如何并行对多个张量进行thergrid？

发布于 2025-02-11 02:10:37 字数 291 浏览 1 评论 0原文

假设我们有一个大小[60,9]的张量X和一个大小的张量[60,9] 是否可以进行操作，例如xx，yy = torch.meshgrid（x，y），以便xx和yy具有大小[60,9,9]和xx [i，：，：]，yy [i，：，：]基本上是torch.meshgrid（x [i]，y [i]）？

内置的torch.meshgrid操作仅接受1D张量，是否可以在不使用循环的情况下执行上述操作（因为不使用GPU的并行操作，这是效率低下的）？

原文

分享到QQ

分享到微博

如果你对这篇内容有疑问，欢迎到本站社区发帖提问参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。

发布评论

需要登录才能够评论，你可以免费注册一个本站的账号。

江湖彼岸 2025-02-18 02:10:37

我不相信您会获得任何收益，因为在GPU上没有进行张量的初始化。因此，提出的方法确实是在x和y或使用 map 作为一个迭代：

grids = map(torch.meshgrid, zip(x,y))

I don't believe you will gain anything since the initialization of the tensors is not done on the GPU. So a proposed approach would indeed be to loop over x and y or using map as an iterable:

grids = map(torch.meshgrid, zip(x,y))

回复收藏 0 原文

如此安好 2025-02-18 02:10:37

在为grid_sample生成网格的情况下，我找到了一种方法，这被称为一批图像。

特别是，您有2个张量grid_x，grid_y带有shape（m，n），其中m是批处理大小，n是每个网格维度的点数。您需要grid_sample的网格是大小（m，n，n，2），您可以通过在第一个维度上循环，应用：

for i in range(M):
    gx, gy = torch.meshgrid(grid_x[i], grid_y[i], indexing='ij')
    grid[i, :, :, 0] = gx
    grid[i, :, :, 1] = gy

一种更有效的方法是：

grid = torch.stack([grid_x.unsqueeze(-1).expand(M, -1, grid_y.shape[1]),
                    grid_y.unsqueeze(1).expand(M, grid_x.shape[1], -1)],
                    dim=-1
)

它的阅读不是很容易，但是它显示了meshgrid的替代方法。

I found a way to do this in the case when a grid is being generated for grid_sample, which is being called for a batch of images.

In particular, you have 2 tensors grid_x, grid_y with shape (M, N), where M is the batch size and N is the number of points per grid dimension. The grid you need for grid_sample is of size (M, N, N, 2), and you could obtain it by looping over the first dimension, applying:

for i in range(M):
    gx, gy = torch.meshgrid(grid_x[i], grid_y[i], indexing='ij')
    grid[i, :, :, 0] = gx
    grid[i, :, :, 1] = gy

A much more efficient way to do this is:

grid = torch.stack([grid_x.unsqueeze(-1).expand(M, -1, grid_y.shape[1]),
                    grid_y.unsqueeze(1).expand(M, grid_x.shape[1], -1)],
                    dim=-1
)

It's not very easy to read, but it shows an alternative to meshgrid.

回复收藏 0 原文

~没有更多了~