def test_reuse_model(self, policy, is_training, classnet_type):
config = saccader_classnet_config.get_config()
num_times = 2
image_shape = (100, 100, 3)
num_classes = 10
config.num_classes = num_classes
config.num_times = num_times
config.classnet_type = classnet_type
batch_size = 3
images = tf.constant(
np.random.rand(*((batch_size,) + image_shape)), dtype=tf.float32)
model = saccader_classnet.SaccaderClassNet(config)
logits1 = model(images, images, num_times=num_times,
is_training_saccader=is_training,
is_training_classnet=is_training,
policy=policy)[0]
num_params = len(tf.all_variables())
l2_loss1 = tf.losses.get_regularization_loss()
# Build twice with different num_times.
logits2 = model(images, images, num_times=num_times+1,
is_training_saccader=is_training,
is_training_classnet=is_training,
policy=policy)[0]
l2_loss2 = tf.losses.get_regularization_loss()
# Ensure variables are reused.
self.assertLen(tf.all_variables(), num_params)
init = tf.global_variables_initializer()
self.evaluate(init)
logits1, logits2 = self.evaluate((logits1, logits2))
l2_loss1, l2_loss2 = self.evaluate((l2_loss1, l2_loss2))
np.testing.assert_almost_equal(l2_loss1, l2_loss2, decimal=5)
def test_build(self, policy, is_training, classnet_type):
config = saccader_classnet_config.get_config()
num_times = 2
image_shape = (100, 100, 3)
num_classes = 10
config.num_classes = num_classes
config.num_times = num_times
config.classnet_type = classnet_type
batch_size = 3
images = tf.constant(np.random.rand(*((batch_size, ) + image_shape)),
dtype=tf.float32)
model = saccader_classnet.SaccaderClassNet(config)
logits = model(images,
images,
num_times=num_times,
is_training_saccader=is_training,
is_training_classnet=is_training,
policy=policy)[0]
init_op = model.init_op
self.evaluate(init_op)
self.assertEqual((batch_size, num_classes),
self.evaluate(logits).shape)