New PR for potjans connectome.

[russelljjarvis]

Break fixing change. The file src/contrib/genPlasticWeights-erdos-renyi-potjans.jl

Doesn't look substantially different from default genPlasticWeights-erdos-renyi.jl but it has the a necessary change in its indexs:

    exc_selected = sort(exc_ordered[1:end])
    inh_selected = sort(inh_ordered[1:end])

The file genPotjansConnectivity.jl does away with SparseArrays now possibly making it more compatible with GPU code.

src/contrib/genPotjansConnectivity.jl

I typed most or all unresolved types of: Vectors, and dictionaries and I got a speed up.

I put the invocation to this new use of the plugin connectomes in the test directory.

I put the plugin files in a contrib directory.

Besides changing plugin paths in test/Potjansdata/init.jl

Init.jl needs to calculate cell size, to pre-allocate sizes that need to be known in other init.jl contexts, for the basic default code to work.

scale =1.0/30.0
function get_Ncell(scale=1.0::Float64)
    ccu = Dict{String, Int32}("23E"=>20683,
            "4E"=>21915, 
            "5E"=>4850, 
            "6E"=>14395, 
            "6I"=>2948, 
            "23I"=>5834,
            "5I"=>1065,
            "4I"=>5479)
    ccu = Dict{String, Int32}((k,ceil(Int64,v*scale)) for (k,v) in pairs(ccu))
    Ncells = Int32(sum([i for i in values(ccu)])+1)
    Ne = Int32(sum([ccu["23E"],ccu["4E"],ccu["5E"],ccu["6E"]]))
    Ni = Int32(Ncells - Ne)
    Ncells, Ne, Ni, ccu

end
Ncells,Ne,Ni, ccu = get_Ncell(scale)   

src/cpu/loop.jl Line 240 of loop has to be this,

if post_ci!=0

for loop to tolerate connection matrices that are ragged (ie ragged arrays packed into a dense rectangle will look sparse, they will have empty, zero values).

Introducing that line is the easiest way I know to resolve the conflict.

I use a package called distributions too, to draw PRNGs that are drawn from a distribution.

I don't reverted an unintentional change to the file: src/init.jl. It still has a superficial formatting white space difference, but not a substantial difference.

I thought they where changes that you had made that got mixed in with my PR.

If that is true I could try to delete them.

Also I think this GitHub GUI has a squash merge option too.

If I run julia from the tests directory, include("potconntest.jl")

After some time the algorithm converges on a correlation value of 0.96 and it does testing too.

Loop no. 100, task no. 1
correlation: 0.9610312565565572
elapsed time: 2.321516990661621 sec
firing rate: 33.96662786185487 Hz

trial #97, task #1: 1.46 sec
trial #98, task #1: 1.44 sec
trial #99, task #1: 1.38 sec
trial #100, task #1: 1.43 sec

[bjarthur]

the ragged adjacency matrices are now vectors of vectors. to save memory for networks for which the number of presynaptic synapses is not uniform.

also, the params.jl file now needs a new variable init_code.

lastly, you shouldn't have to change the loop code. why do you make sure that post_ci!=0? that should never be the case.

[russelljjarvis]

"lastly, you shouldn't have to change the loop code. why do you make sure that post_ci!=0? that should never be the case." --at the time it was because presynaptic synapses where not uniform, and it was the only way to get the TrainSpikingNet.jl to evaluate post loading a non uniform presynaptic network, while I was waiting for the network patches you discuss here.

[russelljjarvis]

I made some changes to make the weight matrix a ragged array format.

Still working on this it is a very slow burning PR.