grid_sample 输出结果与torch的不一致

[zhiminzhang0830]

相同的输入数据条件下，paddle和torch的grid_sample输出结果不一致，diff 为13。

gpu diff: 13.0
cpu diff: 13.0

import paddle
import numpy as np
import torch

input = np.load('input.npy')
grid = np.load('vgrid.npy')

# paddle cpu
cpu_place = paddle.CPUPlace()
input_paddle = paddle.to_tensor(input, place=cpu_place)
grid_paddle = paddle.to_tensor(grid, place=cpu_place)
output_paddle = paddle.nn.functional.grid_sample(
        x=input_paddle, grid=grid_paddle, padding_mode='border', mode='nearest', align_corners=True
    )
# paddle gpu
input_paddle = paddle.to_tensor(input)
grid_paddle = paddle.to_tensor(grid)
output_paddle_gpu = paddle.nn.functional.grid_sample(
        x=input_paddle, grid=grid_paddle, padding_mode='border', mode='nearest', align_corners=True
    )



# torch cpu
input_torch = torch.from_numpy(input)
grid_torch = torch.from_numpy(grid)
output_torch = torch.nn.functional.grid_sample(input_torch, grid_torch,  padding_mode='border', mode='nearest', align_corners=True)

# torch gpu
device = torch.device('cuda:0')
input_torch = torch.from_numpy(input).to(device)
grid_torch = torch.from_numpy(grid).to(device)
output_torch_gpu = torch.nn.functional.grid_sample(input_torch, grid_torch,  padding_mode='border', mode='nearest', align_corners=True)

diff = abs(output_torch_gpu.cpu().numpy() - output_paddle_gpu.cpu().numpy()).max()
print(' gpu diff:', diff)

diff = abs(output_torch.numpy() - output_paddle.cpu().numpy()).max()
print(' cpu diff:', diff)

其他补充信息 Additional Supplementary Information

torch版本：1.12.1
paddle版本：2.5.2
input.npy 和 vgrid.npy 可以从此处下载

[wangna11BD]

升级torch版本至2.1.0，可以对齐～

[AndSonder]

@zhiminzhang0830 @wangna11BD 可以请问一下为什么无法和torch 2.1.0 之前的结果对齐吗，我在给 openvino 支持 paddle2.5.1 的时候也遇到了这个问题 当 padding_mode='border' 的时候，有时候 grid_sample 的值无法和 openvino 端输出的值对齐，在 openvino 中 paddle 和 torch 对于 grid_sample 这个算子的写法是一样的，torch 就可以通过相关单侧

[paddle-bot]

Since you haven't replied for more than a year, we have closed this issue/pr.
If the problem is not solved or there is a follow-up one, please reopen it at any time and we will continue to follow up.
由于您超过一年未回复，我们将关闭这个issue/pr。
若问题未解决或有后续问题，请随时重新打开，我们会继续跟进。