Reinforcement learning of targeted movement (Chadderdon et al. 2012)

 Download zip file   Auto-launch 
Help downloading and running models
Accession:144538
"Sensorimotor control has traditionally been considered from a control theory perspective, without relation to neurobiology. In contrast, here we utilized a spiking-neuron model of motor cortex and trained it to perform a simple movement task, which consisted of rotating a single-joint “forearm” to a target. Learning was based on a reinforcement mechanism analogous to that of the dopamine system. This provided a global reward or punishment signal in response to decreasing or increasing distance from hand to target, respectively. Output was partially driven by Poisson motor babbling, creating stochastic movements that could then be shaped by learning. The virtual forearm consisted of a single segment rotated around an elbow joint, controlled by flexor and extensor muscles. ..."
Reference:
1 . Chadderdon GL, Neymotin SA, Kerr CC, Lytton WW (2012) Reinforcement learning of targeted movement in a spiking neuronal model of motor cortex. PLoS One 7:e47251 [PubMed]
Citations  Citation Browser
Model Information (Click on a link to find other models with that property)
Model Type: Realistic Network;
Brain Region(s)/Organism: Neocortex;
Cell Type(s): Neocortex fast spiking (FS) interneuron; Neocortex spiking regular (RS) neuron; Neocortex spiking low threshold (LTS) neuron;
Channel(s):
Gap Junctions:
Receptor(s): GabaA; AMPA; NMDA;
Gene(s):
Transmitter(s): Dopamine; Gaba; Glutamate;
Simulation Environment: NEURON;
Model Concept(s): Simplified Models; Synaptic Plasticity; Long-term Synaptic Plasticity; Reinforcement Learning; Reward-modulated STDP;
Implementer(s): Neymotin, Sam [Samuel.Neymotin at nki.rfmh.org]; Chadderdon, George [gchadder3 at gmail.com];
Search NeuronDB for information about:  GabaA; AMPA; NMDA; Dopamine; Gaba; Glutamate;
/
arm1d
README
drspk.mod *
infot.mod *
intf6_.mod *
intfsw.mod *
misc.mod *
nstim.mod *
stats.mod *
updown.mod *
vecst.mod *
arm.hoc
basestdp.hoc
col.hoc *
colors.hoc *
declist.hoc *
decmat.hoc *
decnqs.hoc *
decvec.hoc *
default.hoc *
drline.hoc *
filtutils.hoc *
geom.hoc
grvec.hoc *
hinton.hoc *
infot.hoc *
init.hoc
intfsw.hoc *
labels.hoc *
local.hoc *
misc.h *
mosinit.hoc
network.hoc
nload.hoc
nqs.hoc *
nqsnet.hoc *
nrnoc.hoc *
params.hoc
run.hoc
samutils.hoc *
sense.hoc *
setup.hoc *
sim.hoc
simctrl.hoc *
stats.hoc *
stim.hoc
syncode.hoc *
units.hoc *
xgetargs.hoc *
                            
// $Id: decnqs.hoc,v 1.38 2011/03/01 19:06:15 billl Exp $

print "Loading decnqs.hoc..."

{load_file("nqs.hoc")}
objref nq[10],pq[10]

//** prl2nqs(NQS[,min,max,nointerp]) -- transfer printlist to NQS
proc rename () {}
// eg proc rename () { sprint($s1,"P%d",objnum($s1)) }
obfunc prl2nqs () { local tstep localobj st,oq
  st=new String2()
  oq=new NQS()
  if (numarg()>=1) min=$1 else min=0
  if (numarg()>=2) max=$2 else max=printlist.count-1
  if (numarg()>=3) interp=$3 else interp=0 // no interp when looking at spk times
  if (interp) oq.resize(max-min+2)
  if (interp) {
    tstep=0.1 // 0.1 ms step size for interpolation
    oq.s[0].s="time"
    oq.v[0].indgen(0,printlist.object(0).tvec.max,tstep)
    for ii=min,max {
      XO=printlist.object(ii)
      oq.s[ii+1-min].s = XO.var
      rename(oq.s[ii+1-min].s)
      oq.v[ii+1-min].resize(oq.v.size)
      oq.v[ii+1-min].interpolate(oq.v[0],XO.tvec,XO.vec)
    }
  } else {
    for ii=min,max {
      XO=printlist.object(ii)
      st.s=XO.name
      sprint(st.t,"%s-time",XO.name)
      rename(st.s) rename(st.t)
      oq.resize(st.s,st.t)
      oq.setcols(XO.vec,XO.tvec)
    }
  }
  return oq
}

//** pvp2nqs(NQS) -- transfer grvec data file to NQS
obfunc pvp2nqs () { local min,max,interp,gvnum,ii,jj,n localobj oq,po,st,xo
  interp=min=max=gvnum=0
  if (argtype(1)==2) { po=gvnew($s1) gvnum=panobjl.count()-1 }
  if (argtype(1)==0) gvnum=$1
  if (numarg()>=2) min=$2
  if (numarg()>=3) max=$3
  if (numarg()>=4) interp=$4 // no interp when looking at spk times
  if (po==nil) po=panobjl.o(gvnum)
  oq=new NQS() st=new String2()
  if (gvnum>0) {
    if (max==0) max=po.llist.count()-1
    for ii=min,max {
      xo=po.llist.object(ii)
      po.tmpfile.seek(xo.loc)
      if (xo.num==-2) {
        sprint(st.s,"%s-time",xo.name)
        jj=oq.resize(st.s)-1
        oq.v[jj].vread(po.tmpfile)
      }
      jj=oq.resize(xo.name)-1
      oq.v[jj].vread(po.tmpfile)
    }
  } else { // from printlist
    if (max==0) max=printlist.count()-1
    for ii=min,max {
      xo=printlist.o(ii)
      jj=oq.resize(xo.name)-1
      oq.v[jj].copy(xo.vec)
      if (xo.pstep==0) {
        sprint(st.s,"%s-time",xo.name)
        jj=oq.resize(st.s)-1
        oq.v[jj].copy(xo.tvec)
      }
    }
  }
  return oq
}

//** veclist2nqs(nqs[,STR1,STR2,...])
proc veclist2nqs () { local flag,i
  if (numarg()==0) {printf("veclist2nqs(nqs[,STR1,STR2,...])\n") return}
  $o1.resize(veclist.count)
  if (numarg()==1+$o1.m) flag=1 else flag=0
  for ltr(XO,veclist) {
    $o1.v[i1].copy(XO)
    if (flag) {i=i1+2 $o1.s[i1].s=$si} else {sprint(tstr,"v%d",i1) $o1.s[i1].s=tstr}
  }
  $o1.cpout()
}

// fudup(vec[,nq,#CUTS,LOGCUT,MIN]) -- use updown() to find spikes
// LOC(0) PEAK(1) WIDTH(2) BASE(3) HEIGHT(4) START(5) SLICES(6) SHARP(7) INDEX(8) FILE(9) NESTED(10)
// other options
pos_fudup=1 // set to 1 to move whole curve up above 0
maxp_fudup=0.95 // draw top sample at 95% of max 
minp_fudup=0.05 // draw bottom sample at 5% of max
over_fudup=1    // turn over and try again if nothing found
allover_fudup=0 // turn over and add these locs (not debugged)
verbose_fudup=0 // give messages, can also turn on DEBUG_VECST for messages from updown()
obfunc fudup () { local a,i,ii,npts,logflag,min,x,sz localobj bq,cq,v1,v2,v3,bb,tl,v5,eq
  if (verbose_fudup) printf("MAXTIME appears to be %g (dt=%g)\n",$o1.size*dt,dt)
  logflag=0  npts=10 // n sample locations by default
  bq=new NQS("LOC","PEAK","WIDTH","BASE","HEIGHT","START","SLICES","SHARP","INDEX","FILE","NESTED") 
  i=2
  if (argtype(i)==1) {i+=1 if ($o2==nil) {cq=new NQS() $o2=cq} else cq=$o2} else cq=new NQS()
  if (cq.m!=11) { cq.resize(0) 
  cq.resize("LOC","PEAK","WIDTH","BASE","HEIGHT","START","SLICES","SHARP","INDEX","FILE","NESTED")}
  if (argtype(i)==0){ npts=$i i+=1
    if (argtype(i)==0){ logflag=$i i+=1
      if (argtype(i)==1) { v5=$oi i+=1
        if (npts!=v5.size) printf("Correcting npts from %d to %d\n",npts,npts=v5.size)
        if (v5.ismono(1)) v5.reverse
        if (! v5.ismono(-1)) {printf("fudup: final arg (%s) must be monotonic\n",v5) return}
      }
    }
  }
  bq.listvecs(bb)
  bq.pad(5000)
  eq=new NQS(-2,npts) a=allocvecs(v1,v2,v3)
  tl=eq.vl
  eq.clear(2e4) vrsz(2e4,v1,v2,v3)
  v1.copy($o1)
  if (pos_fudup) {
    min=v1.min
    v1.sub(min) // make it uniformly positive
  } else min=0
  if (numarg()>4) v2.copy(v5) else {
    v2.indgen(2,2+npts-1,1)   // sampling at npts points, start at 2 to avoid log(1)=0
    if (logflag) v2.log() // log sampling
    v2.scale(-maxp_fudup*v1.max,-minp_fudup*v1.max) v2.mul(-1)
  }
  v1.updown(v2,tl,bb)
  if (pos_fudup) { bq.v[1].add(min) bq.v[3].add(min) }
  cq.append(bq)
  sz=bq.size(1)
  if (allover_fudup) { // do it upside down as well
    v1.mul(-1) // v2 will be upside-down
    if (pos_fudup) {min=v1.min v1.sub(min)}
    if (0) {  // can't see a rationale to recalc sampling points
      v2.indgen(2,2+npts-1,1)   // sampling at npts points
      if (logflag) v2.log() // log sampling
      v2.scale(-0.95*v1.max,-0.05*v1.max) v2.mul(-1)
    }
    v1.updown(v2,tl,bb)
    bq.v[8].add(sz) bq.v[4].mul(-1) // turn HEIGHT upside down
    cq.append(bq)
  } else if (over_fudup && sz==0) { // turn it over an try again
    print "fudup() checking upside-down"
    v1.mul(-1) // v2 will be upside-down
    v1.updown(v2,tl,bb)
  } 
  for case(&x,0,2,5) cq.v[x].mul(dt)
  nqsdel(bq,eq)
  dealloc(a)
  return cq
}

//** listsort(LIST[,START,REV]) sorts list of strings numerically
// optional start gives a regexp to start at
proc listsort () { local x,rev localobj nq,st,xo
  if (numarg()==0) { 
    print "listsort(LIST[,RXP,REV]) numerically, optional RXP starts after there" return}
  if (numarg()==3) if ($3) rev=-1 else rev=0
  nq=new NQS("STR","NUM") nq.strdec("STR")
  st=new String()
  for ltr(xo,$o1) {
    if (numarg()>=2) sfunc.tail(xo.s,$s2,st.s) else st.s=xo.s
    if (sscanf(st.s,"%g",&x)!=1) print "listsort ERR: num not found in ",st.s
    nq.append(xo.s,x)
  } 
  nq.sort("NUM",rev)
  $o1.remove_all
  for nq.qt(st.s,"STR") $o1.append(new String(st.s))
}

// stat(VEC) print stats for the vector
// stat(VEC1,VEC2) append stats of VEC1 on VEC2
// stat(VEC1,NQS) append stats of VEC1 on NQS (create if necessary)
proc stat () { local sz
  sz=$o1.size
  if (sz<=1) {printf("decnqs::stat() WARN: %s size %d\n",$o1,$o1.size) return}
  if (numarg()==1 && sz>1) {
    printf("Sz:%d\tmax=%g; min=%g; mean=%g; stdev=%g\n",$o1.size,$o1.max,$o1.min,$o1.mean,$o1.stdev)
  } else {
    if (!isassigned($o2)) { $o2=new NQS("SIZE","MAX","MIN","MEAN","STDEV")
    } else if (isobj($o2,"NQS")) { 
      if ($o2.m!=5) {$o2.resize(0) $o2.resize("SIZE","MAX","MIN","MEAN","STDEV")}
    } else if (isobj($o2,"Vector")) revec($o2)
    if (sz>2) {$o2.append($o1.size,$o1.max,$o1.min,$o1.mean,$o1.stdev) // .append for Vector or NQS
    } else   $o2.append($o1.size,$o1.max,$o1.min,$o1.min,0) // no sdev
  }
}

//* fil2nqs(FILE,NQS) reads lines of file and places all numbers in NQS
func fil2nqs () { local a,n localobj v1
  $o2.clear
  a=allocvecs(v1)
  tmpfile.ropen($s1)
  for (n=1;tmpfile.gets(tstr)!=-1;n+=1) {
    if (n%1e3==0) printf("%d ",n)
    parsenums(tstr,v1)
    if (v1.size!=$o2.m) {
      printf("Wrong size at line %d (%d)  ",n,v1.size)  vlk(v1)
      return
    }
    $o2.append(v1)
  }
  dealloc(a)
  return $o2.size(1)
}

//** plnqs(file,NQS) reads output of txt2num.pl
// format ascii 'rows cols' then binary contents
proc plnqs () { local a,rows,cols localobj v1,v2
  a=allocvecs(v1,v2)
  tmpfile.ropen($s1)
  tmpfile.gets(tstr)
  sscanf(tstr,"%d %d",&rows,&cols)
  printf("%s: %d rows x %d cols\n",$s1,rows,cols)
  v1.fread(tmpfile,rows*cols) // could now use .transpose
  v2.indgen(0,rows*cols,cols)
  $o2.resize(cols,rows)
  for ii=0,cols-1 {
    v2.add(ii)
    $o2.v[ii].index(v1,v2)
  }
  dealloc(a)
}
    
// DEST=maxem(SRC,MIN,WIDTH) -- keep looking for maxima till get down to min
obfunc maxem () { local a,min,wid,ii,ix,beg,end localobj v1,aq
  a=allocvecs(v1)
  aq=new NQS("max","loc") aq.clear(v1.size/2)
  v1.copy($o1)
  min=$2 wid=$3
  while(v1.max>min) {
   aq.append(v1.max,ix=v1.max_ind)
   beg=ix-wid if (beg<0) beg=0
   end=ix+wid if (end>=v1.size) end=v1.size-1
   for ii=beg,end v1.x[ii]=-1e9
  }
  dealloc(a)
  return aq
}

// nqo=percl(nq,"COLA", ..) generates NQS of percentile values (10..90) for these cols
// nqo=percl(nq,min,max,step,"COLA", ..) -- eg percl(nq,50,70,5,"COLA","COLB")
obfunc percl () {  local i,ii,a,p localobj v1,v2,v3,aq,xo
  a=allocvecs(v1,v2,v3)
  aq=new NQS(numarg())
  if (argtype(2)==0) {
    v3.indgen($2,$3,$4) 
    aq.resize(aq.size(1)-3)
    i=5 j=4 // start at arg i and aq col #j
  } else {
    v3.indgen(10,90,10)
    i=2 j=1
  }
  aq.setcol(0,"PERCL",v3)
  for (;i<=numarg();i+=1) {
    $o1.getcol($si,v1)
    v1.sort()
    v2.resize(0)
    for vtr(&ii,v3) {ii/=100 v2.append(v1.x[round(ii*v1.size)])}
    aq.setcol(i-j,$si,v2)
  }
  dealloc(a)
  return aq
}

// pqunq(NQS) returns columns of sorted nique values corresponding to the arg
// NB: does not produce a rectangular array
obfunc pqunq () { local a localobj v1,v2,aq
  aq=new NQS()
  a=allocvecs(v1,v2,1e5)
  aq.sethdrs($o1)
  aq.resize(-2) 
  for ii=0,aq.m-1 {
    $o1.getcol(ii,v1)
    v1.sort 
    v2.redundout(v1)
    aq.v[ii].copy(v2)
  }
  dealloc(a)
  return aq
}

//** aa=seqind(ind) -- find beginning and end of sequential indices with 
obfunc seqind () { local a,n,skip,ii,x,last localobj vi,oq
  vi=$o1
  if (numarg()>=2) skip=$2+1 else skip=1
  if (numarg()>=3) oq=$o3
  if (!isassigned(oq)) {oq=new NQS() if (numarg()>=3) $o3=oq}
  if (oq.m!=3) { oq.resize(0) oq.resize("beg","end","diff") }
  oq.clear()
  n=last=0
  for ii=1,vi.size(1)-1 {
    if (vi.x[ii]-vi.x[ii-1]>skip) {
      if (n>0) oq.append(vi.x[last],vi.x[ii-1],0)
      last=ii
      n=0
    } else n+=1
  }
  if (n>0) oq.append(vi.x[last],vi.x[ii-1],0)
  oq.pad()
  oq.calc("<diff>.copy(<end>.c.sub(<beg>))")
  return oq
}

//** list_transpose
proc list_transpose () { localobj aq,mat,xo,inlist,outlist
  aq=new NQS() inlist=$o1 outlist=$o2
  if (!isojt(outlist,inlist)) {outlist=new List() $o2=outlist}
  for ltr(xo,inlist) aq.resize("",xo)
  mat=aq.tomat(1) // transpose
  aq.frmat(mat)
  outlist.remove_all
  aq.listvecs(outlist)
  delnqs(aq)
}

//* Sam's additions -- moved from nqs_utils.hoc
//get row of Vectors
//$o1 = nqs
//$2 = row number
//$s3 - $snumarg() - name of cols to get values for
//returns list with associated Vectors
obfunc getobjrow(){ local i,rowid localobj nq,vt,ls
  nq=$o1 rowid=$2
  ls=new List()
  vt=new Vector()
  for(i=3;i<=numarg();i+=1){
    nq.get($si,rowid,vt)
    ls.append(vt)
  }
  return ls
}

//get column of objects as Vector using oform
//$o1 = nqs
//$s2 = col name
//$3=iff==1 return list of Vectors in column, else return vector of oform of each row in column
obfunc getobjcol(){ local idx,getl localobj nq,vt,vt2,ls
  nq=$o1
  if(numarg()>2) getl=$3 else getl=0
  if(getl){
    ls=new List()
    vt=new Vector()
    for idx=0,nq.size-1{
      nq.get($s2,idx,vt)
      ls.append(vt)
    }
    return ls
  } else {
    vt=new Vector(nq.size)
    vt.resize(0)
    vt2=new Vector()
    for idx=0,nq.size-1{
      nq.get($s2,idx,vt2)
      vt.append(oform(vt2))
    }
    return vt
  }
}

//get correlation between 2 columns of an NQS
//$o1=nqs
//$s2=column name
//$s3=column name
//$4=pearson correlation iff == 1 (default), otherwise spearman
func nqcor(){ local pc localobj nq1,v1,v2
  if(numarg()>3) pc=$4 else pc=1
  nq1=$o1
  v1=nq1.getcol($s2)  v2=nq1.getcol($s3)
  if (pc) return v1.pcorrel(v2) else return v1.scorrel(v2)
}

//func nqgrslice(){ local startidx,endidx localobj nq,vtmp
//  nq=$o1 startidx=$2 endidx=$3
//  vtmp=new Vector($3-$2+1)
//  gg(
//}

func MIN(){ if($1<$2)return $1 else return $2 }

//get correlation matrix/nqs of all columns to all columns
//$o1 = NQS
//$2 = num columns 0 - NQS.m
//$3 = start row/index
//$4 = end row/index
//$5 = index increment
//$6 = window size , iff <= 0, do full columns against each other
obfunc nqcolcor(){ local startidx,endidx,inct,wint,c1,c2,ncol localobj vhr,vhl,nqc,nqf
 if(numarg()<1){
   printf("nqcolcor usage: \n\t$o1 = NQS\
                           \n\t$2 = num columns 0 - NQS.m\
                           \n\t$3 = start row/index\
                           \n\t$4 = end row/index\
                           \n\t$5 = index increment\
                           \n\t$6 = window size , iff <= 0, do full columns against each other\n")
   return nil
 }
 nqf=$o1
 if(numarg()>1) ncol=$2 else ncol=nqf.m
 if(numarg()>2) startidx=$3 else startidx=0
 if(numarg()>3) endidx=$4 else endidx=nqf.size
 if(numarg()>4) inct=$5 else inct=50*2//50ms
 if(numarg()>5) wint=$6 else wint=100*2//50ms

 vhr=new Vector(wint) vhl=new Vector(wint)

 if(wint<=0){ //full column cross-correlation
   nqc=new NQS("ID0","ID1","cor") 
   for c1=0,ncol-1{
     for c2=c1+1,ncol-1{
       nqc.append(c1,c2,nqf.v[c1].pcorrel(nqf.v[c2]))
     }
   }
 } else { //cross correlation using slices of column
   nqc=new NQS("ID0","ID1","start","end","cor") 
   for c1=0,ncol-1{
     for c2=c1+1,ncol-1{
       for(startidx=0;startidx<endidx;startidx+=inct){
         vhl.copy(nqf.v[c1],startidx,MIN(startidx+wint,endidx-1))
         vhr.copy(nqf.v[c2],startidx,MIN(startidx+wint,endidx-1))
         nqc.append(c1,c2,startidx,startidx+wint,vhl.pcorrel(vhr))
       }
     }
   }
 }

 return nqc
}

 
//read wmf ascii file (just skips header and calls rdcol)
//$s1 = wmf file path
//$2 = # of columns
// obfunc rdwmf(){ local idx,jdx,hdrlines localobj nq,myf,myftmp,strf,str,strtmp,lcols
// myf=new File() myftmp=new File() strf=new StringFunctions() str=new String() lcols=new List()
// strtmp=new String() hdrlines=6
// myf.ropen($s1)
// if(!myf.is_open()){
//  printf("rdwmf ERRA: couldn't open wmf file %s for read\n",$s1)
//  return nil
// }
// for idx=0,hdrlines-1{
//  if(myf.gets(str.s)==-1){
//    printf("rdwmf ERRB: corrupt header\n")
//    return nil
//  } else if(idx==2){
//    jdx=strf.tail(str.s,"",strtmp.s)            
//  }
// }
// myf.close()
// return nq
// }

//draw regression line
//$o1 = nqs, $s2 = column 1, $s3 = column 2
// or
//$o1 = Vector 1 , $o2 = Vector 2
//returns vo
obfunc drawregline(){ local x0,y0,x1,y1,gvtmp,r localobj vo,nq,v1,v2,vx,vy,str
  vo=new Vector(5)
  if(numarg()==3){
    nq = $o1
    v1=new Vector()  v2=new Vector()
    nq.getcol($s2,v1)
    nq.getcol($s3,v2)
  } else {
    v1=$o1 v2=$o2
  }
  v1.vstats(v2,vo)
  x0 = v2.min
  y0 = x0*vo.x(0)+vo.x(1)
  x1 = v2.max
  y1 = x1*vo.x(0)+vo.x(1)
  vx=new Vector(2)
  vx.x(0)=x0
  vx.x(1)=x1
  vy=new Vector(2)
  vy.x(0)=y0
  vy.x(1)=y1
  gvtmp=gvmarkflag
  gvmarkflag=0
  gg(vy,vx)
  gvmarkflag=gvtmp
  str=new String()
  r=v1.pcorrel(v2)
  if(name_declared("rpval_stats")){
    sprint(str.s,"r = %.2f, p = %g, N = %d",r,rpval_stats(v1.size,r),v1.size)
    g.label(0,0,str.s)
  }
  return vo
}

// select from a vector handled as a matrix -- see matrix.mod
halfmat=0
obfunc mindsel () { local a,x,r,c localobj vm,vi,oq
  vm=$o1 
  a=allocvecs(vi)
  if (numarg()==4) vi.indvwhere(vm,$s2,$3,$4) else vi.indvwhere(vm,$s2,$3)
  oq=new NQS("row","col","val") 
  oq.clear(vi.size)
  for vtr(&x,vi) {
    r=int(x/COLS) c=x-r*COLS
    if (!halfmat || c>r) oq.append(r,c,vm.x[x])
  }
  dealloc(a)
  print oq.size(1)
  return oq
}

//* return row $2 of nqs $o1 
obfunc nqrow () { local row,col localobj vout,nq
  nq=$o1
  row=$2
  vout=new Vector(nq.m)
  for col=0,nq.m-1 vout.x(col)=nq.v[col].x(row)
  return vout
}

//* find row $o2 (vector) in $o1 (nqs) and return index
//  if not there return -1
func nqfindrow () { local idx,jdx,sz localobj nq,vf,vrow
  nq=$o1  vf=$o2
  sz=nq.size(-1)
  for idx=0,sz-1 {
    vrow = nqrow(nq,idx)
    if(vrow.eq(vf)) return idx
  }
  return -1
}

//* nquniq(NQS) -- return a new NQS with unique rows in $o1
obfunc nquniq () { local sz,idx,outrow,jdx localobj nqin,nqout,vrow
  nqin=$o1
  nqout=new NQS()
  for idx=0,nqin.m-1{
    nqout.resize(nqin.s[idx].s)
    nqout.v[nqout.m-1].resize(nqin.v[idx].size)
  }
  sz=nqin.size(-1)
  jdx=0
  outrow=0
  for idx=0,sz-1{
    vrow=nqrow(nqin,idx)
    if(nqfindrow(nqout,vrow)==-1){
      for jdx=0,nqin.m-1 nqout.v[jdx].x(outrow) = vrow.x(jdx)
      outrow += 1
    }
  }
  for idx=0,nqout.m-1 nqout.v[idx].resize(outrow)
  return nqout
}