bug描述 Describe the Bug

pytorch的这个算子torchvision.ops.roi_align可以支持batch！=1时的操作
paddle使用paddle.vision.ops.roi_align算子时，batch！=1时出现问题
paddle版本为2.3.1
如下代码中，batch==4，报错

import paddle
from paddle.vision.ops import roi_align
x=paddle.rand(shape=[4,3,512,512])
bbox=paddle.rand(shape=[8,4],dtype='float32')
num=paddle.to_tensor([2],dtype='int32')
print(roi_align(x,boxes=bbox,boxes_num=num,aligned=True,output_size=80))

报错为

Traceback (most recent call last):
  File "c:\Users\Administrator\Desktop\GFPGAN-master\GFPGAN-master_paddle\gfpgan\data\a.py", line 14, in <module>
    print(roi_align(x,boxes=bbox,boxes_num=num,aligned=True,output_size=80))
  File "C:\ProgramData\Anaconda3\lib\site-packages\paddle\vision\ops.py", line 1252, in roi_align
    align_out = _C_ops.roi_align(
ValueError: (InvalidArgument) The batch size of rois and the batch size of images  must be the 
same. But received the batch size of rois is 1, and the batch size of images is 4
  [Hint: Expected boxes_batch_size == batch_size, but received boxes_batch_size:1 != batch_size:4.] (at C:\home\workspace\Paddle_release\paddle\phi\kernels\cpu\roi_align_kernel.cc:217)      
  [operator < roi_align > error]

batch等于1时不报错

import paddle
from paddle.vision.ops import roi_align
x=paddle.rand(shape=[1,3,512,512])
bbox=paddle.rand(shape=[2,4],dtype='float32')
num=paddle.to_tensor([2],dtype='int32')
print(roi_align(x,boxes=bbox,boxes_num=num,aligned=True,output_size=80))

其他补充信息 Additional Supplementary Information

No response