Bool mask indexing

[x66ccff]

I want to implement an operation like this pytorch code

https://github.com/intell-sci-comput/PTS/blob/07088bda3e4c48fe8ac7eded9341855caa02b545/model/models.py#L46-L48

Note: the dr_mask here is a constant bool mask

This is super fast in pytorch, but i can not figure out why Reactant.jl / Julia can not do this.

I know this is not a problem unique to Reactant.jl, but I would like to know what the solution is on Julia. I'm just getting started with Julia. Many thanks!

[wsmoses]

I think you should be able to basically the same thing in Julia

Just write array[:, mask]

@avik-pal

[x66ccff]

well, but i want to do this on GPU. and the array is pretty large, like shape = (1, 100000). I'm afraid that transfering the array between GPU and CPU is slow.

And, i need to do this multiple times in a loop.

[x66ccff]

What I want to do is to build a forward propagation network like this

https://github.com/intell-sci-comput/PTS/blob/07088bda3e4c48fe8ac7eded9341855caa02b545/model/models.py#L190-L253

But i don't now how to implement the DRlayer here in Julia...

[avik-pal]

This is a bug, we need to special case boolean indexing

[avik-pal]

If the mask is constant, i.e. a julia Array then #457 will solve it. If the mask is dynamic, we need to go down the route of constructing dynamically sized tensors, that will need some more ops like findall to be supported.