问题

我想对张量（图像集合）执行位移操作，例如使用torch.roll或其他方法。但我要对每个张量应用的每个位移都不同（我将有另一个包含位移的张量）。是否有一种有效的方法可以"一次性"执行，而不必循环遍历每个图像？

例如，我有以下简单的张量，它对应于两幅图像：

torch.tensor([1, 2, 3, 4, 5, 6, 7, 8]).view(2, 2, 2)

tensor([[[1, 2],
         [3, 4]],

        [[5, 6],
         [7, 8]]])

我需要对第一张图像在x轴上进行1px的位移，并对第二张图像在y轴上进行位移：

tensor([[[2, 1],
         [4, 3]],

        [[7, 8],
         [5, 6]]])

torch.tensor([1, 2, 3, 4, 5, 6, 7, 8]).view(2,2,2)

tensor([[[1, 2],
         [3, 4]],

        [[5, 6],
         [7, 8]]])

tensor([[[2, 1],
         [4, 3]],

        [[7, 8],
         [5, 6]]])

答案1

得分: 1

扩展评论，我认为使用for循环可能是实现这个目标的最佳方式

```python
import torch
data = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8]).view(2, 2, 2)

# 为每个元素定义在y轴和x轴上的位移量
shifts = [(0, 1), (1, 0)]

# 分别移动每个图像并将它们连接起来
data_shifted = []
for d, s in zip(data, shifts):
    data_shifted.append(torch.roll(d, s, dims=(0, 1)))
result = torch.stack(data_shifted, dim=0)

print(result)

其结果为

tensor([[[2, 1],
         [4, 3]],

        [[7, 8],
         [5, 6]]])

如果你愿意，可以用一行的推导式代替这个循环

result = torch.stack([torch.roll(d, s, dims=(0, 1)) for d, s in zip(data, shifts)], dim=0)

所有这些都是自动求导兼容的，供你参考。

<details>
<summary>英文:</summary>

Expanding on the comments I think a for loop is likely the best way to do this

```python
import torch
data = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8]).view(2, 2, 2)

# define shift amount in y-axis and x-axis for each element
shifts = [(0, 1), (1, 0)]

# shift each image separately and concatenate them
data_shifted = []
for d, s in zip(data, shifts):
    data_shifted.append(torch.roll(d, s, dims=(0, 1)))
result = torch.stack(data_shifted, dim=0)

print(result)

which results in

tensor([[[2, 1],
         [4, 3]],

        [[7, 8],
         [5, 6]]])

The loop could be replaced with a one-line comprehension if you prefer

result = torch.stack([torch.roll(d, s, dims=(0, 1)) for d, s in zip(data, shifts)], dim=0)

All of this is autograd compatible just FYI