def test_grad_div_backward(get_clients) -> None:
parties = get_clients(2)
session = Session(parties=parties)
session.autograd_active = True
SessionManager.setup_mpc(session)
x_secret = torch.tensor([1.0, 2.1, 3.0, -4.13], requires_grad=True)
x = MPCTensor(secret=x_secret, session=session, requires_grad=True)
y_secret = torch.tensor([-2.0, 3.0, 4.39, 5.0], requires_grad=True)
y = MPCTensor(secret=y_secret, session=session, requires_grad=True)
z = x_secret / y_secret
z.backward(torch.tensor([1, 1, 1, 1]))
grad = torch.tensor([1, 1, 1, 1])
grad_mpc = MPCTensor(secret=grad, session=session, requires_grad=True)
ctx = {"x": x, "y": y, "result": x / y}
grad_x, grad_y = GradDiv.backward(ctx, grad_mpc)
expected_grad_x = x_secret.grad
expected_grad_y = y_secret.grad
res_x = grad_x.reconstruct()
res_y = grad_y.reconstruct()
assert np.allclose(res_x, expected_grad_x, rtol=1e-2)
assert np.allclose(res_y, expected_grad_y, rtol=1e-2)
def test_backward_without_requires_grad(get_clients):
clients = get_clients(4)
session = Session(parties=clients)
session.autograd_active = True
SessionManager.setup_mpc(session)
x_secret = torch.tensor([[0.125, -1.25], [-4.25, 4], [-3, 3]])
y_secret = torch.tensor([[4.5, -2.5], [5, 2.25], [-3, 3]])
x = MPCTensor(secret=x_secret, session=session)
y = MPCTensor(secret=y_secret, session=session)
res_mpc = x - y
s_mpc = res_mpc.sum()
s_mpc.backward()
assert not res_mpc.requires_grad
assert res_mpc.grad is None
assert x.grad is None
assert y.grad is None
def test_backward(get_clients):
clients = get_clients(4)
session = Session(parties=clients)
session.autograd_active = True
SessionManager.setup_mpc(session)
x_secret = torch.tensor([[0.125, -1.25], [-4.25, 4], [-3, 3]],
requires_grad=True)
y_secret = torch.tensor([[4.5, -2.5], [5, 2.25], [-3, 3]],
requires_grad=True)
x = MPCTensor(secret=x_secret, session=session, requires_grad=True)
y = MPCTensor(secret=y_secret, session=session, requires_grad=True)
res_mpc = x * y
res = x_secret * y_secret
s_mpc = res_mpc.sum()
s = torch.sum(res)
s_mpc.backward()
s.backward()
assert np.allclose(x.grad.get(), x_secret.grad, rtol=1e-3)
assert np.allclose(y.grad.get(), y_secret.grad, rtol=1e-3)
def test_backward_with_one_requires_grad(get_clients):
clients = get_clients(4)
session = Session(parties=clients)
session.autograd_active = True
SessionManager.setup_mpc(session)
x_secret = torch.tensor([[0.125, -1.25], [-4.25, 4], [-3, 3]],
requires_grad=True)
y_secret = torch.tensor([[4.5, -2.5], [5, 2.25], [-3, 3]])
x = MPCTensor(secret=x_secret, session=session, requires_grad=True)
y = MPCTensor(secret=y_secret, session=session)
res_mpc = x - y
res = x_secret - y_secret
s_mpc = res_mpc.sum()
s = torch.sum(res)
s_mpc.backward()
s.backward()
# TODO: add assert for res_mpc.grad and res.grad
assert res_mpc.requires_grad
assert np.allclose(x.grad.get(), x_secret.grad, rtol=1e-3)
assert y.grad is None
def test_forward(get_clients) -> None:
model = LinearSyNet(torch)
clients = get_clients(2)
session = Session(parties=clients)
session.autograd_active = True
SessionManager.setup_mpc(session)
mpc_model = model.share(session=session)
x_secret = torch.tensor([[0.125, -1.25, -4.25], [-3, 3, 8], [-3, 3, 8]],
requires_grad=True)
x_mpc = MPCTensor(secret=x_secret, session=session, requires_grad=True)
out_torch = model(x_secret)
out_mpc = mpc_model(x_mpc)
s_torch = torch.sum(out_torch)
s_mpc = out_mpc.sum()
s_torch.backward()
s_mpc.backward()
assert np.allclose(x_mpc.grad.get(), x_secret.grad, rtol=1e-2)
