def test_allow_unused_true(self):
def func(x, y):
return x * x
numerical_jacobian = _compute_numerical_batch_jacobian(
func, [self.x, self.y], self.numerical_delta, self.np_dtype)
self.x.stop_gradient = False
self.y.stop_gradient = False
jacobian = paddle.autograd.batch_jacobian(func, [self.x, self.y],
allow_unused=True)
assert np.allclose(jacobian[0].numpy(), numerical_jacobian[0][0],
self.rtol, self.atol)
assert jacobian[1] is None
def test_batch_multi_input_and_batch_multi_output(self):
def func(x, y):
return x * y, x * y
numerical_jacobian = _compute_numerical_batch_jacobian(
func, [self.x, self.y], self.numerical_delta, self.np_dtype)
self.x.stop_gradient = False
self.y.stop_gradient = False
batch_jacobian = paddle.autograd.batch_jacobian(func, [self.x, self.y])
for i in range(len(batch_jacobian)):
assert np.allclose(batch_jacobian[i], numerical_jacobian[i],
self.rtol, self.atol)
def test_create_graph_true(self):
def func(x, y):
return x * y
numerical_jacobian = _compute_numerical_batch_jacobian(
func, [self.x, self.y], self.numerical_delta, self.np_dtype)
self.x.stop_gradient = False
self.y.stop_gradient = False
jacobian = paddle.autograd.batch_jacobian(func, [self.x, self.y],
create_graph=True)
for j in range(len(jacobian)):
assert jacobian[j].stop_gradient == False
assert np.allclose(jacobian[j].numpy(), numerical_jacobian[0][j],
self.rtol, self.atol)
double_grad = paddle.grad(jacobian[0], [self.x, self.y])
assert double_grad is not None
def test_batch_single_input_and_batch_multi_output(self):
def func(x):
return paddle.matmul(paddle.matmul(x, self.weight), self.y), x * x
numerical_jacobian = _compute_numerical_batch_jacobian(
func, [self.x], self.numerical_delta, self.np_dtype)
self.x.stop_gradient = False
batch_jacobian = paddle.autograd.batch_jacobian(
func,
self.x,
)
for i in range(len(batch_jacobian)):
assert np.allclose(batch_jacobian[i].numpy(),
numerical_jacobian[i][0], self.rtol, self.atol)
def test_batch_single_input_and_batch_single_output(self):
def func(x):
return paddle.matmul(paddle.matmul(x, self.weight), self.y)
numerical_jacobian = _compute_numerical_batch_jacobian(
func, [self.x], self.numerical_delta, self.np_dtype)
self.x.stop_gradient = False
batch_jacobian = paddle.autograd.batch_jacobian(
func,
self.x,
)
self.assertTrue(
np.allclose(batch_jacobian.numpy().all(),
numerical_jacobian[0][0].all()))
def test_create_graph_false(self):
def func(x, y):
return x * y
numerical_jacobian = _compute_numerical_batch_jacobian(
func, [self.x, self.y], self.numerical_delta, self.np_dtype)
self.x.stop_gradient = False
self.y.stop_gradient = False
jacobian = paddle.autograd.batch_jacobian(func, [self.x, self.y])
for j in range(len(jacobian)):
assert jacobian[j].stop_gradient == True
assert np.allclose(jacobian[j].numpy(), numerical_jacobian[0][j],
self.rtol, self.atol)
try:
paddle.grad(jacobian[0], [self.x, self.y])
except RuntimeError as e:
error_msg = cpt.get_exception_message(e)
assert error_msg.find("has no gradient") > 0