Indexing a multi-dimensional tensor with a tensor in PyTorch

Question

I have the following code:

a = torch.randint(0,10,[3,3,3,3])
b = torch.LongTensor([1,1,1,1])

I have a multi-dimensional index b and want to use it to select a single cell in a. If b wasn't a tensor, I could do:

a[1,1,1,1]

Which returns the correct cell, but:

a[b]

Doesn't work, because it just selects a[1] four times.

How can I do this? Thanks

dennlinger · Accepted Answer · 2018-08-30 08:54:00Z

13

A more elegant (and simpler) solution might be to simply cast b as a tuple:

a[tuple(b)]
Out[10]: tensor(5.)

I was curious to see how this works with "regular" numpy, and found a related article explaining this quite well here.

answered Aug 30, 2018 at 8:54

dennlinger

11.7k2 gold badges47 silver badges65 bronze badges

Sign up to request clarification or add additional context in comments.

5 Comments

Chum-Chum Scarecrows Over a year ago

Is there any way of making this solution work with a list of indices?

dennlinger Over a year ago

Turns out a[list(b)] works, too. Interesting. Or are you referring to a "list of elements in the list" (i.e., something like b = [[1,1,1,1], [1,1,1,2], [2,3,1,2]]?

drevicko Over a year ago

Um... Can we do this without converting our index tensor to a tuple? (let's say it's large and resides on GPU, making a tuple of it pulls all the values to CPU, both an overhead and forcing the GPU to wait on the CPU or vice versa).

dennlinger Over a year ago

I haven't worked with it in a while, so I cannot confidently answer your question. My intuition tells me that it isn't possible, and you are going to have to move data. I'm happy to be proven wrong, though, so maybe this could be a separate question?

user30850 Over a year ago

It is worth mentioning that this even works when you want to select more than one element: b=tuple(torch.tensor([[1,1,1,1],[2,1,1,1]]).T), a[b].

Shai · Accepted Answer · 2018-08-30 08:37:47Z

11

You can split b into 4 using chunk, and then use the chunked b to index the specific element you want:

>> a = torch.arange(3*3*3*3).view(3,3,3,3)
>> b = torch.LongTensor([[1,1,1,1], [2,2,2,2], [0, 0, 0, 0]]).t()
>> a[b.chunk(chunks=4, dim=0)]   # here's the trick!
Out[24]: tensor([[40, 80,  0]])

What's nice about it is that it can be easily generalized to any dimension of a, you just need to make number of chucks equal the dimension of a.

answered Aug 30, 2018 at 8:37

Shai

115k39 gold badges259 silver badges398 bronze badges

3 Comments

Chum-Chum Scarecrows Over a year ago

Added bonus for being able to use multiple indices at once which I didn't account for in my question. Tested this and it works, although it's worth noting I need to squeeze the output. Thanks!

Shai Over a year ago

@Chum-ChumScarecrows thank you for accepting, but AFAIK dennlinger's answer also generalized to multiple indices. I suppose you should accept his.

Addison Klinke Over a year ago

Thank you! I have been looking far and wide for this kind of generalized indexing. Very surprised PyTorch doesn't support this natively, i.e. a[b] without all the extra operations

Collectives™ on Stack Overflow

Indexing a multi-dimensional tensor with a tensor in PyTorch

2 Answers 2

5 Comments

3 Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

5 Comments

3 Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related