def test_build_graph(self, dataset_name, regression):
"""Test whether build_graph works as expected."""
data_x, data_y, _ = data_utils.load_dataset(dataset_name)
data_gen = data_utils.split_training_dataset(
data_x, data_y, n_splits=5, stratified=not regression)
(x_train, y_train), (x_validation, y_validation) = next(data_gen)
sess = tf.InteractiveSession()
graph_tensors_and_ops, metric_scores = graph_builder.build_graph(
x_train=x_train,
y_train=y_train,
x_test=x_validation,
y_test=y_validation,
activation='exu',
learning_rate=1e-3,
batch_size=256,
shallow=True,
regression=regression,
output_regularization=0.1,
dropout=0.1,
decay_rate=0.999,
name_scope='model',
l2_regularization=0.1)
# Run initializer ops
sess.run(tf.global_variables_initializer())
sess.run([
graph_tensors_and_ops['iterator_initializer'],
graph_tensors_and_ops['running_vars_initializer']
])
for _ in range(2):
sess.run(graph_tensors_and_ops['train_op'])
self.assertIsInstance(metric_scores['train'](sess), float)
sess.close()
def _create_computation_graph(
x_train, y_train, x_validation,
y_validation, batch_size
):
"""Build the computation graph."""
graph_tensors_and_ops = []
metric_scores = []
for n in range(FLAGS.n_models):
graph_tensors_and_ops_n, metric_scores_n = graph_builder.build_graph(
x_train=x_train,
y_train=y_train,
x_test=x_validation,
y_test=y_validation,
activation=FLAGS.activation,
learning_rate=FLAGS.learning_rate,
batch_size=batch_size,
shallow=FLAGS.shallow,
output_regularization=FLAGS.output_regularization,
l2_regularization=FLAGS.l2_regularization,
dropout=FLAGS.dropout,
num_basis_functions=FLAGS.num_basis_functions,
units_multiplier=FLAGS.units_multiplier,
decay_rate=FLAGS.decay_rate,
feature_dropout=FLAGS.feature_dropout,
regression=FLAGS.regression,
use_dnn=FLAGS.use_dnn,
trainable=True,
name_scope=f'model_{n}')
graph_tensors_and_ops.append(graph_tensors_and_ops_n)
metric_scores.append(metric_scores_n)
return graph_tensors_and_ops, metric_scores