def test_network_creation(self, input_size, model_id,
endpoint_filter_scale):
"""Test creation of RevNet family models."""
tf.keras.backend.set_image_data_format('channels_last')
network = revnet.RevNet(model_id=model_id)
inputs = tf.keras.Input(shape=(input_size, input_size, 3),
batch_size=1)
endpoints = network(inputs)
network.summary()
self.assertAllEqual([
1, input_size / 2**2, input_size / 2**2,
128 * endpoint_filter_scale
], endpoints['2'].shape.as_list())
self.assertAllEqual([
1, input_size / 2**3, input_size / 2**3,
256 * endpoint_filter_scale
], endpoints['3'].shape.as_list())
self.assertAllEqual([
1, input_size / 2**4, input_size / 2**4,
512 * endpoint_filter_scale
], endpoints['4'].shape.as_list())
self.assertAllEqual([
1, input_size / 2**5, input_size / 2**5,
832 * endpoint_filter_scale
], endpoints['5'].shape.as_list())
def test_input_specs(self, input_dim):
"""Test different input feature dimensions."""
tf.keras.backend.set_image_data_format('channels_last')
input_specs = tf.keras.layers.InputSpec(shape=[None, None, None, input_dim])
network = revnet.RevNet(model_id=56, input_specs=input_specs)
inputs = tf.keras.Input(shape=(128, 128, input_dim), batch_size=1)
_ = network(inputs)
def test_serialize_deserialize(self):
# Create a network object that sets all of its config options.
kwargs = dict(
model_id=56,
activation='relu',
use_sync_bn=False,
norm_momentum=0.99,
norm_epsilon=0.001,
kernel_initializer='VarianceScaling',
kernel_regularizer=None,
)
network = revnet.RevNet(**kwargs)
expected_config = dict(kwargs)
self.assertEqual(network.get_config(), expected_config)
# Create another network object from the first object's config.
new_network = revnet.RevNet.from_config(network.get_config())
# Validate that the config can be forced to JSON.
_ = new_network.to_json()
# If the serialization was successful, the new config should match the old.
self.assertAllEqual(network.get_config(), new_network.get_config())
