cpp mx::slice

[mureva]

I have a tensor as (batches, rows, cols, channels) which I want to slice, but I'm slightly confused by the logic for taking a whole axis. I've been managing so far by just knowing the size of the whole axis, thus if I have an input of (16,128,128,3) and I want to get just the middle channel, I could easily do:

mx::Shape begin(0,0,0,1);
mx::Shape end(16,128,128,2);
auto ch = mx::slice( data, begin, end );

But I don't want to have to explicitly state the end for axes that don't change, because it can confuse things when I later want to say, use a different batch size during inference than I did during training.

My impression from the docs was that I could maybe use -1, as in:

mx::Shape begin(0,0,0,1);
mx::Shape end(-1,-1,-1,2);
auto ch = mx::slice( data, begin, end );

but then my output shape seems to be (15,127,127,1) - so I've lost a whole image from my batch, and also a whole row and column from each channel. That suggested to me that maybe I want to use mx::Shape end(0,0,0,2) but then my output shape is (-1,-1,-1,1).

So how do I say "use this whole axis"?

Note, I know I could in this trivial case use slice_axis, but my real task is more like:

mx::Shape step(1,2,2,1);
mx::Shape begin(0,0,0,0);
mx::Shape end(64,128,128,1);
mx::slice( data, begin, end, step );

(I'm also confused by how -ve indices are said to be possible when mx::Shape appears to use t_index as a datatype, and t_index appears to be defined as typedef unsigned t_index)