def build(self, placeholders, n_labels, phase):
x, y, lr = fetch_placeholders(placeholders)
n_conjunctions_arr = self.config['n_conjunctions_arr']
conjunctions_depth_arr = self.config['conjunctions_depth_arr']
n_formulas = self.config['n_formulas']
total_number_of_literals = compute_total_number_of_literals(n_formulas, n_conjunctions_arr, conjunctions_depth_arr)
total_number_of_conjunctions = compute_total_number_of_conjunctions(n_formulas, n_conjunctions_arr)
out = tf.layers.dense(x, total_number_of_literals, activation=tf.tanh, kernel_initializer=tf.contrib.layers.xavier_initializer())
out = tf.layers.dense(out, total_number_of_conjunctions, activation=tf.tanh, kernel_initializer=tf.contrib.layers.xavier_initializer())
out = tf.layers.dense(out, n_formulas, activation=tf.tanh, kernel_initializer=tf.contrib.layers.xavier_initializer())
output = tf.layers.dense(out, n_labels, activation=None, kernel_initializer=tf.contrib.layers.xavier_initializer())
with tf.variable_scope('loss'):
if n_labels > 1:
loss = tf.losses.softmax_cross_entropy(onehot_labels=y, logits=output)
output = tf.nn.softmax(output)
else:
loss = tf.losses.sigmoid_cross_entropy(multi_class_labels=y, logits=output)
output = tf.nn.sigmoid(output)
loss += tf.losses.get_regularization_loss()
if phase == 'train':
print('Build optimizer')
optimizer = tf.train.AdamOptimizer(learning_rate=lr).minimize(loss)
else:
optimizer = None
build_dict = {'optimizer': optimizer, 'loss': loss, 'output': output, 'features': None, 'input_masks': None}
return build_dict
def build(self, placeholders, n_labels, phase):
x, y, lr = fetch_placeholders(placeholders)
input_dim = self.config['input_dim']
n_conjunctions_arr = self.config['n_conjunctions_arr']
conjunctions_depth_arr = self.config['conjunctions_depth_arr']
n_formulas = self.config['n_formulas']
total_number_of_literals = compute_total_number_of_literals(n_formulas, n_conjunctions_arr, conjunctions_depth_arr)
total_number_of_conjunctions = compute_total_number_of_conjunctions(n_formulas, n_conjunctions_arr)
n_literals_per_formula_arr = compute_n_literals_per_formula(n_conjunctions_arr, conjunctions_depth_arr)
conjunctions_indicator_matrix = create_conjunctions_indicator_matrix(total_number_of_literals, total_number_of_conjunctions, conjunctions_depth_arr)
formulas_indicator_matrix = create_formulas_indicator_matrix(n_formulas, n_conjunctions_arr)
conjunctions_indicator_sparse_matrix = denseNDArrayToSparseTensor(conjunctions_indicator_matrix)
formulas_indicator_sparse_matrix = denseNDArrayToSparseTensor(formulas_indicator_matrix)
and_bias = np.sum(conjunctions_indicator_matrix, axis=0)
or_bias = np.sum(formulas_indicator_matrix, axis=0)
learnable_binary_mask, literals_random_mask, elastic_net_reg = feature_selection(input_dim, self.config['keep_feature_prob_arr'], n_literals_per_formula_arr, n_formulas, self.config['elastic_net_beta'])
bias = tf.get_variable('literals_bias', [total_number_of_literals], initializer=tf.zeros_initializer())
weight = tf.get_variable('literals_weights', [input_dim, total_number_of_literals], initializer=tf.contrib.layers.xavier_initializer())
weight = tf.multiply(weight, tf.constant(literals_random_mask))
out_literals = tf.tanh(tf.add(tf.matmul(x, tf.multiply(weight, learnable_binary_mask)), bias))
out_conjunctions = and_operator(tf_matmul_sparse_dense(conjunctions_indicator_sparse_matrix, out_literals), d=and_bias)
out_DNNFs = or_operator(tf_matmul_sparse_dense(formulas_indicator_sparse_matrix, out_conjunctions), d=or_bias)
out_DNNFs = tf.reshape(out_DNNFs, shape=(-1, n_formulas))
loc = broadcast_exp(x, n_formulas, self.config['input_dim'])
temperature = tf.get_variable('temperature', initializer=tf.constant(value=2.))
loc = tf.nn.softmax(tf.sigmoid(temperature) * loc)
out_DNNFs = tf.multiply(out_DNNFs, loc)
output = tf.layers.dense(out_DNNFs, n_labels, activation=None, kernel_initializer=tf.contrib.layers.xavier_initializer())
with tf.variable_scope('loss'):
if n_labels > 1:
loss = tf.losses.softmax_cross_entropy(onehot_labels=y, logits=output)
output = tf.nn.softmax(output)
else:
loss = tf.losses.sigmoid_cross_entropy(multi_class_labels=y, logits=output)
output = tf.nn.sigmoid(output)
loss += elastic_net_reg
loss += tf.losses.get_regularization_loss()
if phase == 'train':
print('Build optimizer')
optimizer = tf.train.AdamOptimizer(learning_rate=lr).minimize(loss)
else:
optimizer = None
build_dict = {'optimizer': optimizer, 'loss': loss, 'output': output, 'features': None, 'input_masks': None}
return build_dict