Advanced Indexing Error

[snowyday]

Grateful thanks for new various features!

Advanced Indexing does not support vector?:

>>> x = torch.Tensor(5)
>>> x[[0, 3]]

Traceback (most recent call last):
File "", line 1, in
TypeError: indexing a tensor with an object of type list. The only supported types are integers, slices,, numpy scalars and torch.LongTensor or torch.ByteTensor as the only argument.

[chenzhekl]

Related to #2305. Use x[[0, 3], ] instead.

[snowyday]

@chenzhekl Thank you!!
But, I think the expression is non-trivial.
Numpy supports both expressions:

>>> numpy.zeros(5)
array([ 0.,  0.,  0.,  0.,  0.])
>>> x = numpy.zeros(5)
>>> x[[0, 3]]
array([ 0.,  0.])
>>> x[[0, 3], ]
array([ 0.,  0.])

Is there any reason not to support the simple expression such as x[[0, 3]]?

[killeent]

@snowyday by definition, advanced indexing requires a sequence, so the behavior without sequences in PyTorch is unchanged. The currently supported way of doing what you want is:

x = torch.Tensor(5)
x[torch.LongTensor([0, 3])]

[fmassa]

@killeent Numpy seems not to use tuples for advanced indexing, but lists or arrays
So

a = np.random.rand(3,3)
a[[1, 2]] == a[np.array([1, 2])] != a[(1, 2)]

Can't we modify the behavior in pytorch to be more in line with numpy's?

[apaszke]

We should fix that to work as in numpy

[killeent]

@fmassa can you clarify what the difference is in behavior for PyTorch?

[killeent]

Actually, to correct myself, I guess this is considered a form of advanced indexing, so I will address this either way...

[vadimkantorov]

A few other things that don't work:

a = torch.zeros(2, 2)

a[[0,1], 0]                        # doesn't work
a[torch.LongTensor([0, 1]), 0]     # doesn't work
a[torch.LongTensor([0, 1])][:, 0]  # works

a[[0, 1]][:, 0]                    # doesn't work
a[[0, 1], :1]                      # doesn't work
a[[0, 1], 0:1]                     # doesn't work
a[[0, 1], 0:2]                     # works

[killeent]

@vadimkantorov this is to be expected. We only support a subset of advanced indexing - including no mixing of basic and advanced indexing, or support for non-empty slice objects.

[snowyday]

@killeent Thank you so much for #2323!

[micklexqg]

@chenzhekl , why x[[0,3],] still does not work....

[fmassa]

@micklexqg are you using pytorch 0.4?

[micklexqg]

@fmassa , no, maybe 0.2. so it need to be under pytorch 0.4?

[fmassa]

Preferably. Full support for advanced indexing was added in 0.4
Previous versions had some support for it, but many cases were missing