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

dev

Browse Source
This commit is contained in:
ton
2023-09-10 11:28:40 +07:00
parent fb3e59a414
commit e7c0228313
5 changed files with 384 additions and 155 deletions
Download Patch File Download Diff File Expand all files Collapse all files

404
src/learn.jl
View File

@@ -267,30 +267,32 @@ end
function learn!(kfn::kfn_1, device=cpu)
# lif learn
kfn.lif_wRec, kfn.lif_neuronInactivityCounter, kfn.lif_synapticActivityCounter =
kfn.lif_wRec, kfn.lif_neuronInactivityCounter, kfn.lif_synapseReconnectDelayCounter =
lifLearn(kfn.lif_wRec,
kfn.lif_exInType,
kfn.lif_wRecChange,
kfn.lif_exInType,
kfn.lif_arrayProjection4d,
kfn.lif_neuronInactivityCounter,
kfn.lif_synapticActivityCounter,
kfn.lif_synapseReconnectDelayCounter,
kfn.lif_synapseConnectionNumber,
kfn.lif_synapticWChangeCounter,
kfn.lif_eta,
kfn.lif_vt,
kfn.zitCumulative,
device)
# alif learn
kfn.alif_wRec, kfn.alif_neuronInactivityCounter, kfn.alif_synapticActivityCounter =
kfn.alif_wRec, kfn.alif_neuronInactivityCounter, kfn.alif_synapseReconnectDelayCounter =
alifLearn(kfn.alif_wRec,
kfn.alif_exInType,
kfn.alif_wRecChange,
kfn.alif_exInType,
kfn.alif_arrayProjection4d,
kfn.alif_neuronInactivityCounter,
kfn.alif_synapticActivityCounter,
kfn.alif_synapseReconnectDelayCounter,
kfn.alif_synapseConnectionNumber,
kfn.alif_synapticWChangeCounter,
kfn.alif_eta,
kfn.alif_vt,
kfn.zitCumulative,
device)
@@ -306,146 +308,178 @@ function learn!(kfn::kfn_1, device=cpu)
# error("DEBUG -> kfn learn! $(Dates.now())")
end
# function lifLearn(wRec,
# exInType,
# wRecChange,
# arrayProjection4d,
# neuronInactivityCounter,
# synapseReconnectDelayCounter,
# synapseConnectionNumber,
# synapticWChangeCounter, #TODO
# eta,
# zitCumulative,
# device)
# # merge learning weight with average learning weight of all batch
# wch = sum(wRecChange, dims=4) ./ (size(wRec, 4)) .* arrayProjection4d
# wRec .= (exInType .* wRec) .+ wch
# arrayProjection4d_cpu = arrayProjection4d |> cpu
# wRec_cpu = wRec |> cpu
# wRec_cpu = wRec_cpu[:,:,:,1] # since every batch has the same neuron wRec, (row, col, n)
# eta_cpu = eta |> cpu
# eta_cpu = eta_cpu[:,:,:,1]
# neuronInactivityCounter_cpu = neuronInactivityCounter |> cpu
# neuronInactivityCounter_cpu = neuronInactivityCounter_cpu[:,:,:,1] # (row, col, n)
# synapseReconnectDelayCounter_cpu = synapseReconnectDelayCounter |> cpu
# synapseReconnectDelayCounter_cpu = synapseReconnectDelayCounter_cpu[:,:,:,1]
# zitCumulative_cpu = zitCumulative |> cpu
# zitCumulative_cpu = zitCumulative_cpu[:,:,1] # (row, col)
# # -W if less than 10% of repeat avg, +W otherwise
# _, _, i3 = size(wRec_cpu)
# for i in 1:i3
# x = 0.1 * (sum(synapseReconnectDelayCounter[:,:,i]) / length(synapseReconnectDelayCounter[:,:,i]))
# mask = GeneralUtils.replaceLessThan.(wRec_cpu[:,:,i], x, -1, 1)
# wRec_cpu[:,:,i] .+= mask .* eta_cpu[:,:,i] .* wRec_cpu[:,:,i]
# end
# # weak / negative synaptic connection will get randomed in neuroplasticity()
# wRec_cpu = GeneralUtils.replaceBetween.(wRec_cpu, 0.0, 0.01, -1.0) # mark with -1.0
# # neuroplasticity, work on CPU side
# wRec_cpu = neuroplasticity(synapseConnectionNumber,
# zitCumulative_cpu,
# wRec_cpu,
# neuronInactivityCounter_cpu,
# synapseReconnectDelayCounter_cpu)
# wRec_cpu = wRec_cpu .* arrayProjection4d_cpu
# wRec = wRec_cpu |> device
# neuronInactivityCounter_cpu = neuronInactivityCounter_cpu .* arrayProjection4d_cpu
# neuronInactivityCounter = neuronInactivityCounter_cpu |> device
# synapseReconnectDelayCounter_cpu = synapseReconnectDelayCounter_cpu .* arrayProjection4d_cpu
# synapseReconnectDelayCounter = synapseReconnectDelayCounter_cpu |> device
# return wRec, neuronInactivityCounter, synapseReconnectDelayCounter
# end
function lifLearn(wRec,
exInType,
wRecChange,
exInType,
arrayProjection4d,
neuronInactivityCounter,
synapticActivityCounter,
synapseReconnectDelayCounter,
synapseConnectionNumber,
synapticWChangeCounter, #TODO
eta,
vt,
zitCumulative,
device)
# merge learning weight with average learning weight of all batch
wch = sum(wRecChange, dims=4) ./ (size(wRec, 4)) .* arrayProjection4d
wRec .= (exInType .* wRec) .+ wch
# transfer data to cpu
arrayProjection4d_cpu = arrayProjection4d |> cpu
wRec_cpu = wRec |> cpu
wRec_cpu = wRec_cpu[:,:,:,1] # since every batch has the same neuron wRec, (row, col, n)
wRecChange_cpu = wRecChange |> cpu
wRecChange_cpu = wRecChange_cpu[:,:,:,1]
eta_cpu = eta |> cpu
eta_cpu = eta_cpu[:,:,:,1]
neuronInactivityCounter_cpu = neuronInactivityCounter |> cpu
neuronInactivityCounter_cpu = neuronInactivityCounter_cpu[:,:,:,1] # (row, col, n)
synapticActivityCounter_cpu = synapticActivityCounter |> cpu
synapticActivityCounter_cpu = synapticActivityCounter_cpu[:,:,:,1]
synapseReconnectDelayCounter_cpu = synapseReconnectDelayCounter |> cpu
synapseReconnectDelayCounter_cpu = synapseReconnectDelayCounter_cpu[:,:,:,1]
zitCumulative_cpu = zitCumulative |> cpu
zitCumulative_cpu = zitCumulative_cpu[:,:,1] # (row, col)
zitCumulative_cpu = zitCumulative_cpu[:,:,1]
# -W if less than 10% of repeat avg, +W otherwise
_, _, i3 = size(wRec_cpu)
for i in 1:i3
x = 0.1 * (sum(synapticActivityCounter[:,:,i]) / length(synapticActivityCounter[:,:,i]))
mask = GeneralUtils.replaceLessThan.(wRec_cpu[:,:,i], x, -1, 1)
wRec_cpu[:,:,i] .+= mask .* eta_cpu[:,:,i] .* wRec_cpu[:,:,i]
end
#TODO neuroplasticity, work on CPU side
wRec_cpu, neuronInactivityCounter_cpu, synapseReconnectDelayCounter_cpu,
= neuroplasticity(synapseConnectionNumber,
zitCumulative_cpu,
wRec_cpu,
wRecChange_cpu,
vt,
neuronInactivityCounter_cpu,
synapseReconnectDelayCounter_cpu)
# weak / negative synaptic connection will get randomed in neuroplasticity()
wRec_cpu = GeneralUtils.replaceBetween.(wRec_cpu, 0.0, 0.01, -1.0) # mark with -1.0
# # merge learning weight with average learning weight of all batch
# wch = sum(wRecChange, dims=4) ./ (size(wRec, 4)) .* arrayProjection4d
# wRec .= (exInType .* wRec) .+ wch
# # (row, col)
# # -W if less than 10% of repeat avg, +W otherwise
# _, _, i3 = size(wRec_cpu)
# for i in 1:i3
# x = 0.1 * (sum(synapseReconnectDelayCounter[:,:,i]) / length(synapseReconnectDelayCounter[:,:,i]))
# mask = GeneralUtils.replaceLessThan.(wRec_cpu[:,:,i], x, -1, 1)
# wRec_cpu[:,:,i] .+= mask .* eta_cpu[:,:,i] .* wRec_cpu[:,:,i]
# end
# # weak / negative synaptic connection will get randomed in neuroplasticity()
# wRec_cpu = GeneralUtils.replaceBetween.(wRec_cpu, 0.0, 0.01, -1.0) # mark with -1.0
# neuroplasticity, work on CPU side
wRec_cpu = neuroplasticity(synapseConnectionNumber,
zitCumulative_cpu,
wRec_cpu,
neuronInactivityCounter_cpu,
synapticActivityCounter_cpu)
# wRec_cpu = neuroplasticity(synapseConnectionNumber,
# zitCumulative_cpu,
# wRec_cpu,
# wRecChange_cpu,
# vt,
# neuronInactivityCounter_cpu,
# synapseReconnectDelayCounter_cpu)
# transfer data backto gpu
wRec_cpu = wRec_cpu .* arrayProjection4d_cpu
wRec = wRec_cpu |> device
neuronInactivityCounter_cpu = neuronInactivityCounter_cpu .* arrayProjection4d_cpu
neuronInactivityCounter = neuronInactivityCounter_cpu |> device
synapticActivityCounter_cpu = synapticActivityCounter_cpu .* arrayProjection4d_cpu
synapticActivityCounter = synapticActivityCounter_cpu |> device
synapseReconnectDelayCounter_cpu = synapseReconnectDelayCounter_cpu .* arrayProjection4d_cpu
synapseReconnectDelayCounter = synapseReconnectDelayCounter_cpu |> device
return wRec, neuronInactivityCounter, synapticActivityCounter
end
function alifLearn(wRec,
exInType,
wRecChange,
arrayProjection4d,
neuronInactivityCounter,
synapticActivityCounter,
synapseConnectionNumber,
synapticWChangeCounter, #TODO
eta,
zitCumulative,
device)
# merge learning weight with average learning weight of all batch
wch = sum(wRecChange, dims=4) ./ (size(wRec, 4)) .* arrayProjection4d
wRec .= (exInType .* wRec) .+ wch
arrayProjection4d_cpu = arrayProjection4d |> cpu
wRec_cpu = wRec |> cpu
wRec_cpu = wRec_cpu[:,:,:,1] # since every batch has the same neuron wRec, (row, col, n)
eta_cpu = eta |> cpu
eta_cpu = eta_cpu[:,:,:,1]
neuronInactivityCounter_cpu = neuronInactivityCounter |> cpu
neuronInactivityCounter_cpu = neuronInactivityCounter_cpu[:,:,:,1] # (row, col, n)
synapticActivityCounter_cpu = synapticActivityCounter |> cpu
synapticActivityCounter_cpu = synapticActivityCounter_cpu[:,:,:,1]
zitCumulative_cpu = zitCumulative |> cpu
zitCumulative_cpu = zitCumulative_cpu[:,:,1] # (row, col)
# -W if less than 10% of repeat avg, +W otherwise
_, _, i3 = size(wRec_cpu)
for i in 1:i3
x = 0.1 * (sum(synapticActivityCounter[:,:,i]) / length(synapticActivityCounter[:,:,i]))
mask = GeneralUtils.replaceLessThan.(wRec_cpu[:,:,i], x, -1, 1)
wRec_cpu[:,:,i] .+= mask .* eta_cpu[:,:,i] .* wRec_cpu[:,:,i]
end
# weak / negative synaptic connection will get randomed in neuroplasticity()
wRec_cpu = GeneralUtils.replaceBetween.(wRec_cpu, 0.0, 0.01, -1.0) # mark with -1.0
# neuroplasticity, work on CPU side
wRec_cpu = neuroplasticity(synapseConnectionNumber,
zitCumulative_cpu,
wRec_cpu,
neuronInactivityCounter_cpu,
synapticActivityCounter_cpu)
wRec_cpu = wRec_cpu .* arrayProjection4d_cpu
wRec = wRec_cpu |> device
neuronInactivityCounter_cpu = neuronInactivityCounter_cpu .* arrayProjection4d_cpu
neuronInactivityCounter = neuronInactivityCounter_cpu |> device
synapticActivityCounter_cpu = synapticActivityCounter_cpu .* arrayProjection4d_cpu
synapticActivityCounter = synapticActivityCounter_cpu |> device
# error("DEBUG -> alifLearn! $(Dates.now())")
return wRec, neuronInactivityCounter, synapticActivityCounter
end
function onLearn!(wOut,
wOutChange,
arrayProjection4d)
# merge learning weight with average learning weight
wOut .+= (sum(wOutChange, dims=4) ./ (size(wOut, 4))) .* arrayProjection4d
# adaptive wOut to help convergence using c_decay
wOut .-= 0.001 .* wOut
#TODO synaptic strength
#TODO neuroplasticity
return wRec, neuronInactivityCounter, synapseReconnectDelayCounter
end
#TODO
function neuroplasticity(synapseConnectionNumber,
zitCumulative, # (row, col)
wRec, # (row, col, n)
wRecChange,
vt,
neuronInactivityCounter,
synapticActivityCounter) # (row, col, n)
synapseReconnectDelayCounter) # (row, col, n)
i1,i2,i3 = size(wRec)
# merge weight
# adjust weight based on vt progress and repeatition (90% +w, 10% -w)
# -w all non-fire connection except mature connection
# prune weak connection
# rewire synapse connection
# for each neuron, find total number of synaptic conn that should draw
# new connection to firing and non-firing neurons pool
subToFireNeuron_toBe = Int(floor(0.7 * synapseConnectionNumber))
@@ -460,7 +494,7 @@ function neuroplasticity(synapseConnectionNumber,
println("neuroplasticity, from $(synapseConnectionNumber*size(totalNewConn, 3)) conn, $(sum(totalNewConn)) are replaced")
# clear -1.0 marker
GeneralUtils.replaceElements!(wRec, -1.0, synapticActivityCounter, -0.99)
GeneralUtils.replaceElements!(wRec, -1.0, synapseReconnectDelayCounter, -0.99)
GeneralUtils.replaceElements!(wRec, -1.0, 0.0) # -1.0 marker is no longer required
for i in 1:i3
@@ -473,7 +507,7 @@ function neuroplasticity(synapseConnectionNumber,
a = similar(w) .= -0.99 # synapseConnectionNumber of this neuron
mask = (!iszero).(w)
GeneralUtils.replaceElements!(mask, 1, a, 0)
synapticActivityCounter[:,:,i] = a
synapseReconnectDelayCounter[:,:,i] = a
else
remaining = 0
if subToFireNeuron_current[1,1,i] < subToFireNeuron_toBe
@@ -482,7 +516,7 @@ function neuroplasticity(synapseConnectionNumber,
# add new conn to firing neurons pool
remaining = addNewSynapticConn!(zitMask[:,:,i], 1,
@view(wRec[:,:,i]),
@view(synapticActivityCounter[:,:,i]),
@view(synapseReconnectDelayCounter[:,:,i]),
toAddConn)
totalNewConn[1,1,i] += remaining
end
@@ -490,12 +524,12 @@ function neuroplasticity(synapseConnectionNumber,
# add new conn to non-firing neurons pool
remaining = addNewSynapticConn!(zitMask[:,:,i], 0,
@view(wRec[:,:,i]),
@view(synapticActivityCounter[:,:,i]),
@view(synapseReconnectDelayCounter[:,:,i]),
totalNewConn[1,1,i])
if remaining > 0 # final get-all round if somehow non-firing pool has not enough slot
remaining = addNewSynapticConn!(zitMask[:,:,i], 1,
@view(wRec[:,:,i]),
@view(synapticActivityCounter[:,:,i]),
@view(synapseReconnectDelayCounter[:,:,i]),
remaining)
end
end
@@ -505,6 +539,146 @@ function neuroplasticity(synapseConnectionNumber,
return wRec
end
function alifLearn(wRec,
wRecChange,
exInType,
arrayProjection4d,
neuronInactivityCounter,
synapseReconnectDelayCounter,
synapseConnectionNumber,
synapticWChangeCounter, #TODO
eta,
vt,
zitCumulative,
device)
# merge learning weight with average learning weight of all batch
wch = sum(wRecChange, dims=4) ./ (size(wRec, 4)) .* arrayProjection4d
wRec .= (exInType .* wRec) .+ wch
arrayProjection4d_cpu = arrayProjection4d |> cpu
wRec_cpu = wRec |> cpu
wRec_cpu = wRec_cpu[:,:,:,1] # since every batch has the same neuron wRec, (row, col, n)
eta_cpu = eta |> cpu
eta_cpu = eta_cpu[:,:,:,1]
neuronInactivityCounter_cpu = neuronInactivityCounter |> cpu
neuronInactivityCounter_cpu = neuronInactivityCounter_cpu[:,:,:,1] # (row, col, n)
synapseReconnectDelayCounter_cpu = synapseReconnectDelayCounter |> cpu
synapseReconnectDelayCounter_cpu = synapseReconnectDelayCounter_cpu[:,:,:,1]
zitCumulative_cpu = zitCumulative |> cpu
zitCumulative_cpu = zitCumulative_cpu[:,:,1] # (row, col)
# -W if less than 10% of repeat avg, +W otherwise
_, _, i3 = size(wRec_cpu)
for i in 1:i3
x = 0.1 * (sum(synapseReconnectDelayCounter[:,:,i]) / length(synapseReconnectDelayCounter[:,:,i]))
mask = GeneralUtils.replaceLessThan.(wRec_cpu[:,:,i], x, -1, 1)
wRec_cpu[:,:,i] .+= mask .* eta_cpu[:,:,i] .* wRec_cpu[:,:,i]
end
# weak / negative synaptic connection will get randomed in neuroplasticity()
wRec_cpu = GeneralUtils.replaceBetween.(wRec_cpu, 0.0, 0.01, -1.0) # mark with -1.0
# neuroplasticity, work on CPU side
wRec_cpu = neuroplasticity(synapseConnectionNumber,
zitCumulative_cpu,
wRec_cpu,
neuronInactivityCounter_cpu,
synapseReconnectDelayCounter_cpu)
wRec_cpu = wRec_cpu .* arrayProjection4d_cpu
wRec = wRec_cpu |> device
neuronInactivityCounter_cpu = neuronInactivityCounter_cpu .* arrayProjection4d_cpu
neuronInactivityCounter = neuronInactivityCounter_cpu |> device
synapseReconnectDelayCounter_cpu = synapseReconnectDelayCounter_cpu .* arrayProjection4d_cpu
synapseReconnectDelayCounter = synapseReconnectDelayCounter_cpu |> device
# error("DEBUG -> alifLearn! $(Dates.now())")
return wRec, neuronInactivityCounter, synapseReconnectDelayCounter
end
function onLearn!(wOut,
wOutChange,
arrayProjection4d)
# merge learning weight with average learning weight
wOut .+= (sum(wOutChange, dims=4) ./ (size(wOut, 4))) .* arrayProjection4d
# adaptive wOut to help convergence using c_decay
wOut .-= 0.001 .* wOut
end
# function neuroplasticity(synapseConnectionNumber,
# zitCumulative, # (row, col)
# wRec, # (row, col, n)
# neuronInactivityCounter,
# synapseReconnectDelayCounter) # (row, col, n)
# i1,i2,i3 = size(wRec)
# # for each neuron, find total number of synaptic conn that should draw
# # new connection to firing and non-firing neurons pool
# subToFireNeuron_toBe = Int(floor(0.7 * synapseConnectionNumber))
# # for each neuron, count how many synap already subscribed to firing-neurons
# zw = zitCumulative .* wRec
# subToFireNeuron_current = sum(GeneralUtils.isBetween.(zw, 0.0, 100.0), dims=(1,2)) # (1, 1, n)
# zitMask = (!iszero).(zitCumulative) # zitMask of firing neurons = 1, non-firing = 0
# projection = ones(i1,i2,i3)
# zitMask = zitMask .* projection # (row, col, n)
# totalNewConn = sum(isequal.(wRec, -1.0), dims=(1,2)) # count new conn mark (-1.0), (1, 1, n)
# println("neuroplasticity, from $(synapseConnectionNumber*size(totalNewConn, 3)) conn, $(sum(totalNewConn)) are replaced")
# # clear -1.0 marker
# GeneralUtils.replaceElements!(wRec, -1.0, synapseReconnectDelayCounter, -0.99)
# GeneralUtils.replaceElements!(wRec, -1.0, 0.0) # -1.0 marker is no longer required
# for i in 1:i3
# if neuronInactivityCounter[1:1:i][1] < -10000 # neuron die i.e. reset all weight
# println("neuron die")
# neuronInactivityCounter[:,:,i] .= 0 # reset
# w = random_wRec(i1,i2,1,synapseConnectionNumber)
# wRec[:,:,i] .= w
# a = similar(w) .= -0.99 # synapseConnectionNumber of this neuron
# mask = (!iszero).(w)
# GeneralUtils.replaceElements!(mask, 1, a, 0)
# synapseReconnectDelayCounter[:,:,i] = a
# else
# remaining = 0
# if subToFireNeuron_current[1,1,i] < subToFireNeuron_toBe
# toAddConn = subToFireNeuron_toBe - subToFireNeuron_current[1,1,i]
# totalNewConn[1,1,i] = totalNewConn[1,1,i] - toAddConn
# # add new conn to firing neurons pool
# remaining = addNewSynapticConn!(zitMask[:,:,i], 1,
# @view(wRec[:,:,i]),
# @view(synapseReconnectDelayCounter[:,:,i]),
# toAddConn)
# totalNewConn[1,1,i] += remaining
# end
# # add new conn to non-firing neurons pool
# remaining = addNewSynapticConn!(zitMask[:,:,i], 0,
# @view(wRec[:,:,i]),
# @view(synapseReconnectDelayCounter[:,:,i]),
# totalNewConn[1,1,i])
# if remaining > 0 # final get-all round if somehow non-firing pool has not enough slot
# remaining = addNewSynapticConn!(zitMask[:,:,i], 1,
# @view(wRec[:,:,i]),
# @view(synapseReconnectDelayCounter[:,:,i]),
# remaining)
# end
# end
# end
# # error("DEBUG -> neuroplasticity $(Dates.now())")
# return wRec
# end
# learningLiquidity(x) = -0.0001x + 1 # -10000 to +10000; f(x) = -5e-05x+0.5
function learningLiquidity(x)