index_put_ take min when there are repeated indices

[TaoHuUMD]

🚀 Feature

Assume we want to get this:
image[x,y]=p, where x, y, p are lists. For example,

x=[1, 1, 2]

y=[3, 3, 4]

p=[1, 2, 3]

It should be mentioned that image[x][y] may be updated twice, since x[0]=x[1]=1, y[0]=y[1]=3, but p[0]=1, p[1]=2. In this case, image[1][3] should be the minimum, 1.

But it seems that torch.where does not consider that one element in list can be updated twice or more.

Is there an efficient way to do this? Better use OP of Pytorch. I can update image in a for loop, but it’s too slow, since the len(x) is big.

Motivation

Pitch

Alternatives

Additional context

[ezyang]

First a quick clarification: you're talking about advanced indexing a.k.a. index_put_, not where_. The documentation of index_put_ says what the behavior is supposed to be in this case:

If accumulate is True, the elements in tensor are added to self. If accumulate is False, the behavior is undefined if indices contain duplicate elements.

So unfortunately, if you want to take the min, you are currently out of luck. I can't think of a good way of do this with our current API, so I'll leave this open as a feature request (I don't know how efficiently implementable this would be, we would need an "atomicMin" operation.) If you ask on https://discuss.pytorch.org/ others may have suggestions about how to efficiently implement this API in userland. Also, it would be helpful to say why you want the min in this case.

[Chen-Suyi]

It may be too late for you now, but I did find a solution with the current PyTorch API, following the clue of "atomicMin" operation. The function you need should be TORCH.TENSOR.SCATTER_REDUCE_

The following example may help you:

x=[1, 1, 2]
y=[3, 3, 4]
p=[1, 2, 3]
shape = image.shape
index = x * shape[1] + y
image = image.flatten()
image.scatter_reduce(dim=0, index=index, src=p, reduce="amin", include_self=False)
image = image.reshape(shape)

.