def test_single_input(self):
def func(x):
return paddle.sum(paddle.matmul(x, x))
numerical_func_output = func(self.x).numpy()
numerical_vhp = _compute_numerical_vhp(func, self.x, self.vx,
self.numerical_delta,
self.np_dtype)
self.x.stop_gradient = False
func_output, vhp = paddle.autograd.vhp(func, self.x, self.vx)
assert np.allclose(func_output.numpy(), numerical_func_output,
self.rtol, self.atol)
assert np.allclose(vhp[0].numpy(), numerical_vhp[0], self.rtol,
self.atol)
def test_v_default(self):
def func(x, y):
return paddle.sum(paddle.matmul(x, y))
numerical_func_output = func(self.x, self.y).numpy()
vx = paddle.ones(self.vx.shape, dtype=self.vx.dtype)
vy = paddle.ones(self.vy.shape, dtype=self.vy.dtype)
numerical_vhp = _compute_numerical_vhp(func, [self.x, self.y],
[vx, vy], self.numerical_delta,
self.np_dtype)
self.x.stop_gradient = False
self.y.stop_gradient = False
func_output, vhp = paddle.autograd.vhp(func, [self.x, self.y])
assert np.allclose(func_output.numpy(), numerical_func_output,
self.rtol, self.atol)
for i in range(len(vhp)):
assert np.allclose(vhp[i].numpy(), numerical_vhp[i], self.rtol,
self.atol)
def test_allow_unused_true(self):
def func(x, y):
return paddle.sum(paddle.matmul(x, x))
numerical_func_output = func(self.x, self.y).numpy()
numerical_vhp = _compute_numerical_vhp(func, [self.x, self.y],
[self.vx, self.vy],
self.numerical_delta,
self.np_dtype)
self.x.stop_gradient = False
self.y.stop_gradient = False
func_output, vhp = paddle.autograd.vhp(func, [self.x, self.y],
[self.vx, self.vy],
allow_unused=True)
assert np.allclose(func_output.numpy(), numerical_func_output,
self.rtol, self.atol)
assert np.allclose(vhp[0].numpy(), numerical_vhp[0], self.rtol,
self.atol)
assert vhp[1] is None
def test_create_graph_true(self):
def func(x):
return paddle.sum(F.sigmoid(x))
numerical_func_output = func(self.x).numpy()
numerical_vhp = _compute_numerical_vhp(func, self.x, self.vx,
self.numerical_delta,
self.np_dtype)
self.x.stop_gradient = False
func_output, vhp = paddle.autograd.vhp(func,
self.x,
self.vx,
create_graph=True)
assert np.allclose(func_output.numpy(), numerical_func_output,
self.rtol, self.atol)
assert vhp[0].stop_gradient == False
assert np.allclose(vhp[0].numpy(), numerical_vhp[0], self.rtol,
self.atol)
triple_grad = paddle.grad(vhp, self.x)
assert triple_grad is not None
def test_create_graph_false(self):
def func(x):
return paddle.sum(F.sigmoid(x))
numerical_func_output = func(self.x).numpy()
numerical_vhp = _compute_numerical_vhp(func, self.x, self.vx,
self.numerical_delta,
self.np_dtype)
self.x.stop_gradient = False
func_output, vhp = paddle.autograd.vhp(func, self.x, self.vx)
assert np.allclose(func_output.numpy(), numerical_func_output,
self.rtol, self.atol)
assert vhp[0].stop_gradient == True
assert np.allclose(vhp[0].numpy(), numerical_vhp[0], self.rtol,
self.atol)
try:
paddle.grad(vhp, self.x)
except RuntimeError as e:
error_msg = cpt.get_exception_message(e)
assert error_msg.find("has no gradient") > 0