segmented_sum

[pevnak]

Hello,
is there a simple way, how I can use segmented_sum in the model? I would like to use Flux to implement multi-instance learning problems as described in
Using Neural Network Formalism to Solve Multiple-Instance Problems.

Thank you very much for help.
Tomas

[MikeInnes]

You may want to mention it to @alha02 who's implementing a lot of tf ops in #28. It's pretty easy to do. I should add some docs for this.

[ghost]

@pevnak I added segment_sum op to my TODO-list in #28

[pevnak]

Thanks a lot. I am curious how flux will look like with it.

[pevnak]

For those interested in, it can be easily added as multiplication with sparse vector:

function sparse_segmented_mean(x,segments::Vector{UnitRange{Int64}})
s = zeros(eltype(x),size(x,2),length(segments))
for (i,segment) in enumerate(segments)
s[segment,i]=1./length(segment)
end
x*sparse(s)
end

or slightly faster version

function fast_segmented_mean(x,segments::Vector{UnitRange{Int64}})
rowval = collect(1:size(x,2))
nzval = zeros(eltype(x),size(x,2))
colptr = zeros(Int,length(segments)+1)
colptr[end] = size(x,2)+1
for (i,segment) in enumerate(segments)
colptr[i] = segment.start
nzval[segment] = 1./length(segment)
end
m = size(x,2)
n = length(segments)
x*SparseMatrixCSC(m,n,colptr,rowval,nzval)
end

[MikeInnes]

FWIW, while I don't know your use case exactly, this seems like a fairly strange construction to me. Often TensorFlow has things like this because it can't handle things like variable-length lists of tensors. So in Flux it might be possible to just use an array of matrices, and map the appropriate function over that list.