def evaluate_task(inputs):
train_inputs, train_outputs, test_inputs, test_outputs = inputs
with tf.variable_scope('shared_features'):
# extract features from train and test data
features_train = feature_extractor_fn(images=train_inputs,
output_size=args.d_theta,
use_batch_norm=True,
dropout_keep_prob=dropout_keep_prob)
features_test = feature_extractor_fn(images=test_inputs,
output_size=args.d_theta,
use_batch_norm=True,
dropout_keep_prob=dropout_keep_prob)
# Infer classification layer from q
with tf.variable_scope('classifier'):
classifier = infer_classifier(features_train, train_outputs, args.d_theta, args.way)
# Local reparameterization trick
# Compute parameters of q distribution over logits
weight_mean, bias_mean = classifier['weight_mean'], classifier['bias_mean']
weight_log_variance, bias_log_variance = classifier['weight_log_variance'], classifier['bias_log_variance']
logits_mean_test = tf.matmul(features_test, weight_mean) + bias_mean
logits_log_var_test = \
tf.log(tf.matmul(features_test ** 2, tf.exp(weight_log_variance)) + tf.exp(bias_log_variance))
logits_sample_test = sample_normal(logits_mean_test, logits_log_var_test, args.samples)
test_labels_tiled = tf.tile(tf.expand_dims(test_outputs, 0), [args.samples, 1, 1])
task_log_py = multinoulli_log_density(inputs=test_labels_tiled, logits=logits_sample_test)
averaged_predictions = tf.reduce_logsumexp(logits_sample_test, axis=0) - tf.log(L)
task_accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(test_outputs, axis=-1),
tf.argmax(averaged_predictions, axis=-1)), tf.float32))
task_score = tf.reduce_logsumexp(task_log_py, axis=0) - tf.log(L)
task_loss = -tf.reduce_mean(task_score, axis=0)
return [task_loss, task_accuracy]
def shapenet_inference(image_features, angles, d_theta, d_psi, num_samples):
"""
Perform inference for the view reconstruction task
:param image_features: tensor (N x d) of training image features.
:param angles: tensor (N x d_angle) of training image angles.
:param d_theta: integer dimensionality of the features.
:param d_psi: integer dimensionality of adaptation parameters.
:param num_samples: number of samples to generate from the distribution.
:return: dictionary containing distribution parameters and samples.
"""
# Concatenate features and angles
h = tf.concat([image_features, angles], axis=-1)
# denses layer before pooling
h = dense_layer(inputs=h, output_size=d_theta, activation=tf.nn.elu, use_bias=True, name='pre_process_dense_1')
h = dense_layer(inputs=h, output_size=d_theta, activation=tf.nn.elu, use_bias=True, name='pre_process_dense_2')
# Pool across dimensions
nu = tf.expand_dims(tf.reduce_mean(h, axis=0), axis=0)
post_processed = _post_process(nu, d_psi)
# Compute means and log variances for the parameter
psi = {
'mu': dense_layer(inputs=post_processed, output_size=d_psi, activation=None, use_bias=True, name='psi_mean'),
'log_variance': dense_layer(inputs=post_processed, output_size=d_psi, activation=None, use_bias=True,
name='psi_log_var')
}
psi['psi_samples'] = sample_normal(psi['mu'], psi['log_variance'], num_samples)
return psi
def vae_prior(noise, is_training):
regularizer = tf.contrib.layers.l2_regularizer(scale=1e-10)
initializer = tf.contrib.layers.xavier_initializer()
with tf.variable_scope("prior", reuse=None):
# layer 1
with tf.variable_scope('layer_1'):
output = tf.layers.conv2d(noise, filters=16, kernel_size=3, padding='same',
kernel_initializer=initializer,
kernel_regularizer=regularizer, name='conv_1')
output = tf.layers.batch_normalization(output, training=is_training, name='bn_1')
output_shortcut = tf.nn.leaky_relu(output, name='relu_1')
for i in range(3):
with tf.variable_scope('layer_%d' % (i * 2 + 2)):
output = tf.layers.conv2d(output_shortcut, num_feature, KernelSize, padding='same',
kernel_initializer=initializer,
kernel_regularizer=regularizer, name=('conv_%d' % (i * 2 + 2)))
output = tf.layers.batch_normalization(output, training=is_training, name=('bn_%d' % (i * 2 + 2)))
output = tf.nn.leaky_relu(output, name=('relu_%d' % (i * 2 + 2)))
with tf.variable_scope('layer_%d' % (i * 2 + 3)):
output = tf.layers.conv2d(output, num_feature, KernelSize, padding='same',
kernel_initializer=initializer,
kernel_regularizer=regularizer, name=('conv_%d' % (i * 2 + 3)))
output = tf.layers.batch_normalization(output, training=is_training, name=('bn_%d' % (i * 2 + 3)))
output = tf.nn.leaky_relu(output, name=('relu_%d' % (i * 2 + 3)))
output_shortcut = tf.add(output_shortcut, output) # shortcut
with tf.variable_scope('layer_final'):
output = tf.layers.conv2d(output_shortcut, 1, KernelSize, padding='same',
kernel_initializer=initializer,
kernel_regularizer=regularizer, name='conv_26')
output = tf.layers.batch_normalization(output, training=is_training, name='bn_26')
mn = tf.layers.conv2d(output, filters=1, kernel_size=3, strides=1, padding="SAME", activation=None)
sd = tf.layers.conv2d(output, filters=1, kernel_size=3, strides=1, padding="SAME", activation=None)
z = sample_normal(mn, sd, FLAGS.num_samples)
return z, mn, sd