[0-size Tensor No.355] Add 0-size Tensor support for unique_consecutive

[ccsuzzh]

PR Category

Execute Infrastructure

PR Types

Improvements

Description

infermeta修改设置维度
修改cpu/gpu kernel
当0-size Tensor时， out tensor的shape与输入保持一致（与torch保持一致）

PaddleAPITest测试通过

[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.

[DanielSun11]

paddleAPITest回测请优先使用accuracy模式（ --accuracy=True），校验输出是否和torch一致

[ccsuzzh]

paddleAPITest回测请优先使用accuracy模式（ --accuracy=True），校验输出是否和torch一致

因为paddle.unique_consecutive(Tensor([0],"float64"), return_inverse=True, return_counts=True, )输出shape刚刚全部都[0]，所以没有回测出来。

[ccsuzzh]

/re-run all-failed

[DanielSun11]

这个地方还是有些疑问。
out->set_dims({0}) 设置完后，tensor的shape是不是[0] ? shape [0] 和shape[ ] 是两个不同的shape，shape [0]的dims().size() 为1，是1d Tensor，尽管是0-size。shape [ ]的dims().size() 为0，是0d Tensor也是个标量。 试试out->set_dims({}),让tensor的shape为[ ]
请把

import torch
x = torch.randn([0,1,2])
output, inverse, counts = torch.unique_consecutive(x, return_inverse=True,return_counts=True)
print("x ",x)
print("output ",output)
print("inverse ",inverse)
print("counts ",counts)

转化为paddle的代码，跑一下，将结果贴到PR描述里。

[DanielSun11]

请注意如下三种情况：

case1

import torch
x = torch.randn([0,1,2])
output, inverse, counts = torch.unique_consecutive(x, return_inverse=True,return_counts=True,dim = 0)
print("x ",x)
print("output ",output)
print("inverse ",inverse)
print("counts ",counts)

输出：

x  tensor([], size=(0, 1, 2))
output  tensor([], size=(0, 1, 2))
inverse  tensor([], dtype=torch.int64)
counts  tensor([], dtype=torch.int64)

case2

import torch
x = torch.randn([0,1,2])
output, inverse, counts = torch.unique_consecutive(x, return_inverse=True,return_counts=True)
print("x ",x)
print("output ",output)
print("inverse ",inverse)
print("counts ",counts)

输出：

x  tensor([], size=(0, 1, 2))
output  tensor([])
inverse  tensor([], size=(0, 1, 2), dtype=torch.int64)
counts  tensor([], dtype=torch.int64)

case3

import torch
x = torch.randn([0,1,2])
output, inverse, counts = torch.unique_consecutive(x, return_inverse=True,return_counts=True,dim=1)
print("x ",x)
print("output ",output)
print("inverse ",inverse)
print("counts ",counts)

输出：报错paddle error

Traceback (most recent call last):
  File "/home/code/test/test_unique.py", line 3, in <module>
    output, inverse, counts = torch.unique_consecutive(x, return_inverse=True,return_counts=True,dim=1)
  File "/usr/local/lib/python3.10/dist-packages/torch/_jit_internal.py", line 622, in fn
    return if_true(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/_jit_internal.py", line 622, in fn
    return if_true(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/functional.py", line 1046, in _unique_consecutive_impl
    output, inverse_indices, counts = _VF.unique_consecutive(  # type: ignore[attr-defined]
RuntimeError: There are 0 sized dimensions, and they aren't selected, so unique cannot be applied

请验证修复后如下三种情形的输出，并告知输出结果～

[DanielSun11]

使用paddleAPITest在CPU和GPU上使用accuracy模式，跑如下测试case，非0size的Tensor不要求进行回测：

paddle.unique_consecutive(Tensor([0],"float64"), return_inverse=True, return_counts=True, )
paddle.unique_consecutive(Tensor([0],"float64"), return_inverse=True, return_counts=True, axis=-1, )
paddle.unique_consecutive(x=Tensor([0, 1],"float32"), return_inverse=True, )
paddle.unique_consecutive(x=Tensor([0, 1],"float32"), return_inverse=True, return_counts=True, )
paddle.unique_consecutive(x=Tensor([0, 4],"float32"), return_inverse=True, return_counts=True, axis=0, )
paddle.unique_consecutive(x=Tensor([3, 0],"float64"), axis=1, )
paddle.unique_consecutive(x=Tensor([4, 0],"float32"), return_inverse=True, )
paddle.unique_consecutive(x=Tensor([6, 0],"float32"), return_inverse=True, return_counts=True, )

[paddle-ci-bot]

Sorry to inform you that 73bb142's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[DanielSun11]

已经被 #74305 修复