Slicing tensor using boolean list

1

I have a boolean Python list that I’d like to use as a “mask” for a tensor (of the same size as the list), returning the entries of the tensor where the list is true.

For instance, given the list mask = [True, False, True] and the tensor x = Tensor([1, 2, 3]), I would like to get the tensor y = Tensor([1, 3]). In numpy, this would be simply y = x[mask], but in PyTorch indexing tensors with lists is not (yet?) supported.

Moreover, I need an efficient implementation for this slicing, since this would be performed in every forward pass of my model. What do you suggest?

Thank you in advance.

2

Well, currently my solution is:

idx = np.argwhere(np.asarray(mask))
y = x[idx]

Is there any better option?

3

Hi Diogo -

We currently have a subset of support for numpy indexing but what I believe you are referring to is Boolean array indexing as described here: https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html#boolean-array-indexing.

The closest PyTorch equivalent is masked select, as described here: http://pytorch.org/docs/master/torch.html?highlight=masked_select#torch.masked_select. If you could instead generate your mask in the ByteTensor format described there, you could achieve what you want.

7 Likes
4

Thanks, @killeent, this perfectly does the job!

5 
y=torch.arange(0,3)
x=torch.Tensor([True,False,True])==True
print(y[x])

this works

6 Likes
6

This solution is better. Assume you have a two dimensional tensor y=torch.range(1, 8).reshape(2, 4), and x=torch.Tensor([True,False])==True. In this case, you can’t use torch.masked_select, but you can easily execute y[x, :].