Sum a tensor over a range of axes

1

I am trying to sum a tensor over its first n axes, where n is a parameter I don’t know in advance. Why something like the following doesn’t work?

v = torch.randn(100, 20, 10)
torch.sum(v, range(2))  # I would write range(n) here.

This gives an error:

TypeError: sum() received an invalid combination of arguments - got (Tensor, range), but expected one of:
 * (Tensor input, torch.dtype dtype)
      didn't match because some of the arguments have invalid types: (Tensor, range)
 * (Tensor input, tuple of names dim, bool keepdim, torch.dtype dtype, Tensor out)
 * (Tensor input, tuple of ints dim, bool keepdim, torch.dtype dtype, Tensor out)

However if I instantiate the range:

v = torch.randn(100, 20, 10)
torch.sum(v, list(range(2)))

Then it works. But this seems inefficient for large n.

2

Hi,

In general Tensors with a very large number of dimensions are not as efficient. Cuda will limit you to 25 dimensions if I recall correctly. And on cpu, you might hit slowdowns.
But to answer your question, sum does expect a list/tuple and does not accept a generator (what range returns).

1 Like
3

Thanks. But there is no fundamental reason why it could not accept a range, right?

I opened a feature request issue:

4

I don’t think there is any reason for it to accept the range because it will check all the values before doing work.

5

Convenience? range(n) is better to type and read than list(range(n)).

6

Right. But that is a different argument than the one you mentionned above which was " But this seems inefficient for large n .".

1 Like
7

Yes, it’s a different argument, sorry. The efficiency seemed more important to me at first but as you explained that’s not so important. But even if that doesn’t matter, convenience is also a plus :wink:. I have updated the Github issue.

1 Like