ton/IronpenGPU
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

building knowledgeFn in GPU format

Browse Source
This commit is contained in:
ton
2023-07-20 15:51:25 +07:00
parent 2e34679f73
commit 6b71450055
3 changed files with 86 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

78
src/forward.jl
View File

@@ -17,14 +17,17 @@ function (kfn::kfn_1)(input::AbstractArray)
end
println(">>> input ", size(input))
println(">>> z_i_t1 ", size(kfn.z_i_t1))
# pass input_data into input neuron.
GeneralUtils.cartesianAssign!(kfn.z_i_t1, input)
println(">>> z_i_t1 ", size(kfn.z_i_t1))
GeneralUtils.cartesianAssign!(kfn.z_i_t, input)
kfn.lif_z_i_t = GeneralUtils.matMul_3Dto4D_batchwise(kfn.z_i_t,
ones(size(kfn.z_i_t)[1], size(kfn.z_i_t)[2], size(kfn.lif_w)[3], size(kfn.z_i_t)[3]))
println(">>> z_i_t ", size(kfn.z_i_t))
println(">>> lif_z_i_t ", size(kfn.lif_z_i_t))
println(">>> lif_recSignal ", size(kfn.lif_recSignal))
println(">>> lif_w ", size(kfn.lif_w))
println(">>> lif_refractoryActive ", size(kfn.lif_refractoryActive))
println(">>> lif_refractoryActive ", size(kfn.lif_refractoryCounter))
println(">>> lif_alpha ", size(kfn.lif_alpha))
println(">>> lif_vt0 ", size(kfn.lif_vt0))
println(">>> lif_vt0 sum ", sum(kfn.lif_vt0))
@@ -33,13 +36,14 @@ function (kfn::kfn_1)(input::AbstractArray)
refractoryStatus!(kfn.lif_refractoryCounter, kfn.lif_refractoryActive, kfn.lif_refractoryInactive)
refractoryStatus!(kfn.alif_refractoryCounter, kfn.alif_refractoryActive, kfn.alif_refractoryInactive)
# LIF forward active neurons
kfn.lif_recSignal .= GeneralUtils.sumAlongDim3(
GeneralUtils.matMul_3Dto4D_batchwise(kfn.z_i_t1, kfn.lif_refractoryActive .* kfn.lif_w))
kfn.lif_vt1 = (kfn.lif_alpha .* kfn.lif_vt0) .+ kfn.lif_recSignal
# for (i, v) in enumerate(kfn.lif_vt1)
# if v <
# LIF forward inactive neurons
#WORKING LIF forward active neurons
# a = kfn.lif_refractoryActive .* kfn.lif_w
# lifForward.(kfn.lif_refractoryCounter, kfn.z_i_t0, kfn.z_i_t1,
# kfn.lif_vt0, kfn.lif_vt1, kfn.lif_alpha, kfn.lif_recSignal)
# kfn.lif_recSignal .= GeneralUtils.sumAlongDim3(
# GeneralUtils.matMul_3Dto4D_batchwise(kfn.z_i_t1, kfn.lif_refractoryActive .* kfn.lif_w))
# kfn.lif_vt1 = (kfn.lif_alpha .* kfn.lif_vt0) .+ kfn.lif_recSignal
@@ -52,19 +56,53 @@ function (kfn::kfn_1)(input::AbstractArray)
error("debug end kfn forward")
end
function lifForward(lif_refractoryCounter, z_i_t0, z_i_t1, lif_w, lif_vt0, lif_vt1, lif_alpha,
lif_recSignal)
error("debug end LIF forward")
# if n.refractoryCounter != 0
# n.refractoryCounter -= 1
# # neuron is in refractory state, skip all calculation
# n.z_t1 = false # used by timestep_forward() in kfn. Set to zero because neuron spike
# # last only 1 timestep follow by a period of refractory.
# n.recSignal = n.recSignal * 0.0
# # decay of v_t1
# n.v_t1 = n.alpha * n.v_t
# n.phi = 0.0
# n.decayedEpsilonRec = n.alpha * n.epsilonRec
# n.epsilonRec = n.decayedEpsilonRec
# else
# n.recSignal = sum(n.wRec .* n.z_i_t) # signal from other neuron that this neuron subscribed
# # computeAlpha!(n)
# n.alpha_v_t = n.alpha * n.v_t
# n.v_t1 = n.alpha_v_t + n.recSignal
# # n.v_t1 = no_negative!(n.v_t1)
# if n.v_t1 > n.v_th
# n.z_t1 = true
# n.refractoryCounter = n.refractoryDuration
# n.firingCounter += 1
# n.v_t1 = n.vRest
# else
# n.z_t1 = false
# end
# # there is a difference from alif formula
# n.phi = (n.gammaPd / n.v_th) * max(0, 1 - (n.v_t1 - n.v_th) / n.v_th)
# n.decayedEpsilonRec = n.alpha * n.epsilonRec
# n.epsilonRec = n.decayedEpsilonRec + n.z_i_t
# end
end

2
src/snnUtil.jl
View File

@@ -13,7 +13,7 @@ function refractoryStatus!(refractoryCounter, refractoryActive, refractoryInacti
if refractoryCounter[1, 1, i, j] > 0 # inactive
view(refractoryActive, 1, 1, i, j) .= 0
view(refractoryInactive, 1, 1, i, j) .= 1
else
else # active
view(refractoryActive, 1, 1, i, j) .= 1
view(refractoryInactive, 1, 1, i, j) .= 0
end

38
src/type.jl
View File

@@ -21,10 +21,14 @@ Base.@kwdef mutable struct kfn_1 <: knowledgeFn
timeStep::AbstractArray = [0]
learningStage::AbstractArray = [0] # 0 inference, 1 start, 2 during, 3 end learning
z_i_t1::Union{AbstractArray, Nothing} = nothing # 2D activation matrix
z_i_t0::Union{AbstractArray, Nothing} = nothing
z_i_t::Union{AbstractArray, Nothing} = nothing # 3D activation matrix
# ---------------------------------------------------------------------------- #
# LIF #
# ---------------------------------------------------------------------------- #
# a projection of kfn.z_i_t into lif dimension for broadcasting later)
lif_z_i_t::Union{AbstractArray, Nothing} = nothing
# LIF
lif_w::Union{AbstractArray, Nothing} = nothing
lif_recSignal::Union{AbstractArray, Nothing} = nothing
lif_vt0::Union{AbstractArray, Nothing} = nothing
@@ -39,7 +43,9 @@ Base.@kwdef mutable struct kfn_1 <: knowledgeFn
lif_delta::AbstractFloat = 1.0
lif_tau_m::AbstractFloat = 20.0
# ALIF
# ---------------------------------------------------------------------------- #
# ALIF #
# ---------------------------------------------------------------------------- #
alif_w::Union{AbstractArray, Nothing} = nothing
alif_recSignal::Union{AbstractArray, Nothing} = nothing
alif_zt0::Union{AbstractArray, Nothing} = nothing
@@ -55,7 +61,8 @@ end
function kfn_1(params::Dict)
kfn = kfn_1()
kfn.params = params
# initialize activation matrix
# ----------------------- initialize activation matrix ----------------------- #
# row*col is a 2D matrix represent all RSNN activation
row, col, batch = kfn.params[:inputPort][:signal][:numbers] # z-axis represent signal batch number
row += kfn.params[:inputPort][:noise][:numbers][1]
col += kfn.params[:inputPort][:signal][:numbers][2]
@@ -63,12 +70,13 @@ function kfn_1(params::Dict)
col += kfn.params[:computeNeuron][:alif][:numbers][2]
# activation matrix
kfn.z_i_t0 = zeros(row, col, batch)
kfn.z_i_t1 = zeros(row, col, batch)
kfn.z_i_t = zeros(row, col, batch)
# LIF
# -------------------------------- LIF config -------------------------------- #
# In 3D LIF matrix, z-axis represent each neuron while each 2D slice represent that neuron's
# synaptic subscription to other neurons (via activation matrix)
z = kfn.params[:computeNeuron][:lif][:numbers][1] * kfn.params[:computeNeuron][:lif][:numbers][2]
kfn.lif_w = zeros(row, col, z) # matrix z-axis represent each neurons
kfn.lif_recSignal = zeros(1, 1, z, batch)
kfn.lif_vt0 = zeros(1, 1, z, batch)
kfn.lif_vt1 = zeros(1, 1, z, batch)
@@ -81,19 +89,23 @@ function kfn_1(params::Dict)
kfn.lif_alpha = ones(1, 1, z, batch) .* (exp(-kfn.lif_delta / kfn.lif_tau_m))
# subscription
row, col, _ = size(kfn.lif_w) # row*col is synaptic subscribe weight for each neuron in z-axis
w = zeros(row, col, z)
synapticConnectionPercent = kfn.params[:computeNeuron][:lif][:params][:synapticConnectionPercent]
synapticConnection = Int(floor(row*col * synapticConnectionPercent/100))
for slice in eachslice(kfn.lif_w, dims=3)
for slice in eachslice(w, dims=3)
pool = shuffle!([1:row*col...])[1:synapticConnection]
for i in pool
slice[i] = randn()/10
slice[i] = randn()/10 # assign weight to synaptic connection
end
end
#WORKING project 3D w into 4D kfn.lif_w
kfn.lif_w = reshape(w, (row, col, z, 1)) .* ones(row, col, z, batch)
println(">>> lif_w ", size(kfn.lif_w))
error("end WORKING")
# ALIF
z = kfn.params[:computeNeuron][:alif][:numbers][1] * kfn.params[:computeNeuron][:alif][:numbers][2]
kfn.alif_w = zeros(row, col, z)
kfn.alif_w = zeros(row, col, z) # matrix z-axis represent each neurons
kfn.alif_recSignal = zeros(1, 1, z, batch)
kfn.alif_zt0 = zeros(1, 1, z, batch)
kfn.alif_zt1 = zeros(1, 1, z, batch)