What's the best way of checking whether a torch::Tensor is empty?

Question

I'm currently using the has_storage() method to check whether a tensor is empty or not, but I wonder if there is anything better other than this! and whether there are any implications involved in using this other than the fact that an initialized torch::Tensor always has a storage while an empty one doesn't!

Answer 1

After some digging, it turns out that the best solution for this is to use .numel() method which returns the number of elements a tensor has.
In summary:

• To know whether a tensor is allocated (type and storage), use defined().
• To know whether an allocated tensor has zero elements, use numel()
• To know whether a tensor is allocated and whether it has zero elements, use defined() and then numel()

Side note:
An empty tensor (that is the one created using torch::Tensor t; for example) returns zero when .numel() is used, while the size/sizes will result in an exception.

This is a perfect check for such cases where an empty tensor (in the sense I just explained above) is returned. One can simply do:

if (!tensor.numel())
{
    std::cout<<"tensor is empty!" << std::endl;
    // do other checks you wish to do
}

reference

Answer 2

Yes there is a small nuance here : all tensors do not have the same underlying implementation, and some implementations will have has_storage return false no matter what. This is in particular the case for sparse tensor (see here).

However I am not aware of any better way. Just be sure to correctly track your sparse tensors if you use them (and your opaque tensors, if you ever need whatever they are ^^)