def testGradient(self):
N = 30
initval = np.arange(1, 0.5 * N * (N + 1) + 1)
v = tf.Variable(initval[None, :])
with tf.Session(''):
f = (vec_to_tri(v) * np.random.randn(N, N)) ** 2.0 # Some function involving vec_to_tri
self.assertLess(compute_gradient_error(v, [1, len(initval)], f, [1, N, N]), 10 ** -10)
def testTriToVec(self):
mats = [
np.arange(1, 4)[None, :],
np.vstack((np.arange(1, 4), np.arange(3, 0, -1))), # Currently, only do matrices
np.arange(1, 16)[None, :]
]
with tf.Session(''):
for m in mats:
# The ops are each others' inverse.
self.assertTrue(np.all(tri_to_vec(vec_to_tri(m)).eval() == m))
def testBasicFunctionality(self):
N = 3
D = 3
reference_matrices = self.getExampleMatrices(D, N)
input_vector_tensor = tf.constant(
self.referenceInverse(reference_matrices))
test_matrices_tensor = vec_to_tri(input_vector_tensor, N)
test_matrices = tf.Session().run(test_matrices_tensor)
np.testing.assert_array_almost_equal(reference_matrices, test_matrices)
def testDifferentiable(self):
N = 3
D = 3
reference_matrices = self.getExampleMatrices(D, N)
input_vector_array = self.referenceInverse(reference_matrices)
input_vector_tensor = tf.constant(input_vector_array)
test_matrices_tensor = vec_to_tri(input_vector_tensor, N)
reduced_sum = tf.reduce_sum(test_matrices_tensor)
gradient = tf.gradients(reduced_sum, input_vector_tensor)[0]
reference_gradient = np.ones_like(input_vector_array)
test_gradient = tf.Session().run(gradient)
np.testing.assert_array_almost_equal(reference_gradient, test_gradient)
def testErrorOnIncorrectSize(self):
with self.assertRaises(tf.errors.InvalidArgumentError):
self.sess.run(vec_to_tri(np.arange(1, 5)[None, :]))
def compare_op(v):
with tf.Session(''):
return np_vec_to_tri(v) == vec_to_tri(v).eval()
