def _test_random_rotation(self, dtype):
aa = np.random.rand(4, 3).astype(dtype)
aa_tensor = tf.constant(aa)
R_tensor = ops.angle_axis_to_rotation_matrix(aa_tensor)
aa2_tensor = ops.rotation_matrix_to_angle_axis(R_tensor)
aa2 = aa2_tensor.eval()
err = np.sum(np.abs(aa - aa2))
print('error', err, flush=True)
self.assertLess(err, 1e-5)
def _test_zero_rotation(self, dtype):
aa = np.zeros(3).astype(dtype)
aa_tensor = tf.constant(aa)
R_tensor = ops.angle_axis_to_rotation_matrix(aa_tensor)
R = R_tensor.eval()
I = np.eye(3)
err = np.sum(np.abs(R - I))
print('error', err, flush=True)
self.assertLess(err, 1e-5)
def test_shape(self):
with self.test_session(use_gpu=False,
force_gpu=False,
**session_params):
input1 = np.empty((8, 3))
input2 = np.empty((2, 4, 3))
input3 = np.empty((2, 2, 2, 3))
inputs = (input1, input2, input3)
expected_shape = [8, 3, 3]
for i in inputs:
output_tensor = ops.angle_axis_to_rotation_matrix(i)
out_shape = output_tensor.get_shape().as_list()
self.assertAllEqual(out_shape, expected_shape)
def _test_grad(self, dtype):
aa = np.random.rand(5, 3).astype(dtype)
shape = aa.shape
data = tf.constant(aa)
output = ops.angle_axis_to_rotation_matrix(data)
err = tf.test.compute_gradient_error(data,
shape,
output,
output.get_shape().as_list(),
x_init_value=aa)
print('error', err, flush=True)
self.assertLess(err, 1e-4)
grad = tf.test.compute_gradient(data,
shape,
output,
output.get_shape().as_list(),
x_init_value=aa,
delta=0.0001)
diff = np.abs(grad[0] - grad[1])