【Hackathon 4th No.29】为 Paddle 新增 paddle.sparse.slice 稀疏 API

[ScottWong98]

PR types

New features

PR changes

APIs

Description

实现第四期 Hackathon 第 29 个任务

中文文档: PaddlePaddle/docs#5895

原始 RFC: PaddlePaddle/community#382

原始 RFC 的修复: PaddlePaddle/community#539

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[paddle-ci-bot]

Sorry to inform you that 428dc2e's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[ScottWong98]

@zkh2016 您好，目前已经修改完毕，麻烦再重新Review以下

[jeff41404]

Why use parameter name of axes_arr, starts_arr and ends_arr? No other kernel are not named like this, just using axes, starts and ends is better?

[ScottWong98]

Thank you very much for your review!

The reason is that the data type of these parameters is IntArray, and I will get their underlying data in implementing the function, like axes_arr.GetData(). For convenience, I use axes_arr as the parameter's name and axes as the variable name indicating the underlying data described above in the function body.

And I found that the names of the starts and ends parameters are also starts_array and ends_array in DenseTensor's slice function:

Paddle/paddle/phi/kernels/impl/slice_kernel_impl.h

Lines 102 to 114 in d1d43c2

	template <typename T, typename Context>
	void SliceKernel(const Context& ctx,
	const DenseTensor& input,
	const std::vector<int64_t>& axes,
	const IntArray& starts_arr,
	const IntArray& ends_arr,
	const std::vector<int64_t>& infer_flags,
	const std::vector<int64_t>& decrease_axis,
	DenseTensor* out) {
	int rank = input.dims().size();
	auto& starts = starts_arr.GetData();
	auto& ends = ends_arr.GetData();

[ScottWong98]

Additionally, the configuration of DenseTensor's slice API in legacy_ops.yaml is:

Paddle/paddle/phi/api/yaml/legacy_ops.yaml

Lines 906 to 913 in cdbf62f

	- op : slice
	args : (Tensor input, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)
	output : Tensor
	infer_meta :
	func : SliceRawInferMeta
	kernel :
	func : slice
	backward : slice_grad

[jeff41404]

The signature of the SliceKernel should mainly refers to

Paddle/paddle/phi/kernels/slice_kernel.h

Lines 25 to 32 in d1d43c2

	void SliceKernel(const Context& ctx,
	const DenseTensor& input,
	const std::vector<int64_t>& axes,
	const IntArray& starts,
	const IntArray& ends,
	const std::vector<int64_t>& infer_flags,
	const std::vector<int64_t>& decrease_axis,
	DenseTensor* out);

The function mentioned above in Paddle/paddle/phi/kernels/impl/slice_kernel_impl.h is its implementation, and it's better to be the same, If you are interested, can propose another PR to correct the parameter names in slice_kernel_impl.h

[ScottWong98]

Thanks for your reply! I will propose a PR to correct it when I have free time.

[ScottWong98]

@jeff41404 Thank you again for your review.

Finally, it might be better to use axes, starts, and ends as parameter names. And I have changed axes_arr, starts_arr, and ends_arr to axes, starts, and ends.

Please see the latest commit for details and review it again!

[sunzhongkai588]

LGTM for docs

[jeff41404]

LGTM