def test_serialize_deserialize(self):
# Create a network object that sets all of its config options.
kwargs = dict(
input_specs=hardnet.HardNet(model_id=70).output_specs,
model_id=70,
activation='relu',
use_sync_bn=False,
norm_momentum=0.99,
norm_epsilon=0.001,
kernel_initializer='VarianceScaling',
kernel_regularizer=None,
bias_regularizer=None,
)
network = hardnet_decoder.HardNetDecoder(**kwargs)
expected_config = dict(kwargs)
self.assertEqual(network.get_config(), expected_config)
# Create another network object from the first object's config.
new_network = hardnet_decoder.HardNetDecoder.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())
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 = hardnet.HardNet(model_id=70, input_specs=input_specs)
inputs = tf.keras.Input(shape=(256, 256, input_dim), batch_size=1)
_ = network(inputs)
def test_network_creation(self, model_id, input_size):
"""Test creation of HardnetDecoder."""
tf.keras.backend.set_image_data_format('channels_last')
inputs = tf.keras.Input(shape=(input_size, input_size, 3),
batch_size=1)
backbone = hardnet.HardNet(model_id=70)
network = hardnet_decoder.HardNetDecoder(
model_id=model_id, input_specs=backbone.output_specs)
endpoints = backbone(inputs)
feats = network(endpoints)
feats = list(feats.values())[0] # only one output
# number of stride downs at stem = 2, size of last downsampling channel at stem = 48
self.assertAllEqual([1, input_size // 2**2, input_size // 2**2, 48],
feats.shape.as_list())
def test_network_creation(self, input_size):
"""Test creation of HardNet family models."""
tf.keras.backend.set_image_data_format('channels_last')
network = hardnet.HardNet(model_id=70)
inputs = tf.keras.Input(shape=(input_size, input_size, 3),
batch_size=1)
endpoints = network(inputs)
self.assertAllEqual([1, input_size / 2**2, input_size / 2**2, 48],
endpoints['0'].shape.as_list())
self.assertAllEqual([1, input_size / 2**3, input_size / 2**3, 78],
endpoints['1'].shape.as_list())
self.assertAllEqual([1, input_size / 2**4, input_size / 2**4, 160],
endpoints['2'].shape.as_list())
self.assertAllEqual([1, input_size / 2**5, input_size / 2**5, 214],
endpoints['3'].shape.as_list())
self.assertAllEqual([1, input_size / 2**6, input_size / 2**6, 286],
endpoints['4'].shape.as_list())