def ConfigTstTrlLog(ss, dt):
dt.SetMetaData("name", "TstTrlLog")
dt.SetMetaData("desc", "Record of testing per input pattern")
dt.SetMetaData("read-only", "true")
dt.SetMetaData("precision", str(LogPrec))
nt = ss.TestEnv.Table.Len()
sch = etable.Schema(
[etable.Column("Trial", etensor.INT64, go.nil, go.nil),
etable.Column("TrialName", etensor.STRING, go.nil, go.nil),
etable.Column("Cycle", etensor.INT64, go.nil, go.nil)]
)
for lnm in ss.TstRecLays:
ly = leabra.Layer(ss.Net.LayerByName(lnm))
sch.append( etable.Column(lnm, etensor.FLOAT64, ly.Shp.Shp, go.nil))
dt.SetFromSchema(sch, nt)
def ApplyInputs(ss, en):
"""
ApplyInputs applies input patterns from given envirbonment.
It is good practice to have this be a separate method with appropriate
args so that it can be used for various different contexts
(training, testing, etc).
"""
ss.Net.InitExt()
# going to the same layers, but good practice and cheap anyway
lays = ["LGNon", "LGNoff"]
for lnm in lays:
ly = leabra.Layer(ss.Net.LayerByName(lnm))
pats = en.State(ly.Nm)
if pats != 0:
ly.ApplyExt(pats)
def LogTrnEpc(ss, dt):
"""
LogTrnEpc adds data from current epoch to the TrnEpcLog table.
computes epoch averages prior to logging.
# this is triggered by increment so use previous value
"""
row = dt.Rows
dt.SetNumRows(row + 1)
epc = ss.TrainEnv.Epoch.Prv
nt = float(ss.TrainEnv.Table.Len()) # number of trials in view
ss.EpcSSE = ss.SumSSE / nt
ss.SumSSE = 0
ss.EpcAvgSSE = ss.SumAvgSSE / nt
ss.SumAvgSSE = 0
ss.EpcPctErr = float(ss.SumErr) / nt
ss.SumErr = 0
ss.EpcPctCor = 1 - ss.EpcPctErr
ss.EpcCosDiff = ss.SumCosDiff / nt
ss.SumCosDiff = 0
if ss.FirstZero < 0 and ss.EpcPctErr == 0:
ss.FirstZero = epc
if ss.EpcPctErr == 0:
ss.NZero += 1
else:
ss.NZero = 0
dt.SetCellFloat("Run", row, float(ss.TrainEnv.Run.Cur))
dt.SetCellFloat("Epoch", row, float(epc))
dt.SetCellFloat("SSE", row, ss.EpcSSE)
dt.SetCellFloat("AvgSSE", row, ss.EpcAvgSSE)
dt.SetCellFloat("PctErr", row, ss.EpcPctErr)
dt.SetCellFloat("PctCor", row, ss.EpcPctCor)
dt.SetCellFloat("CosDiff", row, ss.EpcCosDiff)
for lnm in ss.LayStatNms:
ly = leabra.Layer(ss.Net.LayerByName(lnm))
dt.SetCellFloat(ly.Nm+" ActAvg", row, float(ly.Pool(0).ActAvg.ActPAvgEff))
# note: essential to use Go version of update when called from another goroutine
ss.TrnEpcPlot.GoUpdate()
if ss.TrnEpcFile != 0:
if ss.TrainEnv.Run.Cur == 0 and epc == 0:
dt.WriteCSVHeaders(ss.TrnEpcFile, etable.Tab)
dt.WriteCSVRow(ss.TrnEpcFile, row, etable.Tab)
def Harmony(ss, nt):
"""
Harmony computes the harmony (excitatory net input Ge * Act)
"""
harm = float(0)
nu = 0
for lyi in nt.Layers:
ly = leabra.Layer(handle=lyi)
if ly.IsOff():
continue
for nrni in ly.Neurons:
nrn = leabra.Neuron(handle=nrni)
harm += nrn.Ge * nrn.Act
nu += 1
if nu > 0:
harm /= float(nu)
return harm
def ApplyInputs(ss, en):
"""
ApplyInputs applies input patterns from given envirbonment.
It is good practice to have this be a separate method with appropriate
# going to the same layers, but good practice and cheap anyway
args so that it can be used for various different contexts
(training, testing, etc).
"""
ss.Net.InitExt()
lays = go.Slice_string(["Name", "Identity", "Color", "FavoriteFood", "Size", "Species", "FavoriteToy"])
for lnm in lays :
ly = leabra.Layer(ss.Net.LayerByName(lnm))
pats = en.State(ly.Nm)
if pats != 0:
ly.ApplyExt(pats)
def TrainTrial(ss):
"""
TrainTrial runs one trial of training using TrainEnv
"""
if ss.NeedsNewRun:
ss.NewRun()
ss.TrainEnv.Step()
# Key to query counters FIRST because current state is in NEXT epoch
# if epoch counter has changed
epc = env.CounterCur(ss.TrainEnv, env.Epoch)
chg = env.CounterChg(ss.TrainEnv, env.Epoch)
if chg:
if ss.ViewOn and ss.TrainUpdt.value > leabra.AlphaCycle:
ss.UpdateView(True)
if epc >= ss.MaxEpcs:
# done with training..
ss.RunEnd()
if ss.TrainEnv.Run.Incr(): # we are done!
ss.StopNow = True
return
else:
ss.NeedsNewRun = True
return
rch = leabra.Layer(ss.Net.LayerByName("Reach"))
if "choice" in ss.TrainEnv.TrialName.Cur:
rch.SetType(emer.Compare)
else:
rch.SetType(emer.Input)
if ss.TrainEnv.GroupName.Cur != ss.PrvGpName: # init at start of new group
ss.Net.InitActs()
ss.PrvGpName = ss.TrainEnv.GroupName.Cur
train = True
if "delay" in ss.TrainEnv.TrialName.Cur:
train = False # don't learn on delay trials
ss.ApplyInputs(ss.TrainEnv)
ss.AlphaCyc(train)
ss.TrialStats(True) # accumulate
ss.LogTrnTrl(ss.TrnTrlLog, ss.TrainEnv.Trial.Cur,
ss.TrainEnv.TrialName.Cur)
def AlphaCyc(ss):
"""
AlphaCyc runs one alpha-cycle (100 msec, 4 quarters) of processing.
External inputs must have already been applied prior to calling,
using ApplyExt method on relevant layers (see TrainTrial, TestTrial).
Handles netview updating within scope of AlphaCycle
"""
if ss.Win != 0:
ss.Win.PollEvents(
) # this is essential for GUI responsiveness while running
viewUpdt = ss.ViewUpdt.value
out = leabra.Layer(ss.Net.LayerByName("Output"))
ss.Net.AlphaCycInit()
ss.Time.AlphaCycStart()
overThresh = False
for qtr in range(4):
for cyc in range(ss.Time.CycPerQtr):
ss.Net.Cycle(ss.Time)
ss.Time.CycleInc()
if viewUpdt == leabra.Cycle:
if cyc != ss.Time.CycPerQtr - 1: # will be updated by quarter
ss.UpdateView()
if viewUpdt == leabra.FastSpike:
if (cyc + 1) % 10 == 0:
ss.UpdateView()
trgact = out.Neurons[1].Act
if trgact > 0.5:
overThresh = True
break
ss.Net.QuarterFinal(ss.Time)
ss.Time.QuarterInc()
if viewUpdt <= leabra.Quarter:
ss.UpdateView()
if viewUpdt == leabra.Phase:
if qtr >= 2:
ss.UpdateView()
if overThresh:
break
ss.UpdateView()
def ConfigTrnTrlLog(ss, dt):
dt.SetMetaData("name", "TrnTrlLog")
dt.SetMetaData("desc", "Record of training per input event (time step)")
dt.SetMetaData("read-only", "true")
dt.SetMetaData("precision", str(LogPrec))
nt = 0
sch = etable.Schema(
[etable.Column("Run", etensor.INT64, go.nil, go.nil),
etable.Column("Epoch", etensor.INT64, go.nil, go.nil),
etable.Column("Trial", etensor.INT64, go.nil, go.nil),
etable.Column("Event", etensor.INT64, go.nil, go.nil),
etable.Column("TD", etensor.FLOAT64, go.nil, go.nil),
etable.Column("RewPred", etensor.FLOAT64, go.nil, go.nil)]
)
for lnm in ss.TstRecLays:
ly = leabra.Layer(ss.Net.LayerByName(lnm))
sch.append( etable.Column(lnm, etensor.FLOAT64, ly.Shp.Shp, go.nil))
dt.SetFromSchema(sch, nt)
def SetParams(ss, sheet, setMsg):
"""
SetParams sets the params for "Base" and then current ParamSet.
If sheet is empty, then it applies all avail sheets (e.g., Network, Sim)
otherwise just the named sheet
if setMsg = true then we output a message for each param that was set.
"""
if sheet == "":
ss.Params.ValidateSheets(go.Slice_string(["Network", "Sim"]))
ss.SetParamsSet("Base", sheet, setMsg)
if ss.ParamSet != "" and ss.ParamSet != "Base":
sps = ss.ParamSet.split()
for ps in sps:
ss.SetParamsSet(ps, sheet, setMsg)
ly = leabra.Layer(ss.Net.LayerByName("NeckerCube"))
ly.Act.Noise.Var = float(ss.Noise)
ly.Act.KNa.On = ss.KNaAdapt
ly.Act.Update()
ss.Time.CycPerQtr = int(ss.CycPerQtr)
def LogTstTrl(ss, dt):
"""
LogTstTrl adds data from current trial to the TstTrlLog table.
log always contains number of testing items
"""
trl = ss.TestEnv.Trial.Cur
row = trl
if dt.Rows <= row:
dt.SetNumRows(row + 1)
dt.SetCellFloat("Trial", row, float(trl))
dt.SetCellString("TrialName", row, ss.TestEnv.TrialName.Cur)
for lnm in ss.TstRecLays:
tsr = ss.ValsTsr(lnm)
ly = leabra.Layer(ss.Net.LayerByName(lnm))
ly.UnitValsTensor(tsr, "Act")
dt.SetCellTensor(lnm, row, tsr)
ss.TstTrlPlot.GoUpdate()
def SetParamsSet(ss, setNm, sheet, setMsg):
"""
SetParamsSet sets the params for given params.Set name.
If sheet is empty, then it applies all avail sheets (e.g., Network, Sim)
otherwise just the named sheet
if setMsg = true then we output a message for each param that was set.
"""
pset = ss.Params.SetByNameTry(setNm)
if sheet == "" or sheet == "Network":
if "Network" in pset.Sheets:
netp = pset.SheetByNameTry("Network")
ss.Net.ApplyParams(netp, setMsg)
hid = leabra.Layer(ss.Net.LayerByName("Hidden"))
fmhid = leabra.Prjn(hid.RcvPrjns.SendName("Hidden"))
fmhid.WtInit.Mean = ss.RecurrentWt
if sheet == "" or sheet == "Sim":
if "Sim" in pset.Sheets:
simp = pset.SheetByNameTry("Sim")
pyparams.ApplyParams(ss, simp, setMsg)
simp.Apply(ss, setMsg)
def LogTstCyc(ss, dt, cyc):
"""
LogTstCyc adds data from current cycle to the TstCycLog table.
log always contains number of testing items
"""
nt = ss.Net()
if dt.Rows <= cyc:
dt.SetNumRows(cyc + 1)
row = cyc
dt.SetCellFloat("Cycle", row, float(cyc))
for lnm in ss.TstRecLays:
ly = leabra.Layer(nt.LayerByName(lnm))
dt.SetCellFloat(lnm + "ActAvg", row,
float(ly.Pool(0).Inhib.Act.Avg))
# note: essential to use Go version of update when called from another goroutine
if cyc % 10 == 0:
ss.TstCycPlot.GoUpdate()
def SetParams(ss, sheet, setMsg):
"""
SetParams sets the params for "Base" and then current ParamSet.
If sheet is empty, then it applies all avail sheets (e.g., Network, Sim)
otherwise just the named sheet
if setMsg = true then we output a message for each param that was set.
"""
if sheet == "":
ss.Params.ValidateSheets(go.Slice_string(["Network", "Sim"]))
ss.SetParamsSet("Base", sheet, setMsg)
if ss.ParamSet != "" and ss.ParamSet != "Base":
sps = ss.ParamSet.split()
for ps in sps:
ss.SetParamsSet(ps, sheet, setMsg)
nt = ss.Net
v1 = leabra.Layer(nt.LayerByName("V1"))
elat = leabra.Prjn(v1.RcvPrjns[2])
elat.WtScale.Rel = ss.ExcitLateralScale
elat.Learn.Learn = ss.ExcitLateralLearn
ilat = leabra.Prjn(v1.RcvPrjns[3])
ilat.WtScale.Abs = ss.InhibLateralScale
def LogTstCyc(ss, dt, cyc):
"""
LogTstCyc adds data from current cycle to the TstCycLog table.
log always contains number of testing items
"""
if dt.Rows <= cyc:
dt.SetNumRows(cyc + 1)
row = cyc
harm = ss.Harmony(ss.Net)
dt.SetCellFloat("Cycle", row, float(cyc))
dt.SetCellString("TrialName", row, ss.TestEnv.TrialName.Cur)
dt.SetCellFloat("Harmony", row, float(harm))
for lnm in ss.TstRecLays:
tsr = ss.ValsTsr(lnm)
ly = leabra.Layer(ss.Net.LayerByName(lnm))
ly.UnitValsTensor(tsr, "Act")
dt.SetCellTensor(lnm, row, tsr)
ss.TstCycPlot.GoUpdate()
def TrialStats(ss, accum):
"""
TrialStats computes the trial-level statistics and adds them to the epoch accumulators if
accum is true. Note that we're accumulating stats here on the Sim side so the
core algorithm side remains as simple as possible, and doesn't need to worry about
different time-scales over which stats could be accumulated etc.
You can also aggregate directly from log data, as is done for testing stats
"""
out = leabra.Layer(ss.Net.LayerByName("Output"))
ss.TrlCosDiff = float(out.CosDiff.Cos)
ss.TrlSSE = out.SSE(0.5) # 0.5 = per-unit tolerance -- right side of .5
ss.TrlAvgSSE = ss.TrlSSE / len(out.Neurons)
if ss.TrlSSE > 0:
ss.TrlErr = 1
else:
ss.TrlErr = 0
if accum:
ss.SumErr += ss.TrlErr
ss.SumSSE += ss.TrlSSE
ss.SumAvgSSE += ss.TrlAvgSSE
ss.SumCosDiff += ss.TrlCosDiff
return
def RunCycles(ss):
"""
RunCycles updates neuron over specified number of cycles
"""
ss.Init()
ss.StopNow = False
ss.Net.InitActs()
ss.SetParams("", False)
ly = leabra.Layer(ss.Net.LayerByName("Neuron"))
nrn = ly.Neurons[0]
inputOn = False
for cyc in range(ss.NCycles):
if ss.Win != 0:
ss.Win.PollEvents(
) # this is essential for GUI responsiveness while running
ss.Cycle = cyc
if cyc == ss.OnCycle:
inputOn = True
if cyc == ss.OffCycle:
inputOn = False
nrn.Noise = float(ly.Act.Noise.Gen(-1))
if inputOn:
nrn.Ge = 1
else:
nrn.Ge = 0
nrn.Ge += nrn.Noise # GeNoise
nrn.Gi = 0
if ss.Spike:
ss.SpikeUpdt(ss.Net, inputOn)
else:
ss.RateUpdt(ss.Net, inputOn)
ss.LogTstCyc(ss.TstCycLog, ss.Cycle)
if ss.Cycle % ss.UpdtInterval == 0:
ss.UpdateView()
if ss.StopNow:
break
ss.UpdateView()
def LogTstCyc(ss, dt, cyc):
"""
LogTstCyc adds data from current cycle to the TstCycLog table.
"""
if dt.Rows <= cyc:
dt.SetNumRows(cyc + 1)
row = cyc
ly = leabra.Layer(ss.Net.LayerByName("Neuron"))
nrn = leabra.Neuron(ly.Neurons[0])
dt.SetCellFloat("Cycle", row, float(cyc))
dt.SetCellFloat("Ge", row, float(nrn.Ge))
dt.SetCellFloat("Inet", row, float(nrn.Inet))
dt.SetCellFloat("Vm", row, float(nrn.Vm))
dt.SetCellFloat("Act", row, float(nrn.Act))
dt.SetCellFloat("Spike", row, float(nrn.Spike))
dt.SetCellFloat("Gk", row, float(nrn.Gk))
dt.SetCellFloat("ISI", row, float(nrn.ISI))
dt.SetCellFloat("AvgISI", row, float(nrn.ISIAvg))
# note: essential to use Go version of update when called from another goroutine
if cyc % ss.UpdtInterval == 0:
ss.TstCycPlot.GoUpdate()
