Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix set value grad #59034

Merged
merged 7 commits into from
Dec 25, 2023
Merged

Fix set value grad #59034

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
29eaaa1
first fix the UT
zoooo0820 Nov 15, 2023
8a5f39f
fix set value grad
zoooo0820 Nov 15, 2023
c833770
polish code
zoooo0820 Nov 16, 2023
6cc1f71
add static mode backward test
zoooo0820 Nov 17, 2023
6dae6f7
Merge branch 'develop' into fix_set_value_grad
zoooo0820 Dec 20, 2023
5005daf
always has input valuetensor
zoooo0820 Dec 20, 2023
f24fd45
add dygraph test
zoooo0820 Dec 21, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 17 additions & 19 deletions paddle/fluid/operators/set_value_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,30 +153,28 @@ class SetValueGradMaker : public framework::SingleGradOpMaker<T> {
void Apply(GradOpPtr<T> op) const override {
if (this->HasInput("ValueTensor")) {
op->SetType("set_value_grad");

op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
op->SetInput("ValueTensor", this->Input("ValueTensor"));
if (this->HasInput("StartsTensorList")) {
op->SetInput("StartsTensorList", this->Input("StartsTensorList"));
}
if (this->HasInput("EndsTensorList")) {
op->SetInput("EndsTensorList", this->Input("EndsTensorList"));
}
if (this->HasInput("StepsTensorList")) {
op->SetInput("StepsTensorList", this->Input("StepsTensorList"));
}

op->SetAttrMap(this->Attrs());

op->SetOutput(framework::GradVarName("ValueTensor"),
this->InputGrad("ValueTensor"));
op->SetOutput(framework::GradVarName("Input"), this->InputGrad("Input"));

} else {
op->SetType("assign");
op->SetInput("X", this->OutputGrad("Out"));
op->SetOutput("Out", this->InputGrad("Input"));
op->SetType("set_value_with_scalar_grad");
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这行看起来依然没有被覆盖到，是不是需要注册一个新Op set_value_with_scalar_grad？

}

op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));

if (this->HasInput("StartsTensorList")) {
op->SetInput("StartsTensorList", this->Input("StartsTensorList"));
}
if (this->HasInput("EndsTensorList")) {
op->SetInput("EndsTensorList", this->Input("EndsTensorList"));
}
if (this->HasInput("StepsTensorList")) {
op->SetInput("StepsTensorList", this->Input("StepsTensorList"));
}

op->SetAttrMap(this->Attrs());

op->SetOutput(framework::GradVarName("Input"), this->InputGrad("Input"));
}
};

Expand Down
6 changes: 3 additions & 3 deletions paddle/phi/api/yaml/legacy_backward.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -612,14 +612,14 @@

- backward_op : set_value_grad
forward : set_value (Tensor x, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes, int64_t[] shape, Scalar[] values) -> Tensor(out)
args : (Tensor out_grad)
args : (Tensor out_grad, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes)
output : Tensor(x_grad)
infer_meta:
func: UnchangedInferMeta
param: [out_grad]
kernel:
func: assign
param: [out_grad]
func: set_value_with_scalar_grad
param: [out_grad, starts, ends, steps, axes, decrease_axes, none_axes]

- backward_op : set_value_with_tensor_grad
forward: set_value_with_tensor (Tensor x, Tensor values, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes) -> Tensor(out)
Expand Down
17 changes: 17 additions & 0 deletions paddle/phi/kernels/cpu/set_value_grad_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,3 +35,20 @@ PD_REGISTER_KERNEL(set_value_grad,
phi::dtype::float16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}

PD_REGISTER_KERNEL(set_value_with_scalar_grad,
CPU,
ALL_LAYOUT,
phi::SetValueWithScalarGradKernel,
float,
double,
int,
int64_t,
bool,
int16_t,
uint8_t,
int8_t,
phi::dtype::bfloat16,
phi::dtype::float16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
17 changes: 17 additions & 0 deletions paddle/phi/kernels/gpu/set_value_grad_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,3 +35,20 @@ PD_REGISTER_KERNEL(set_value_grad,
phi::dtype::bfloat16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}

PD_REGISTER_KERNEL(set_value_with_scalar_grad,
GPU,
ALL_LAYOUT,
phi::SetValueWithScalarGradKernel,
float,
double,
int,
int64_t,
bool,
int16_t,
uint8_t,
int8_t,
phi::dtype::float16,
phi::dtype::bfloat16,
phi::dtype::complex<float>,
phi::dtype::complex<double>) {}
22 changes: 22 additions & 0 deletions paddle/phi/kernels/impl/set_value_grad_kernel_impl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -341,4 +341,26 @@ void SetValueGradKernel(const Context& dev_ctx,
}
}

template <typename T, typename Context>
void SetValueWithScalarGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const IntArray& starts,
const IntArray& ends,
const IntArray& steps,
const std::vector<int64_t>& axes,
const std::vector<int64_t>& decrease_axes,
const std::vector<int64_t>& none_axes,
DenseTensor* x_grad) {
SetValueGradKernel<T, Context>(dev_ctx,
out_grad,
starts,
ends,
steps,
axes,
decrease_axes,
none_axes,
x_grad,
nullptr);
}

} // namespace phi
10 changes: 10 additions & 0 deletions paddle/phi/kernels/set_value_grad_kernel.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,4 +32,14 @@ void SetValueGradKernel(const Context& dev_ctx,
DenseTensor* x_grad,
DenseTensor* value_grad);

template <typename T, typename Context>
void SetValueWithScalarGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const IntArray& starts,
const IntArray& ends,
const IntArray& steps,
const std::vector<int64_t>& axes,
const std::vector<int64_t>& decrease_axes,
const std::vector<int64_t>& none_axes,
DenseTensor* x_grad);
} // namespace phi
31 changes: 31 additions & 0 deletions paddle/phi/kernels/xpu/set_value_grad_kernel.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -397,6 +397,28 @@ void SetValueGradKernel(const Context& dev_ctx,
}
}

template <typename T, typename Context>
void SetValueWithScalarGradKernel(const Context& dev_ctx,
const DenseTensor& out_grad,
const IntArray& starts,
const IntArray& ends,
const IntArray& steps,
const std::vector<int64_t>& axes,
const std::vector<int64_t>& decrease_axes,
const std::vector<int64_t>& none_axes,
DenseTensor* x_grad) {
SetValueGradKernel<T, Context>(dev_ctx,
out_grad,
starts,
ends,
steps,
axes,
decrease_axes,
none_axes,
x_grad,
nullptr);
}

} // namespace phi

PD_REGISTER_KERNEL(set_value_grad,
Expand All @@ -407,3 +429,12 @@ PD_REGISTER_KERNEL(set_value_grad,
phi::dtype::float16,
int,
int64_t) {}

PD_REGISTER_KERNEL(set_value_with_scalar_grad,
XPU,
ALL_LAYOUT,
phi::SetValueWithScalarGradKernel,
float,
phi::dtype::float16,
int,
int64_t) {}
54 changes: 54 additions & 0 deletions test/legacy_test/test_set_value_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1978,5 +1978,59 @@ def test_check_grad(self):
self.check_grad_with_place(place, ['Input'], 'Out', check_dygraph=False)


class TestSetValueWithScalarInStatic(unittest.TestCase):
def setUp(self):
paddle.enable_static()
self.shape = (10, 2)
self.exe = paddle.static.Executor()
self.train_program = paddle.static.Program()
self.startup_program = paddle.static.Program()

def test_value_input_is_scalar(self):
with paddle.static.program_guard(
self.train_program, self.startup_program
):
x = paddle.ones(self.shape)
x.stop_gradient = False
y = x * 1

# mock test case x[0, 0] = 10 with no ValueTensor input
inputs = {
'Input': y,
}
attrs = {
'axes': [0, 1],
'starts': [0, 0],
'ends': [1, 1],
'steps': [1, 1],
'values': [10],
'shape': [1],
}

helper = LayerHelper("set_value")
out = helper.create_variable_for_type_inference(dtype=y.dtype)

helper.append_op(
type="set_value",
inputs=inputs,
outputs={'Out': out},
attrs=attrs,
)

np_data = np.ones(self.shape).astype('float32')

paddle.static.append_backward(out.sum())
res = self.exe.run(
self.train_program, fetch_list=[out, x.grad_name]
)

np_data[0, 0] = 10
expected_x_grad = np.ones(self.shape)
expected_x_grad[0, 0] = 0

np.testing.assert_array_equal(res[0], np_data)
np.testing.assert_array_equal(res[1], expected_x_grad)


if __name__ == '__main__':
unittest.main()
2 changes: 2 additions & 0 deletions test/legacy_test/test_zero_dim_tensor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -830,6 +830,7 @@ def test_setitem(self):
np.testing.assert_allclose(out[1, 2, 3, 4], np.array(10))
self.assertEqual(x.grad.shape, [2, 3, 4, 5])
x_grad_expected = np.ones((2, 3, 4, 5)) * 2
x_grad_expected[1, 2, 3, 4] = 0
np.testing.assert_allclose(x.grad, x_grad_expected)

# case2: 0-D Tensor indice in some axis
Expand All @@ -847,6 +848,7 @@ def test_setitem(self):
self.assertEqual(out.shape, x.shape)
np.testing.assert_allclose(out[1, 1], np.ones((4, 5)) * 0.5)
x_grad_expected = np.ones((2, 3, 4, 5))
x_grad_expected[1, 1] = 0
np.testing.assert_allclose(x.grad, x_grad_expected)

# case3：0-D Tensor indice in some axis, value is a Tensor
Expand Down
1 change: 1 addition & 0 deletions test/xpu/test_zero_dim_tensor_xpu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -441,6 +441,7 @@ def test_setitem(self):
np.testing.assert_allclose(out[1, 2, 3, 4], np.array(10))
self.assertEqual(x.grad.shape, [2, 3, 4, 5])
x_grad_expected = np.ones((2, 3, 4, 5)) * 2
x_grad_expected[1, 2, 3, 4] = 0
np.testing.assert_allclose(x.grad, x_grad_expected)

# case2: 0-D Tensor indice in some axis
Expand Down