IndexError: tensors used as indices must be long, byte or bool tensors

[JinglUO703]

[fuzuoyi]

res[indices.cpu().numpy()] = updates that will work

[StevenLiuWen]

@JinglUO703

1. The fast way is

   res[indices.long()] = updates

2. (Optional way) Also, since indices here are not dynamic index, np.ndarray might be better in compatibility. It can be fixed by following codes:

   
   # indices = torch.from_numpy(np.reshape(
   #     np.reshape(np.arange(0, batch_size) * 9, [-1, 1]) + col_inds,
   #     newshape=(-1,))).to(device)

For better compatibility， since if indices is torch.tensor, it must be long dtype.

For fixed index, np.ndarray might be better.

indices = np.reshape(np.reshape(np.arange(0, batch_size) * 9, [-1, 1]) + col_inds, newshape=(-1, ))

We will fix this soon, and the solution proposed by @fuzuoyi is also feasible.

[JinglUO703]

res[indices.cpu().numpy()] = updates 那可行

[JinglUO703]

res[indices.cpu().numpy()] = updates 那可行

Thanks,at first I thought it was numpy version

[JinglUO703]

@ JinglUO703

1. 快速的方法是

res [indices.long（）] =更新

1. （可选方式）而且，由于indices这里不是动态索引，因此np.ndarray的兼容性可能更好。可以通过以下代码进行修复：

＃索引= torch.from_numpy（np.reshape（
＃      np.reshape（np.arange（0，batch_size）* 9，[-1，1]）+ col_inds，
＃      newshape =（-1，）））。to（设备）

＃为了更好的兼容性，由于索引为torch.tensor，因此必须为long dtype。
＃对于固定索引，np.ndarray可能更好。
指数 = np.reshape（np.reshape（np.arange（ 0，的batch_size） *  9，[ - 1， 1 ]） + col_inds， newshape =（ - 1））

我们将尽快修复此问题，@ fuzuoyi提出的解决方案也是可行的。

3Q,I have solved this problem through @fuzuoyi method.