def test_should_not_add_dropout_not_in_training_mode_using_constant(self):
with tf.Graph().as_default():
encoder_inputs = tf.ones((1, 8, 8, 3))
encoder_layer_specs = [5, 10]
decoder_layer_specs = [(5, 0.5), (3, 0.0)]
outputs = create_encoder_decoder(encoder_inputs,
encoder_layer_specs,
decoder_layer_specs,
is_training=False)
with tf.Session() as session:
session.run(tf.global_variables_initializer())
# without dropout, the outputs are expected to the same for every run
assert_all_close(session.run(outputs), session.run(outputs))
def test_should_add_dropout_in_training_mode_using_placeholder(self):
with tf.Graph().as_default():
is_training = tf.placeholder(tf.bool)
encoder_inputs = tf.ones((1, 8, 8, 3))
encoder_layer_specs = [5, 10]
decoder_layer_specs = [(5, 0.5), (3, 0.0)]
outputs = create_encoder_decoder(encoder_inputs,
encoder_layer_specs,
decoder_layer_specs,
is_training=is_training)
with tf.Session() as session:
session.run(tf.global_variables_initializer())
feed_dict = {is_training: True}
# with dropout, the outputs are expected to be different for every run
assert_all_not_close(session.run(outputs, feed_dict=feed_dict),
session.run(outputs, feed_dict=feed_dict))
def test_should_allow_undefined_batch_size(self):
with tf.Graph().as_default():
input_shape = [None, 8, 8, 3]
encoder_inputs = tf.placeholder(tf.float32, input_shape)
encoder_layer_specs = [5, 10]
decoder_layer_specs = [(5, 0.5), (3, 0.0)]
outputs = create_encoder_decoder(encoder_inputs,
encoder_layer_specs,
decoder_layer_specs,
is_training=False)
assert outputs.get_shape().as_list() == input_shape
with tf.Session() as session:
session.run(tf.global_variables_initializer())
feed_dict = {encoder_inputs: np.ones([1] + input_shape[1:])}
assert_all_close(session.run(outputs, feed_dict=feed_dict),
session.run(outputs, feed_dict=feed_dict))
