[Solved] Most efficient way to index a tensor from a 1-d mask?

I have a tensor/vector of 1s and 0s which show which examples are ‘alive’, like:

a = torch.FloatTensor([1, 0, 1])

then I have a tensor I want to somehow retrieve these indexes rows from, in as efficient a way as possible, whilst ideally having not too convoluted code. eg I have tensor:

In [7]: b = torch.rand(3, 2)

In [8]: b

 0.5253  0.4571
 0.9760  0.0465
 0.3184  0.7277
[torch.FloatTensor of size 3x2]

I came up with two candidate ways to index into this:

In [46]: b.index_select(0, a.nonzero().long().view(-1))

 0.5253  0.4571
 0.3184  0.7277
[torch.FloatTensor of size 2x2]

This one needs two ops: the nonzero and then the index_select. There’s also a long cast, which will presumably involve a data copy.


In [48]: b.masked_select(a.view(-1, 1).byte())

[torch.FloatTensor of size 4]

This is only a single op (I’m happy to store the mask vector as byte, so that cast could be removed). On the downside the output needs to be re-reshaped back again. Which is free, but does mean an extra variable somewhere, to store the original shape.

It looks like the second is probably more efficient?

  • what are standard ways of doing this?
  • most concise ways of writing?
  • most data efficient?

With the new release, you can also do:

a = torch.FloatTensor([1,2])
b = torch.rand(3,2)

It gives the same output than the one you obtained with index_select.
However, I have no idea of what is the most efficient way.

1 Like

discovered that :slight_smile: . I think finally I will use this method, ie b[a_idxes], since:

  • actually, I only need to form a_idxes once, but there are a zillion different b tensors, all of which can be indexed using the exact same a_idxes
  • using b[a_idxes] notation allows update of the original tensors, at the end of the batch

=> going to rename to ‘[Closed]’

1 Like