def test_perturb(): C = SO2.exp(torch.Tensor([np.pi / 4])) C_copy = copy.deepcopy(C) phi = torch.Tensor([0.1]) C.perturb(phi) assert utils.allclose( C.as_matrix(), (SO2.exp(phi).dot(C_copy)).as_matrix())
def test_left_jacobians_batch(): phis = torch.Tensor([0., np.pi / 2]) left_jacobian = SO2.left_jacobian(phis) inv_left_jacobian = SO2.inv_left_jacobian(phis) assert utils.allclose(torch.bmm(left_jacobian, inv_left_jacobian), torch.eye(2).unsqueeze_(dim=0).expand(2, 2, 2))
def test_left_jacobians(): phi_small = torch.Tensor([0.]) phi_big = torch.Tensor([np.pi / 2]) assert torch.allclose(torch.tensor([[1, 0], [0, 1]], dtype=torch.float32), SO2.left_jacobian(phi_small)) print(SO2.left_jacobian(phi_big))
def test_from_matrix(): C_good = SO2.from_matrix(torch.eye(2)) print(C_good.mat) print(SO2.is_valid_matrix(C_good.mat)) C_bad = SO2.from_matrix(torch.eye(2).add_(1e-3), normalize=True) print(isinstance(C_bad, SO2)) print(C_bad.mat)
def test_identity_batch(): C = SO2.identity(5) assert isinstance(C, SO2) \ and C.mat.dim() == 3 \ and C.mat.shape == (5, 2, 2) C_copy = SO2.identity(5, copy=True) assert isinstance(C_copy, SO2) \ and C_copy.mat.dim() == 3 \ and C_copy.mat.shape == (5, 2, 2)
def test_normalize_batch(): C = SO2.exp(torch.Tensor([-1., 0., 1.])) assert SO2.is_valid_matrix(C.mat).all() C.mat.add_(0.1) assert (SO2.is_valid_matrix(C.mat) == torch.ByteTensor([0, 0, 0])).all() C.normalize(inds=[0, 2]) assert (SO2.is_valid_matrix(C.mat) == torch.ByteTensor([1, 0, 1])).all() C.normalize() assert SO2.is_valid_matrix(C.mat).all()
def test_from_matrix(): C_good = SO2.from_matrix(torch.eye(2)) assert isinstance(C_good, SO2) \ and C_good.mat.dim() == 2 \ and C_good.mat.shape == (2, 2) \ and SO2.is_valid_matrix(C_good.mat).all() C_bad = SO2.from_matrix(torch.eye(2).add_(1e-3), normalize=True) assert isinstance(C_bad, SO2) \ and C_bad.mat.dim() == 2 \ and C_bad.mat.shape == (2, 2) \ and SO2.is_valid_matrix(C_bad.mat).all()
def test_dot_batch(): C1 = SO2(torch.Tensor([[0, -1], [1, 0]]).expand(5, 2, 2)) C2 = SO2(torch.Tensor([[-1, 0], [0, -1]])) pt1 = torch.Tensor([1, 2]) pt2 = torch.Tensor([4, 5]) pt3 = torch.Tensor([7, 8]) pts = torch.cat([pt1.unsqueeze(dim=0), pt2.unsqueeze(dim=0), pt3.unsqueeze(dim=0)], dim=0) # 3x2 ptsbatch = pts.unsqueeze(dim=0).expand(5, 3, 2) C1C1 = torch.bmm(C1.mat, C1.mat) C1C1_SO2 = C1.dot(C1).mat assert C1C1_SO2.shape == C1.mat.shape and utils.allclose(C1C1_SO2, C1C1) C1C2 = torch.matmul(C1.mat, C2.mat) C1C2_SO2 = C1.dot(C2).mat assert C1C2_SO2.shape == C1.mat.shape and utils.allclose(C1C2_SO2, C1C2) C1pt1 = torch.matmul(C1.mat, pt1) C1pt1_SO2 = C1.dot(pt1) assert C1pt1_SO2.shape == (C1.mat.shape[0], pt1.shape[0]) \ and utils.allclose(C1pt1_SO2, C1pt1) C1pt2 = torch.matmul(C1.mat, pt2) C1pt2_SO2 = C1.dot(pt2) assert C1pt2_SO2.shape == (C1.mat.shape[0], pt2.shape[0]) \ and utils.allclose(C1pt2_SO2, C1pt2) C1pts = torch.matmul(C1.mat, pts.transpose(1, 0)).transpose(2, 1) C1pts_SO2 = C1.dot(pts) assert C1pts_SO2.shape == (C1.mat.shape[0], pts.shape[0], pts.shape[1]) \ and utils.allclose(C1pts_SO2, C1pts) \ and utils.allclose(C1pt1, C1pts[:, 0, :]) \ and utils.allclose(C1pt2, C1pts[:, 1, :]) C1ptsbatch = torch.bmm(C1.mat, ptsbatch.transpose(2, 1)).transpose(2, 1) C1ptsbatch_SO2 = C1.dot(ptsbatch) assert C1ptsbatch_SO2.shape == ptsbatch.shape \ and utils.allclose(C1ptsbatch_SO2, C1ptsbatch) \ and utils.allclose(C1pt1, C1ptsbatch[:, 0, :]) \ and utils.allclose(C1pt2, C1ptsbatch[:, 1, :]) C2ptsbatch = torch.matmul(C2.mat, ptsbatch.transpose(2, 1)).transpose(2, 1) C2ptsbatch_SO2 = C2.dot(ptsbatch) assert C2ptsbatch_SO2.shape == ptsbatch.shape \ and utils.allclose(C2ptsbatch_SO2, C2ptsbatch) \ and utils.allclose(C2.dot(pt1), C2ptsbatch[:, 0, :]) \ and utils.allclose(C2.dot(pt2), C2ptsbatch[:, 1, :])
def test_perturb_batch(): C = SO2.exp(torch.Tensor([-1., 0., 1.])) C_copy1 = copy.deepcopy(C) C_copy2 = copy.deepcopy(C) phi = torch.Tensor([0.1]) C_copy1.perturb(phi) assert utils.allclose(C_copy1.as_matrix(), (SO2.exp(phi).dot(C)).as_matrix()) phis = torch.Tensor([0.1, 0.2, 0.3]) C_copy2.perturb(phis) assert utils.allclose(C_copy2.as_matrix(), (SO2.exp(phis).dot(C)).as_matrix())
def test_from_matrix_batch(): C_good = SO2.from_matrix(torch.eye(2).repeat(5, 1, 1)) assert isinstance(C_good, SO2) \ and C_good.mat.dim() == 3 \ and C_good.mat.shape == (5, 2, 2) \ and SO2.is_valid_matrix(C_good.mat).all() C_bad = copy.deepcopy(C_good.mat) C_bad[3].add_(0.1) C_bad = SO2.from_matrix(C_bad, normalize=True) assert isinstance(C_bad, SO2) \ and C_bad.mat.dim() == 3 \ and C_bad.mat.shape == (5, 2, 2) \ and SO2.is_valid_matrix(C_bad.mat).all()
def test_left_jacobians(): phi_small = torch.Tensor([0.]) phi_big = torch.Tensor([np.pi / 2]) left_jacobian_small = SO2.left_jacobian(phi_small) inv_left_jacobian_small = SO2.inv_left_jacobian(phi_small) assert utils.allclose( torch.mm(left_jacobian_small, inv_left_jacobian_small), torch.eye(2)) left_jacobian_big = SO2.left_jacobian(phi_big) inv_left_jacobian_big = SO2.inv_left_jacobian(phi_big) assert utils.allclose( torch.mm(left_jacobian_big, inv_left_jacobian_big), torch.eye(2))
def propagate_cov(self, u_odo, u_fog, dt, G_cor): F = self.F G = self.G v, _ = self.encoder2speed(u_odo, dt) J = torch.Tensor([[0, 1], [-1, 0]]) Rot = torch.Tensor([[self.x[5].cos(), self.x[5].sin()], [-self.x[5].sin(), self.x[5].cos()]]) F[:2, 5] = J.mv(Rot.mv(torch.Tensor([v * dt, 0]))) A = torch.zeros(2, 3) A[0, 0] = 1 A[0, 1] = self.x[3].sin() * self.x[4].tan() A[0, 2] = self.x[3].cos() * self.x[4].tan() A[1, 1] = self.x[3].cos() A[1, 2] = -self.x[3].sin() F[3:5, 6:9] = A * dt B = torch.Tensor([ [1 / 2, 1 / 2], # v_l, v_r to v_forward [0, 0] ]) F[3, 3] = 1 + self.x[7] * self.x[3].sin() * self.x[4].tan() * dt F[3, 4] = 1 + (self.x[7] * self.x[3].sin() + self.x[8] * self.x[3].cos()) * self.x[4].tan() * dt F[4, 3] = 1 - self.x[8] * self.x[3].sin() * dt G[:2, :2] = SO2.from_angle( self.x[5].unsqueeze(0)).as_matrix().mm(B) * dt # add Jacobian correction Q = torch.zeros_like(self.P) for i in range(G_cor.shape[0]): Q += (G + G_cor[i]).mm(self.Q).mm((G + G_cor[i]).t()) self.P = F.mm(self.P).mm(F.t()) + Q
def test_dot(): C = SO2(torch.Tensor([[0, -1], [1, 0]])) pt = torch.Tensor([1, 2]) CC = C.mat.mm(C.mat) assert utils.allclose(C.dot(C).mat, CC) Cpt = C.mat.matmul(pt) assert utils.allclose(C.dot(pt), Cpt)
def test_from_angle_batch(): angles = torch.Tensor([np.pi / 2., np.pi]) assert torch.allclose( torch.tensor([ [[0, -1], [1, 0]], [[-1, 0], [0, 1]], ], dtype=torch.float32), SO2.from_angle(angles).mat)
def test_exp_log(): C_big = SO2.exp(torch.Tensor([np.pi / 4])) torch.testing.assert_close(torch.tensor([ [0.7071, -0.7071], [0.7071, 0.7071], ]), C_big.mat, rtol=1e-5, atol=1e-3) C_small = SO2.exp(torch.Tensor([0])) torch.testing.assert_close(torch.tensor([ [1, -0], [0, 1], ], dtype=torch.float32), C_small.mat, rtol=1e-5, atol=1e-3)
def test_wedge(): phi = torch.Tensor([1]) Phi = SO2.wedge(phi) assert (Phi == -Phi.t()).all()
def test_from_angle_to_angle_batch(): angles = torch.Tensor([-1., 0, 1.]) assert utils.allclose(SO2.from_angle(angles).to_angle(), angles)
def test_from_angle_to_angle(): angle = torch.Tensor([np.pi / 2.]) assert utils.allclose(SO2.from_angle(angle).to_angle(), angle)
def test_wedge_vee(): phi = torch.Tensor([1]) Phi = SO2.wedge(phi) assert (phi == SO2.vee(Phi)).all()
def test_identity(): C = SO2.identity() assert isinstance(C, SO2) \ and C.mat.dim() == 2 \ and C.mat.shape == (2, 2)
def test_adjoint_batch(): C = SO2.exp(torch.Tensor([-1., 0., 1.])) assert (C.adjoint() == torch.ones(C.mat.shape[0])).all()
def test_adjoint(): C = SO2.exp(torch.Tensor([np.pi / 4])) assert (C.adjoint() == torch.Tensor([1.])).all()
def test_inv_batch(): C = SO2.exp(torch.Tensor([-1., 0., 1.])) assert utils.allclose(C.dot(C.inv()).mat, SO2.identity(C.mat.shape[0]).mat)
def test_exp_log(): C_big = SO2.exp(torch.Tensor([np.pi / 4])) assert utils.allclose(SO2.exp(SO2.log(C_big)).mat, C_big.mat) C_small = SO2.exp(torch.Tensor([0])) assert utils.allclose(SO2.exp(SO2.log(C_small)).mat, C_small.mat)
def test_normalize(): C = SO2.exp(torch.Tensor([np.pi / 4])) C.mat.add_(0.1) C.normalize() assert SO2.is_valid_matrix(C.mat).all()
def test_exp_log_batch(): C = SO2.exp(torch.Tensor([-1., 0., 1.])) assert utils.allclose(SO2.exp(SO2.log(C)).mat, C.mat)
def test_wedge(): phi = torch.Tensor([1]) assert torch.allclose(torch.tensor([[0, -1], [1, 0]], dtype=torch.float32), SO2.wedge(phi))
def test_wedge_batch(): phis = torch.Tensor([1, 2]) Phis = SO2.wedge(phis) assert (Phis[0, :, :] == SO2.wedge(torch.Tensor([phis[0]]))).all() assert (Phis[1, :, :] == SO2.wedge(torch.Tensor([phis[1]]))).all()
def test_inv(): C = SO2.exp(torch.Tensor([np.pi / 4])) assert utils.allclose(C.dot(C.inv()).mat, SO2.identity().mat)
def test_wedge_vee_batch(): phis = torch.Tensor([1, 2]) Phis = SO2.wedge(phis) assert (phis == SO2.vee(Phis)).all()