[Zero-Dim] support 0-D output and 0-D as indice in __getitem__/__setitem__

[zoooo0820]

PR types

Others

PR changes

APIs

Describe

Pcard-66985

This PR is to support 0-D Tensor in Tensor __getitem__. From now, 0-D Tensor is semantically equivalent to scalar while 1-D Tensor is vector.

There will be two incompatible behavior changes in __getitem__:

• all axis has a scalar indice, the output will be a 0-D tensor.

>>> x = paddle.ones((2,3,4))
>>> y = x[1,1,1]

# before this PR:
>>> y
Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True,
       1.)

# after this PR:
>>> y
Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True,
       1.)

• In index element, 0-D Tensor will be treated as scalar, and 1-D Tensor will be treat as vector

>>> x = paddle.ones((2,3,4))
>>> y = x[1,1,1]
>>> y
Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True,
       1.)

# is same as y = x[1,1,1]
>>> indice = paddle.ones((), dtype='int32')
>>> x[indice, indice, indice]
Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True,
       1.)

# behavior changes for 1-D Tensor
>>> x = paddle.ones((2,3,4))
>>> indice = paddle.ones([1], dtype='int32')

# before this PR, 1-D Tensor is treated as scalar
>>> x[indice].shape
[3, 4]

# after this PR, 1-D Tensor is treated as vector
>>> x[indice].shape
[1, 3, 4]

NOTE : Combine indexing is not supported now, and 1-d Tensor is treated as vector, so following usages are not supported, we will try to support these usages as soon as possible.
x[1, paddle.ones([1], dtype='int32')] / x[:, paddle.ones([1], dtype='int32')]

For __setitem__ , now we can also use 0-D Tensor as indice

>>> x = paddle.ones((2,3,2))
>>> indice = paddle.ones([], dtype='int32')
>>> x[indice, indice] = 10
>>> x
Tensor(shape=[2, 3, 2], dtype=float32, place=Place(gpu:0), stop_gradient=True,
       [[[1. , 1. ],
         [1. , 1. ],
         [1. , 1. ]],

        [[1. , 1. ],
         [10., 10.],
         [1. , 1. ]]])

0-D Tensor can also be used as a value to fill.

>>> x = paddle.ones((2,3,2))
>>> indice = paddle.ones([], dtype='int32')
>>> value = paddle.ones([]) * 10
>>> x[indice, indice] = value
>>> x
Tensor(shape=[2, 3, 2], dtype=float32, place=Place(gpu:0), stop_gradient=True,
       [[[1. , 1. ],
         [1. , 1. ],
         [1. , 1. ]],

        [[1. , 1. ],
         [10., 10.],
         [1. , 1. ]]])

[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.

[zhwesky2010]

这个感觉可以判断out_dims_mapping如果维度>0，就判断 assert ref_dims_mapping[0] == out_dims_mapping[0]，这样之前的检查逻辑也可以保留了

[zoooo0820]

这里特化了ref_dims_mapping==[]的场景，在下面else分支中会逐轴比对ref_dims_mapping[i] / out_dims_mapping[i]

[zhwesky2010]

test_set_value_op_xpu、test_slice_op_xpu挂了

[jeff41404]

LGTM

[luotao1]

LGTM for paddle_enforce

[XiaoguangHu01]

LGTM