Is_floating_point method/property

1

I wanted to do a check whether a tensor is a floating point (torch.half, torch.float, or torch.double) and noticed there is a method torch.Tensor.is_floating_point() and a property torch.Tensor.dtype.is_floating_point. Is one of these preferred to the other? Also is there are reason these are not documented?

2

This is my first time hearing about both of them. I’ve filed an issue at https://github.com/pytorch/pytorch/issues/15700 to get them documented.

Looking at the implementations, my guess is that torch.Tensor.dtype.is_floating_point might be faster than torch.Tensor.is_floating_point() but otherwise I don’t see a difference

3

There are three ways of checking is_floating_point. They have been listed below in the increasing order of time taken to return

  • torch.Tensor.dtype.is_floating_point
  • torch.Tensor.is_floating_point() (2x slower than the fastest)
  • torch.is_floating_point() (3x slower than the fastest)
1 Like