def conv_encoder_32_large(inputs,
z_dim,
nonlinearity=None,
bn=True,
kernel_initializer=None,
kernel_regularizer=None,
is_training=False,
counters={}):
name = get_name("conv_encoder_32_large", counters)
print("construct", name, "...")
with tf.variable_scope(name):
with arg_scope([conv2d, dense],
nonlinearity=nonlinearity,
bn=bn,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
is_training=is_training,
counters=counters):
outputs = inputs
outputs = conv2d(outputs, 32, 1, 1, "SAME")
outputs = conv2d(outputs, 32, 1, 1, "SAME")
outputs = conv2d(outputs, 64, 4, 2, "SAME")
outputs = conv2d(outputs, 128, 4, 2, "SAME")
outputs = conv2d(outputs, 256, 4, 2, "SAME")
outputs = conv2d(outputs, 512, 4, 1, "VALID")
outputs = tf.reshape(outputs, [-1, 512])
z_mu = dense(outputs, z_dim, nonlinearity=None, bn=False)
z_log_sigma_sq = dense(outputs, z_dim, nonlinearity=None, bn=False)
return z_mu, z_log_sigma_sq
def aggregator(r,
num_c,
z_dim,
method=tf.reduce_mean,
nonlinearity=None,
bn=True,
kernel_initializer=None,
kernel_regularizer=None,
is_training=False,
counters={}):
name = get_name("aggregator", counters)
print("construct", name, "...")
with tf.variable_scope(name):
with arg_scope([dense],
nonlinearity=nonlinearity,
bn=bn,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
is_training=is_training,
counters=counters):
r_pr = method(r[:num_c], axis=0, keepdims=True)
r = method(r, axis=0, keepdims=True)
r = tf.concat([r_pr, r], axis=0)
size = 256
r = dense(r, size)
r = dense(r, size)
r = dense(r, size)
z_mu = dense(r, z_dim, nonlinearity=None, bn=False)
z_log_sigma_sq = dense(r, z_dim, nonlinearity=None, bn=False)
return z_mu[:1], z_log_sigma_sq[:1], z_mu[1:], z_log_sigma_sq[1:]
def conditional_decoder(x,
z,
nonlinearity=None,
bn=True,
kernel_initializer=None,
kernel_regularizer=None,
is_training=False,
counters={}):
name = get_name("conditional_decoder", counters)
print("construct", name, "...")
with tf.variable_scope(name):
with arg_scope([dense],
nonlinearity=nonlinearity,
bn=bn,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
is_training=is_training):
size = 256
batch_size = tf.shape(x)[0]
x = tf.tile(x, tf.stack([1, int_shape(z)[1]]))
z = tf.tile(z, tf.stack([batch_size, 1]))
# xz = x + z * tf.get_variable(name="coeff", shape=(), dtype=tf.float32, initializer=tf.constant_initializer(2.0))
xz = x
a = dense(xz, size, nonlinearity=None) + dense(
z, size, nonlinearity=None)
outputs = tf.nn.tanh(a) * tf.sigmoid(a)
for k in range(4):
a = dense(outputs, size, nonlinearity=None) + dense(
z, size, nonlinearity=None)
outputs = tf.nn.tanh(a) * tf.sigmoid(a)
outputs = dense(outputs, 1, nonlinearity=None, bn=False)
outputs = tf.reshape(outputs, shape=(batch_size, ))
return outputs
def fc_encoder(X,
y,
r_dim,
nonlinearity=None,
bn=True,
kernel_initializer=None,
kernel_regularizer=None,
is_training=False,
counters={}):
inputs = tf.concat([X, y[:, None]], axis=1)
name = get_name("fc_encoder", counters)
print("construct", name, "...")
with tf.variable_scope(name):
with arg_scope([dense],
nonlinearity=nonlinearity,
bn=bn,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
is_training=is_training,
counters=counters):
size = 256
outputs = dense(inputs, size)
outputs = nonlinearity(
dense(outputs, size, nonlinearity=None) +
dense(inputs, size, nonlinearity=None))
inputs = outputs
outputs = dense(outputs, size)
outputs = nonlinearity(
dense(outputs, size, nonlinearity=None) +
dense(inputs, size, nonlinearity=None))
outputs = dense(outputs, size)
outputs = dense(outputs, r_dim, nonlinearity=None, bn=False)
return outputs
def conv_decoder_32_large(inputs,
output_features=False,
nonlinearity=None,
bn=True,
kernel_initializer=None,
kernel_regularizer=None,
is_training=False,
counters={}):
name = get_name("conv_decoder_32_large", counters)
print("construct", name, "...")
with tf.variable_scope(name):
with arg_scope([deconv2d, dense],
nonlinearity=nonlinearity,
bn=bn,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
is_training=is_training,
counters=counters):
outputs = dense(inputs, 512)
outputs = tf.reshape(outputs, [-1, 1, 1, 512])
outputs = deconv2d(outputs, 256, 4, 1, "VALID")
outputs = deconv2d(outputs, 128, 4, 2, "SAME")
outputs = deconv2d(outputs, 64, 4, 2, "SAME")
outputs = deconv2d(outputs, 32, 4, 2, "SAME")
if output_features:
return outputs
outputs = deconv2d(outputs,
3,
1,
1,
"SAME",
nonlinearity=tf.sigmoid,
bn=False)
outputs = 2. * outputs - 1.
return outputs
def conv_decoder_28_binary(inputs,
nonlinearity=None,
bn=True,
kernel_initializer=None,
kernel_regularizer=None,
is_training=False,
counters={}):
name = get_name("conv_decoder_28_binary", counters)
print("construct", name, "...")
with tf.variable_scope(name):
with arg_scope([deconv2d, dense],
nonlinearity=nonlinearity,
bn=bn,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
is_training=is_training,
counters=counters):
outputs = dense(inputs, 128)
outputs = tf.reshape(outputs, [-1, 1, 1, 128])
outputs = deconv2d(outputs, 128, 4, 1, "VALID")
outputs = deconv2d(outputs, 64, 4, 1, "VALID")
outputs = deconv2d(outputs, 64, 4, 2, "SAME")
outputs = deconv2d(outputs, 32, 4, 2, "SAME")
outputs = deconv2d(outputs,
1,
1,
1,
"SAME",
nonlinearity=None,
bn=False)
return outputs
def mlp(X,
scope="mlp",
params=None,
nonlinearity=None,
bn=True,
kernel_initializer=None,
kernel_regularizer=None,
is_training=False,
counters={}):
name = get_name(scope, counters)
print("construct", name, "...")
if params is not None:
params.reverse()
with tf.variable_scope(name):
default_args = {
"nonlinearity": nonlinearity,
"bn": bn,
"kernel_initializer": kernel_initializer,
"kernel_regularizer": kernel_regularizer,
"is_training": is_training,
"counters": counters,
}
with arg_scope([dense], **default_args):
batch_size = tf.shape(X)[0]
size = 256
outputs = X
for k in range(4):
if params is not None:
outputs = dense(outputs,
size,
W=params.pop(),
b=params.pop())
else:
outputs = dense(outputs, size)
if params is not None:
outputs = dense(outputs,
1,
nonlinearity=None,
W=params.pop(),
b=params.pop())
else:
outputs = dense(outputs, 1, nonlinearity=None)
outputs = tf.reshape(outputs, shape=(batch_size, ))
return outputs
