[ 组合替代实现 ]torch.cdist¶

torch.cdist ¶

          torch.cdist(x1, x2, p=2.0, compute_mode='use_mm_for_euclid_dist_if_necessary')

         

Paddle 无此 API，需要组合实现。

转写示例¶

当 x1 为 2D tensor 时¶

           # Pytorch 写法
y = torch.cdist(x1, x2, p)

# Paddle 写法
dist_list = []
for i in range(x1.shape[0]):
    for j in range(x2.shape[0]):
        dist_list.append(paddle.dist(x1[i, :], x2[j, :], p=p).item())
y = paddle.to_tensor(dist_list).reshape([x1.shape[0], x2.shape[0]])

          

当 x1 为 3D tensor 时¶

           # Pytorch 写法
y = torch.cdist(x1, x2, p)

# Paddle 写法
dist_list = []
for b in range(x1.shape[0]):
    for i in range(x1.shape[1]):
        for j in range(x2.shape[1]):
            dist_list.append(paddle.dist(x1[b, i, :], x2[b, j, :], p=p).item())
y = paddle.to_tensor(dist_list).reshape([x1.shape[0], x1.shape[1],x2.shape[1]])

          