Tensor math logical operations: any() and all() functions

Martin_Mundt · August 23, 2017, 9:22am

Coming from Torch7 I am missing logical operations such as the any() and all() functions to check whether a Boolean Tensor contains either only zeros or ones (e.g. after another logical operation):

github.com

torch/torch7/blob/master/doc/maths.md#torchanya

<a name="torch.maths.dok"></a>
# Math Functions #

Torch provides MATLAB-like functions for manipulating [`Tensor`](tensor.md) objects.  Functions fall into several types of categories:

  * [Constructors](#torch.construction.dok) like [`zeros`](#torch.zeros), [`ones`](#torch.ones);
  * Extractors like [`diag`](#torch.diag)  and [`triu`](#torch.triu);
  * [Element-wise](#torch.elementwise.dok) mathematical operations like [`abs`](#torch.abs) and [`pow`](#torch.pow);
  * [BLAS](#torch.basicoperations.dok) operations;
  * [Column or row-wise operations](#torch.columnwise.dok) like [`sum`](#torch.sum) and [`max`](#torch.max);
  * [Matrix-wide operations](#torch.matrixwide.dok) like [`trace`](#torch.trace) and [`norm`](#torch.norm);
  * [Convolution and cross-correlation](#torch.conv.dok) operations like [`conv2`](#torch.conv2);
  * [Basic linear algebra operations](#torch.linalg.dok) like [`eig`](#torch.eig);
  * [Logical operations](#torch.logical.dok) on `Tensor`s.

By default, all operations allocate a new `Tensor` to return the result.
However, all functions also support passing the target `Tensor`(s) as the first argument(s), in which case the target `Tensor`(s) will be resized accordingly and filled with result.
This property is especially useful when one wants have tight control over when memory is allocated.

The *Torch* package adopts the same concept, so that calling a function directly on the `Tensor` itself using an object-oriented syntax is equivalent to passing the `Tensor` as the optional resulting `Tensor`.

This file has been truncated. show original

I understand I could probably do something like

a = torch.rand(3,3)
torch.sum(torch.lt(a,1)) / torch.numel(a)

and check whether this is 1 or 0 (result is a byte tensor so this should be integer division), but it feels rather inconvenient.

Is there some equivalent in PyTorch and if not, would it be possible to add these functions?

smth · August 23, 2017, 9:43am

any and all functions are defined on Byte Tensors.

torch.randn(10).byte().any()

Martin_Mundt · August 23, 2017, 10:40am

thanks for the super quick response. That’s exactly what I need.

Sorry if I didn’t look hard enough and my question was perhaps trivial, but I don’t seem to be able to find it in the documentation neither under tensor, nor where all the other comparison ops are listed in math:

http://pytorch.org/docs/master/torch.html#comparison-ops

smth · August 23, 2017, 8:48pm

it’s not in the docs. we’ll add it (I recently got to know that it’s exposed but we missed the docs on them).