def test_group_norm_connection_exception(self):
message = "Cannot divide 11 input channels into 4 groups"
with self.assertRaisesRegexp(LayerConnectionError, message):
layers.join(
layers.Input((10, 10, 11)),
layers.GroupNorm(4),
)
message = ("Group normalization layer expects 4 "
"dimensional input, got 3 instead.")
with self.assertRaisesRegexp(LayerConnectionError, message):
layers.join(
layers.Input((10, 12)),
layers.GroupNorm(4),
)
def test_group_norm_unknown_shape(self):
network = layers.join(
layers.Input((None, None, 16)),
layers.GroupNorm(4),
)
input = np.random.random((7, 6, 6, 16))
actual_output = self.eval(network.output(input))
self.assertEqual(actual_output.shape, (7, 6, 6, 16))
def test_group_norm(self):
network = layers.join(
layers.Input((10, 10, 12)),
layers.GroupNorm(4),
)
self.assertShapesEqual(network.input_shape, (None, 10, 10, 12))
self.assertShapesEqual(network.output_shape, (None, 10, 10, 12))
input = np.random.random((7, 10, 10, 12))
actual_output = self.eval(network.output(input))
self.assertEqual(actual_output.shape, (7, 10, 10, 12))
def test_group_norm_weight_init_exception(self):
network = layers.join(
layers.Input((10, 10, None)),
layers.GroupNorm(4),
)
message = ("Cannot initialize variables when number of "
"channels is unknown.")
with self.assertRaisesRegexp(WeightInitializationError, message):
network.create_variables()
with self.assertRaisesRegexp(WeightInitializationError, message):
network.outputs
def test_group_norm_repr(self):
layer = layers.GroupNorm(4)
self.assertEqual(
str(layer),
("GroupNorm(n_groups=4, beta=Constant(0), "
"gamma=Constant(1), epsilon=1e-05, name='group-norm-1')"))