def test_build_matrices_jacobian_random(self, size, channels):
"""Tests the Jacobian of the build_matrices function."""
tensor_shape = np.random.randint(size, 6, size=3)
image_init = np.random.uniform(0.0,
1.0,
size=tensor_shape.tolist() + [channels])
with self.subTest(name="laplacian"):
self.assert_jacobian_is_correct_fn(
lambda image: matting.build_matrices(image, size=size)[0],
[image_init])
with self.subTest(name="pseudo_inverse"):
self.assert_jacobian_is_correct_fn(
lambda image: matting.build_matrices(image, size=size)[1],
[image_init])
def test_loss_same_images(self, size):
"""Tests that passing same images results in a loss close to 0.0."""
tensor_shape = np.random.randint(size, 6, size=3).tolist()
image = np.random.uniform(0.0, 1.0, size=tensor_shape + [1])
laplacian, _ = matting.build_matrices(image, size=size)
loss = matting.loss(image, laplacian)
self.assertAllClose(loss, 0.0, atol=1e-4)
def test_loss_positive(self, size):
"""Tests that the loss is always greater or equal to 0.0."""
tensor_shape = np.random.randint(size, 6, size=3).tolist()
image = tf.random.uniform(minval=0.0, maxval=1.0, shape=tensor_shape + [3])
matte = tf.random.uniform(minval=0.0, maxval=1.0, shape=tensor_shape + [1])
laplacian, _ = matting.build_matrices(image, size=size)
loss = matting.loss(matte, laplacian)
self.assertAllGreaterEqual(loss, 0.0)
def test_linear_coefficients_reconstruction_opposite_images(self, size):
"""Tests that the matte can be reconstructed by using the coefficients ."""
tensor_shape = np.random.randint(size, 6, size=3).tolist()
image = np.random.uniform(0.0, 1.0, size=tensor_shape + [1])
_, pseudo_inverse = matting.build_matrices(image, size=size)
a, b = matting.linear_coefficients(1.0 - image, pseudo_inverse)
reconstructed = matting.reconstruct(image, a, b)
self.assertAllClose(1.0 - image, reconstructed, atol=1e-4)
def test_build_matrices_laplacian_versions(self, size, channels):
"""Compares two ways of computing the laplacian matrix."""
tensor_shape = np.random.randint(size, 6, size=3)
image_init = np.random.uniform(
0.0, 1.0, size=tensor_shape.tolist() + [channels])
image = tf.convert_to_tensor(value=image_init)
laplacian_v1, _ = matting.build_matrices(image, size=size)
laplacian_v2 = _laplacian_matrix(image, size=size)
self.assertAllClose(laplacian_v1, laplacian_v2)
def test_build_matrices_jacobian_random(self, size, channels):
"""Tests the Jacobian of the build_matrices function."""
tensor_shape = np.random.randint(size, 6, size=3)
image_init = np.random.uniform(
0.0, 1.0, size=tensor_shape.tolist() + [channels])
image = tf.convert_to_tensor(value=image_init)
laplacian, pseudo_inverse = matting.build_matrices(image, size=size)
with self.subTest(name="laplacian"):
self.assert_jacobian_is_correct(image, image_init, laplacian)
with self.subTest(name="pseudo_inverse"):
self.assert_jacobian_is_correct(image, image_init, pseudo_inverse)
def test_build_matrices_laplacian_zero_rows_and_columns(self, size, channels):
"""Tests that the laplacian matrix rows and columns sum to zero."""
tensor_shape = np.random.randint(size, 6, size=3)
image_init = np.random.uniform(
0.0, 1.0, size=tensor_shape.tolist() + [channels])
image = tf.convert_to_tensor(value=image_init)
laplacian, _ = matting.build_matrices(image, size=size)
rows = tf.reduce_sum(input_tensor=laplacian, axis=-2)
columns = tf.reduce_sum(input_tensor=laplacian, axis=-1)
with self.subTest(name="rows"):
self.assertAllClose(rows, tf.zeros_like(rows))
with self.subTest(name="columns"):
self.assertAllClose(columns, tf.zeros_like(columns))
def test_build_matrices_raised(self, error_msg, size, *shapes):
"""Tests that the shape exceptions are properly raised."""
build_matrices = lambda image: matting.build_matrices(image, size=size)
self.assert_exception_is_raised(build_matrices, error_msg, shapes)
def test_build_matrices_not_raised(self, size, *shapes):
"""Tests that the shape exceptions are not raised."""
build_matrices = lambda image: matting.build_matrices(image, size=size)
self.assert_exception_is_not_raised(build_matrices, shapes)
