如何为相等大小的图块创建相等的重叠量？

Question

这些功能将图像张量分割为相等大小的重叠图块。目前，行/列中的最终图块的x / y重叠值可能会有所不同。我需要所有x重叠值都相同，并且所有y重叠值都相同，但是我尝试解决的所有问题似乎都没有。

import torch

def calc_tiles(d, td):
   num_tiles = d // td + 1
   overlap = (d - td ) // (num_tiles - 1)
   tile_idx = [x * overlap for x in range(0, num_tiles)]
   return tile_idx, num_tiles, overlap

def split_tensor_equal(tensor, tile_size=256, offset_x=0, offset_y=0):
    tensor = tensor.clone()
    tile_h, tile_w = tile_size +offset_y, tile_size +offset_x
    h, w = tensor.size(2), tensor.size(3)   

    rows, ovlp = [0,0], [0,0]   
    x_idx, rows[0], ovlp[0] = calc_tiles(w, tile_w)
    y_idx, rows[1], ovlp[1] = calc_tiles(h, tile_h) 

    tile_list = []
    x_max, y_max = 0, 0
    for y in y_idx:
        y_max += tile_h
        y_min = y_max-tile_h       
        for x in x_idx:     
            x_max += tile_w
            x_min = x_max-tile_w
            if x_max > w:
                x_max = w
                x_min = w-tile_w    
            if y_max > h:
                y_max = h
                y_min = h-tile_h                

            #print(y_min,y_max, x_min,x_max)         
            tile = tensor[:, :, y_min:y_max, x_min:x_max]
            tile_list.append(tile)
        x_max = tile_w

    return tile_list, rows, ovlp


test_input = torch.randn(3, 1024, 768).unsqueeze(0)
split_tensor_equal(test_input, tile_size=560)