_shape().as_list()[0]
embed_dim = params.embed_dim
noise_dim = params.noise_dim
outputs = dict()
with slim.arg_scope(
[slim.fully_connected],
weights_initializer=tf.truncated_normal_initializer(stddev=0.002, seed=1)):
# Vector concat
net = tf.concat([his_style, fut_style], axis=1)
for i in xrange(params.latent_fc_layers):
net = resnet_block(net, embed_dim, afn=tf.nn.relu, nfn=None)
net = layer_norm(net)
mu = slim.fully_connected(net, noise_dim, activation_fn=None)
logs2 = slim.fully_connected(net, noise_dim, activation_fn=None)
outputs['mu'] = mu
outputs['logs2'] = logs2
if is_training or (not hasattr(params, 'sample_pdf')) or (not hasattr(params, 'sample_temp')):
temp = 1.0
else:
temp = params.sample_temp
noise_vec = tf.random_normal(shape=[quantity, noise_dim], dtype=tf.float32)
if use_prior:
outputs['latent'] = noise_vec * temp
else:
outputs['latent'] = mu + tf.multiply(noise_vec, tf.exp(0.5 * logs2)) * temp
return outputs
def add_fn(net, prev_input, params):
out_dim = net.get_shape().as_list()[1]
net = tf.concat([net, prev_input], axis=1)
for layer_i in xrange(params.latent_fc_layers):
net = resnet_block(net, out_dim, afn=tf.nn.relu, nfn=None)
net = layer_norm(net)
net = slim.fully_connected(net,
out_dim,
activation_fn=None,
normalizer_fn=None)
return net
def model(latent_input, prev_content, prev_style, params, is_training):
embed_dim = params.embed_dim
state_size = params.dec_rnn_size * 2
style_dim = params.embed_dim - params.content_dim
#
outputs = dict()
with slim.arg_scope([slim.fully_connected],
weights_initializer=tf.truncated_normal_initializer(
stddev=0.002, seed=1)):
net = latent_input
inv_z = slim.fully_connected(net,
style_dim,
activation_fn=None,
normalizer_fn=layer_norm)
net = tf.concat([inv_z, prev_style], 1)
for layer_i in xrange(params.latent_fc_layers):
net = resnet_block(net, style_dim, afn=tf.nn.relu, nfn=None)
net = layer_norm(net)
new_style = slim.fully_connected(net,
style_dim,
activation_fn=None,
normalizer_fn=None)
if hasattr(params, 'T_layer_norm') and params.T_layer_norm > 0:
new_style = layer_norm(new_style)
outputs['new_style'] = tf.identity(new_style)
#
net = tf.concat([prev_content, new_style], axis=1)
for layer_i in xrange(params.dec_fc_layers):
net = resnet_block(net, embed_dim, afn=tf.nn.relu, nfn=None)
net = layer_norm(net)
#
h0 = slim.fully_connected(net,
state_size / 2,
activation_fn=tf.nn.tanh)
c0 = slim.fully_connected(net, state_size / 2, activation_fn=None)
outputs['dec_embedding'] = tf.concat([c0, h0], 1)
return outputs
def model(latent_input, prev_content, prev_style, params, is_training):
embed_dim = params.embed_dim
state_size = params.dec_rnn_size * 2
style_dim = params.embed_dim - params.content_dim
outputs = dict()
with slim.arg_scope([slim.fully_connected],
weights_initializer=tf.truncated_normal_initializer(
stddev=0.002, seed=1)):
net = latent_input
inv_z = slim.fully_connected(net,
style_dim,
activation_fn=None,
normalizer_fn=layer_norm)
interaction_fn = get_interaction_fn(params.dec_interaction)
slim.summaries.add_histogram_summary(inv_z,
name='delta_z_activation',
prefix='summaries')
T_new = interaction_fn(inv_z * params.use_latent, prev_style, params)
if hasattr(params, 'T_layer_norm') and params.T_layer_norm > 0:
T_new = layer_norm(T_new)
new_style = T_new + prev_style
if hasattr(params, 'T_layer_norm') and params.T_layer_norm > 0:
new_style = layer_norm(new_style)
outputs['new_style'] = tf.identity(new_style)
#
net = tf.concat([prev_content, new_style], axis=1)
for layer_i in xrange(params.dec_fc_layers):
net = resnet_block(net, embed_dim, afn=tf.nn.relu, nfn=None)
net = layer_norm(net)
#
h0 = slim.fully_connected(net,
state_size / 2,
activation_fn=tf.nn.tanh)
c0 = slim.fully_connected(net, state_size / 2, activation_fn=None)
outputs['dec_embedding'] = tf.concat([c0, h0], 1)
return outputs