def test_repetitions_increases_number_parameters(self):
input_shape = (32, 32, 3)
num_classes = 10
small_resnet = resnet_models.create_resnet(
input_shape, num_classes, repetitions=[1, 1])
big_resnet = resnet_models.create_resnet(
input_shape, num_classes, repetitions=[2, 2])
self.assertLess(small_resnet.count_params(), big_resnet.count_params())
def test_basic_block_has_fewer_parameters_than_bottleneck(self):
input_shape = (32, 32, 3)
num_classes = 10
basic_resnet = resnet_models.create_resnet(
input_shape,
num_classes,
residual_block=resnet_models.ResidualBlock.basic)
bottleneck_resnet = resnet_models.create_resnet(
input_shape,
num_classes,
residual_block=resnet_models.ResidualBlock.bottleneck)
self.assertLess(basic_resnet.count_params(),
bottleneck_resnet.count_params())
def test_initial_kernel_size_with_negative_values_raises(self):
with self.assertRaisesRegex(
ValueError, 'initial_kernel_size must be an iterable of length 2 '
'containing only positive integers'):
resnet_models.create_resnet(
input_shape=(32, 32, 3), initial_kernel_size=(3, -1))
def test_initial_strides_with_length_not_2_raises(self):
with self.assertRaisesRegex(
ValueError, 'initial_strides must be an iterable of length 2 containing'
' only positive integers'):
resnet_models.create_resnet(
input_shape=(32, 32, 3), initial_strides=(3, 3, 4))
def test_negative_initial_filters_raises(self):
with self.assertRaisesRegex(ValueError,
'initial_filters must be a positive integer'):
resnet_models.create_resnet(input_shape=(32, 32, 3), initial_filters=-2)
def test_repetitions_with_negative_values_raises(self):
with self.assertRaisesRegex(
ValueError, 'repetitions must be None or an iterable containing '
'positive integers'):
resnet_models.create_resnet(input_shape=(32, 32, 3), repetitions=[2, -1])
def test_unsupported_residual_block_raises(self):
with self.assertRaisesRegex(
ValueError, 'residual_block must be of type `ResidualBlock`'):
resnet_models.create_resnet(
input_shape=(32, 32, 3), residual_block='bad_block')
def test_negative_num_classes_raises(self):
with self.assertRaisesRegex(ValueError,
'num_classes must be a positive integer'):
resnet_models.create_resnet(input_shape=(32, 32, 3), num_classes=-5)
def test_non_length_3_input_shape_raises(self):
with self.assertRaisesRegex(
ValueError, 'input_shape must be an iterable of length 3 containing '
'only positive integers'):
resnet_models.create_resnet(input_shape=(10, 10))
def test_resnet_constructs_with_group_norm(self):
group_resnet = resnet_models.create_resnet(
input_shape=(32, 32, 3),
num_classes=10,
norm_layer=resnet_models.NormLayer.group_norm)
self.assertIsInstance(group_resnet, tf.keras.Model)
def test_unsupported_norm_raises(self):
with self.assertRaisesRegex(
ValueError, 'norm_layer must be of type `NormLayer`'):
resnet_models.create_resnet(
input_shape=(32, 32, 3), norm_layer='bad_norm')
