Support for slice assign

[louisfd]

I'm bringing back to life issue #663
The proposed solution of slice-scatter in #927 does not seem to be enough.
I've tried turning things around but I don't succeed in accomplishing the following:

let a =  Tensor::new([[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]]);
let b = Tensor::new( [[88.0, 99.0]]);
let new_a = a.slice_assign(b, [1..2, 1..3])
assert_eq!(new_a, [[0.0, 1.0, 2.0], [3.0, 88.0, 99.0]])

I think I'm not able to achieve this with slice_scatter because it only takes into account one dimension at a time.

[EricLBuehler]

Just saw this, but relates to #1279.

[LaurentMazare]

let a =  Tensor::new([[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]]);
let b = Tensor::new( [[88.0, 99.0]]);
let new_a = a.slice_assign(b, [1..2, 1..3])
assert_eq!(a, [[0.0, 1.0, 2.0], [3.0, 88.0, 99.0]])

Thanks for bringing this up. Just to be sure, should the assert_eq in the example be about new_a or about a? Tensors are immutable by design in candle so it's unlikely that we will get the snippet you gave to work on a (we have some thoughts about providing in-place operations for variables rather than tensors but would really prefer avoiding it).
If it's about new_a it's probably easier to have but it would be good to have some example of models using this to see if it isn't easy to get some workaround there? (it seems like a bit of a convoluted operation so may be easier if we can avoid it and use simpler ops instead but that really depends on what you're trying to achieve here)

[louisfd]

Hi Laurent,
Of course I meant new_a, sorry for the typo. I fixed it now.
In Burn we use slice_assign for instance in autoregressive attention and LSTM.
But even more important is that it's in the backward pass of our equivalent of Candle's narrow operator.

[EricLBuehler]

In Burn we use slice_assign for instance in autoregressive attention and LSTM.

candle-sampling needs this slice_assign feature for the implementation of Beam search.

[LaurentMazare]

Hi Laurent, Of course I meant new_a, sorry for the typo. I fixed it now. In Burn we use slice_assign for instance in autoregressive attention and LSTM. But even more important is that it's in the backward pass of our equivalent of Candle's narrow operator.

Ok, I've merged #1377 that adds some pretty naive implementation for slice_assign, it's quite inefficient but avoids introducing some "base op". If this ends up being a bottleneck for performance, we could consider making something more efficient (but I kind of doubt that this will end up being in the performance critical op).

[louisfd]

Thanks a lot, it works fine!

[LaurentMazare]

Closing again as hopefully all good now :)