Network *PerceptronModel::createNetwork(Episode *first_episode) const
{
Network *network = new Network(first_episode->encodedStateSize());
Dense *hidden = new Dense(_hidden_neurons, 1e-2);
TanhActivation *hidden_activation = new TanhActivation;
Dense *dense2 = new Dense(first_episode->valueSize(), 1e-2);
hidden->setInput(network->inputPort());
hidden_activation->setInput(hidden->output());
dense2->setInput(hidden_activation->output());
network->addNode(hidden);
network->addNode(hidden_activation);
network->addNode(dense2);
return network;
}
void route_edges(Layouter &state,plugin& pg, double scale, int iter, double temp, int debug){
int n=state.nw.nodes.size();
int i,j,k;
vector<NodeLinePair> gs;
Rect bb=state.nw.getBB();
bb.extend(state.avgsize/10);
Network nv;
vector<VI> nodemap(n);
vector<vector<VI> > edge_crossings(n+4,vector<VI>(n+4)); // this stores for each voronoi line between node i and j the crosspoints; includes 4 virtual nodes
printf("finding Voronoi separator lines\n");
for (i=0;i<n;i++){
gs.clear();
double x=state.nw.nodes[i].x;
double y=state.nw.nodes[i].y;
double w=state.nw.nodes[i].width;
double h=state.nw.nodes[i].height;
Rect ri=state.nw.nodes[i].rect();
//if (i==iter) debugrect(ri,0,0,255);
gs.push_back(NodeLinePair(n,ParamEdge(Point(x,bb.ymin),bb.TL()))); // boundary lines (should point to left); adds virtual nodes n ... n+3
gs.push_back(NodeLinePair(n+1,ParamEdge(Point(bb.xmin,y),bb.BL())));
gs.push_back(NodeLinePair(n+2,ParamEdge(Point(x,bb.ymax),bb.BR())));
gs.push_back(NodeLinePair(n+3,ParamEdge(Point(bb.xmax,y),bb.TR())));
double dclosest=DBL_MAX;
int iclosest;
for (j=0;j<n;j++){
if (i==j) continue;
double x2=state.nw.nodes[j].x;
double y2=state.nw.nodes[j].y;
double w2=state.nw.nodes[j].width;
double h2=state.nw.nodes[j].height;
double dx=x2-x;
double dy=y2-y;
double mx=(dx>0 ? (x+w/2+x2-w2/2)/2 : (x-w/2+x2+w2/2)/2); // point thru which separation line should go (in the middle between the two nodes)
double my=(dy>0 ? (y+h/2+y2-h2/2)/2 : (y-h/2+y2+h2/2)/2);
if (my<min(y,y2)) my=min(y,y2); // if nodes "overlap" in one direction, these limits need to be applied
if (mx<min(x,x2)) mx=min(x,x2);
if (my>max(y,y2)) my=max(y,y2);
if (mx>max(x,x2)) mx=max(x,x2);
Point m(mx,my);
double gx=dy; // line sould point perpendicular to distance vector; points to left
double gy=-dx;
//if (i==iter) debugline(m.x,m.y,m.x+gx,m.y+gy,200,100,100,true);
// if (i==iter) debugline(x,y,x2,y2,100,100,200,true);
double ord_angles[4];
double &alpha=ord_angles[0];
alpha=angle(Point(gx,gy));
// find suitable 0, 45 and 90° lines and sort them by proximity to (gx,gy) line
double &nearest=ord_angles[1]; // just references to iterate through candidates later on
double &second=ord_angles[2];
double &third=ord_angles[3];
nearest=PI/4*round(4*alpha/PI);
double left=PI/2*floor(2*alpha/PI);
double right=PI/2*ceil(2*alpha/PI);
// find second nearest
second=PI/4*floor(4*alpha/PI);
if (second==nearest) second=PI/4*ceil(4*alpha/PI);
// find third nearest
if (nearest==left){
third=right;
} else if (nearest==right){
third=left;
} else if (second==left){
third=right;
} else {
third=left;
}
/* for (k=0;k<3;k++){
if (i==1) debugline(m.x,m.y,m.x+100*Point(ord_angles[k]).x,m.y+100*Point(ord_angles[k]).y,100,100,200,true);
}*/
// check candidates in the defined order whether they collide with one of the two nodes (i or j)
Rect rj=state.nw.nodes[j].rect();
Point g(0,0);
bool found=false;
for (k=0;k<4;k++){
g=Point(ord_angles[k]);
if (prod(g,rj.TL()-m)>0 && prod(g,rj.TR()-m)>0 && prod(g,rj.BR()-m)>0 && prod(g,rj.BL()-m)>0 &&
prod(g,ri.TL()-m)<0 && prod(g,ri.TR()-m)<0 && prod(g,ri.BR()-m)<0 && prod(g,ri.BL()-m)<0) { // node j completely right of g && i completely left
found=true;
break;
}
}
if (!found) g=Point(ord_angles[0]); // ups? do nodes overlap?
// generate the line and save it
ParamEdge ge(m,m+g); // the voronoi line separating node i from node j
Point pclosest=ge.dist_vec(state.nw.nodes[i]); // closest point on line to node i
ge.re_ref(state.nw.nodes[i]+pclosest); // reference point of line needs to be the closest point to node i
gs.push_back(NodeLinePair(j,ge));
if (norm(pclosest)<dclosest){ // find closest line to node i on the fly
dclosest=norm(pclosest);
iclosest=gs.size()-1;
}
//if (i==iter) debugline(ge.from(),ge.p(state.avgsize*4),100,100,100,true);
}
//find set of voronoi lines which minimally surround node i
int curidx=iclosest;
int minidx;
int lastidx=-1;
bool first=true;
int gl=gs.size();
VI used(gl,0);
while (first || curidx!=iclosest){
ParamEdge &cur=gs[curidx].line;
double minc=DBL_MAX;
for (j=0;j<gl;j++){ // find the nearest crosspoint of one of all voronoi lines; searching to the left from last cross point
if (j==curidx) continue;
if (used[j]) continue;
if (j==lastidx) continue; // finds the last voronoi line again; should not happen as this line should be in used[] already; except for iclosest
double c=cur.cross_param(gs[j].line);
if (c==0) { // special case, check whether new line points inwards
Point ctr90=to_left(Point(x,y)-cur.p(c),PI);
Point cln90=to_left(cur.unit(),PI);
Point dirj=gs[j].line.unit();
if (!((scalar(ctr90,dirj)<0) && (scalar(cln90,dirj)>0))){ // new line not between center line and current line
continue;
}
}
if (c>=0 && c<minc){
minc=c;
minidx=j;
//if (i==iter) debugpoint(cur.p(c),state.avgsize/10,0,0,255);
}
}
used[minidx]=1;
Point cp=cur.cross_point(gs[minidx].line);
//if (i==iter) debugpoint(cp,state.avgsize/10,0,255,0);
gs[minidx].line.re_ref(cp); // setting ref point of next line to the current cross point
//if (i==iter) debugline(gs[minidx].line.from(),gs[minidx].line.p(state.avgsize),255,0,255);
int newnode=nv.nodes.size();
nv.addNode(newnode,other,"",1,1,cp.x,cp.y,0);
edge_crossings[min(i,gs[curidx].node)][max(i,gs[curidx].node)].push_back(newnode); // add the new node to the two veronoi lines it belongs to
edge_crossings[min(i,gs[minidx].node)][max(i,gs[minidx].node)].push_back(newnode);
//nodemap[i].push_back(newnode);
lastidx=curidx;
curidx=minidx;
first=false;
//if (i==iter) debugline(cur.from(),cp,0,0,0);
}
}
printf("finding relevant cross points on seperator lines and building network\n");
// go through all voronoi lines and connect all registered cross points
CmpCrossPoints ccp(nv);
for (i=0;i<n+4;i++){ // includes 4 virtual nodes
for (j=i+1;j<n+4;j++){
if (edge_crossings[i][j].size() ==0 ) continue;
sort(edge_crossings[i][j].begin(),edge_crossings[i][j].end(),ccp); // sort cross points on line
for (k=0;k<(int)edge_crossings[i][j].size();k++){
int n1=edge_crossings[i][j][k];
if (i<n) nodemap[i].push_back(n1); // register voronoi node to be adjacent to original node i (if not virtual node)
if (j<n) nodemap[j].push_back(n1);
if (k==((int)edge_crossings[i][j].size())-1) break; // for the last crosspoint we do not create an edge
int n2=edge_crossings[i][j][k+1];
nv.addEdge(n1,n2,undirected);
if (iter<=1) debugline(nv.nodes[n1].x,nv.nodes[n1].y,nv.nodes[n2].x,nv.nodes[n2].y,0,0,0,true);
}
}
}
printf("finding shortest path through network for each original edge\n");
if (iter>=1){ // this is just for showing 1 step show only veronoi lines, 2nd show splines
nv.calcEdgeLengths();
int m=state.nw.edges.size();
for (i=0;i<m;i++){
printf(".");fflush(stdout);
Edge &e=state.nw.edges[i];
int n1=e.from;
int n2=e.to;
BFS bfs(nv,nodemap[n1]);
int nn=bfs.next();
while (nn>=0 && find(nodemap[n2].begin(),nodemap[n2].end(),nn)==nodemap[n2].end()){
nn=bfs.next();
}
if (nn<0) printf("Ups, no route for edge\n");
VI path=bfs.path();
smooth_path(nv,path,state.avgsize/4);
e.splinepoints.clear();
e.splinehandles.clear();
double beta;
Point vec;
double dd1=(path.size()>1 ?
(norm(nv.nodes[path[0]]-state.nw.nodes[n1])-max(state.nw.nodes[n1].width,state.nw.nodes[n1].height)/2)/2 :
(norm(state.nw.nodes[n2]-state.nw.nodes[n1])-max(state.nw.nodes[n1].width,state.nw.nodes[n1].height)/2-max(state.nw.nodes[n2].width,state.nw.nodes[n2].height)/2)/2);
dd1=min(dd1,state.avgsize/2);
double d1=max(state.nw.nodes[n1].width,state.nw.nodes[n1].height)/2+dd1;
double dd2=(path.size()>1 ?
(norm(nv.nodes[path.back()]-state.nw.nodes[n2])-max(state.nw.nodes[n2].width,state.nw.nodes[n2].height)/2)/2 :
(norm(state.nw.nodes[n2]-state.nw.nodes[n1])-max(state.nw.nodes[n1].width,state.nw.nodes[n1].height)/2-max(state.nw.nodes[n2].width,state.nw.nodes[n2].height)/2)/2);
dd2=min(dd2,state.avgsize/2);
double d2=max(state.nw.nodes[n2].width,state.nw.nodes[n2].height)/2+state.avgsize/2;
Point sp1,sp2;
switch(e.type){
case substrate:
reverse(path.begin(),path.end());
swap(n1,n2);
swap(d1,d2);
sp1=(path.size()>1 ? nv.nodes[path[0]] : state.nw.nodes[n2]);
e.splinehandles.push_back(unit(sp1-state.nw.nodes[n1])*d1);
e.splinehandles.push_back(Point(state.nw.nodes[n2].dir+PI/2)*d2);
break;
case product:
e.splinehandles.push_back(Point(state.nw.nodes[n1].dir-PI/2)*d1);
sp2=(path.size()>1 ? nv.nodes[path.back()] : state.nw.nodes[n1]);
e.splinehandles.push_back(unit(sp2-state.nw.nodes[n2])*d2);
break;
case activator:
case inhibitor:
case catalyst:
reverse(path.begin(),path.end());
swap(n1,n2);
swap(d1,d2);
vec=state.nw.nodes[n1]-state.nw.nodes[n2]; // vector pointing from n2 (reaction) to n1 (catalyst,etc)
beta=0;
if (scalar(vec,Point(state.nw.nodes[n2].dir+PI))>scalar(vec,Point(state.nw.nodes[n2].dir))) beta=PI;
sp1=(path.size()>1 ? nv.nodes[path[0]] : state.nw.nodes[n2]);
e.splinehandles.push_back(unit(sp1-state.nw.nodes[n1])*d1);
e.splinehandles.push_back(Point(state.nw.nodes[n2].dir+beta)*d2);
break;
default:
sp1=(path.size()>1 ? nv.nodes[path[0]] : state.nw.nodes[n2]);
e.splinehandles.push_back(unit(sp1-state.nw.nodes[n1])*d1);
sp2=(path.size()>1 ? nv.nodes[path.back()] : state.nw.nodes[n1]);
e.splinehandles.push_back(unit(sp2-state.nw.nodes[n2])*d2);
}
if (path.size()>1){
//debugline(state.nw.nodes[n1].x,state.nw.nodes[n1].y,nv.nodes[path.front()].x,nv.nodes[path.front()].y,255,100,100);
for (j=0;j<(int)path.size();j++){
Point before=(j==0 ? state.nw.nodes[n1] : nv.nodes[path[j-1]]);
Point &after=(j==((int)path.size())-1 ? state.nw.nodes[n2] : nv.nodes[path[j+1]]);
Point &cur=nv.nodes[path[j]];
if (after==cur) continue;
if (before==cur) {
if (j-1<0) continue;
before=(j-1==0 ? state.nw.nodes[n1] : nv.nodes[path[j-2]]);
}
Point d=unit(before-cur)+unit(after-cur);
if (d.is_null()) continue;
d=unit(d)*min(min(norm(before-cur),norm(after-cur))/2,state.avgsize/2);
e.splinehandles.insert(--e.splinehandles.end(),to_left(d,PI/2)*sign(scalar(to_left(d,PI/2),before-cur)));
e.splinepoints.push_back(cur+d);
//debugline(cur+d,cur+d+to_left(d,PI/2)*sign(scalar(to_left(d,PI/2),before-cur)),0,0,255,true);
//if (j<path.size()-1) debugline(nv.nodes[path[j]].x,nv.nodes[path[j]].y,nv.nodes[path[j+1]].x,nv.nodes[path[j+1]].y,255,100,100);
}
//debugline(nv.nodes[path.back()].x,nv.nodes[path.back()].y,state.nw.nodes[n2].x,state.nw.nodes[n2].y,255,100,100);
}
}
}
/* NetDisplay nd(nv);
nd.waitKeyPress=true;
nd.show();*/
}
int main (int argc, char** argv) {
//srand(time(NULL));
//manual config...
//------------- output layer
OutputNode output;
output._name = "f";
Sigmoid sigmoid;
output._activationFunc = &sigmoid;
SquaredError squaredError;
output._criterion = &squaredError;
//------------- interal layer 1
InternalNode internal_1, internal_2;
internal_1._name = "i1";
internal_2._name = "i2";
Sigmoid sigmoid_i1, sigmoid_i2;
internal_1._activationFunc = &sigmoid_i1;
internal_2._activationFunc = &sigmoid_i2;
//------------- input layer
InputNode input_1, input_2, input_3;
input_1._name = "a";
input_2._name = "b";
input_3._name = "c";
Constant input_1_value, input_2_value, input_3_value;
input_1._activationFunc = &input_1_value;
input_2._activationFunc = &input_2_value;
input_3._activationFunc = &input_3_value;
//------------- make connections
output.addInput(&internal_1);
output.addInput(&internal_2);
internal_1.addInput(&input_1);
internal_1.addInput(&input_2);
internal_1.addInput(&input_3);
internal_2.addInput(&input_1);
internal_2.addInput(&input_2);
internal_2.addInput(&input_3);
Network network;
network.addNode(&output);
network.addNode(&input_1);
network.addNode(&input_2);
network.addNode(&input_3);
network.addNode(&internal_1);
network.addNode(&internal_2);
//--- training
float a[NUM_TRAINING_SAMPLE] = {0.5, 0.3, 1, 0.25, 0.9, 0.5, 0.41, 0.6, 0.3, 0.7};
float b[NUM_TRAINING_SAMPLE] = {0.3, 0.5, 0.2, 0.1, 0.8, 0.5, 0.4, 0.61, 0.31, 0.6};
float targetValue[NUM_TRAINING_SAMPLE] = {1, 0, 1, 1, 1, 0, 1, 0, 0, 1};
ForwardPass forwardPass;
BackwardPropagation backPropagation;
UpdateWeights updateWeights;
updateWeights._learningRate = 0.75;
for (size_t epoch=0; epoch<1000; ++epoch) {
for (size_t i=0; i<NUM_TRAINING_SAMPLE; ++i) {
input_1_value._value = a[i];
input_2_value._value = b[i];
input_3_value._value = -0.5;
output._criterion->_targetValue = targetValue[i];
forwardPass(&network);
backPropagation(&network);
updateWeights(&network);
}
}
//--- testing
float test_a[NUM_TESTING_SAMPLE] = {0.5, 0.3, 1, 0.2, 0.9, 0.5, 0.51, 0.61, 0.2, 0.4};
float test_b[NUM_TESTING_SAMPLE] = {0.3, 0.5, 0.2, 1, 0.8, 0.5, 0.5, 0.60, 0.1, 0.5};
float test_targetValue[NUM_TESTING_SAMPLE] = {1, 0, 1, 0, 1, 0, 1, 1, 1, 0};
int numCorrect = 0;
for (size_t i=0; i<NUM_TESTING_SAMPLE; ++i) {
input_1_value._value = test_a[i];
input_2_value._value = test_b[i];
input_3_value._value = -0.5;
output._criterion->_targetValue = test_targetValue[i];
forwardPass(&network);
if ( output.getValue() > 0.5
&& static_cast<int>(test_targetValue[i]) == 1) {
++numCorrect;
}
else if ( output.getValue() <= 0.5
&& static_cast<int>(test_targetValue[i]) == 0) {
++numCorrect;
}
}
cout << "accuracy: " << (numCorrect / 10.0f) << endl;
PrintNetwork printNetwork;
printNetwork(&network);
}
int main(){
cout << "Demonstrating Omurtag et al. (2000)" << endl;
Number n_bins = 500;
Potential V_min = 0.0;
NeuronParameter
par_neuron
(
1.0,
0.0,
0.0,
0.0,
50e-3
);
OdeParameter
par_ode
(
n_bins,
V_min,
par_neuron,
InitialDensityParameter(0.0,0.0)
);
double min_bin = 0.01;
LifNeuralDynamics dyn(par_ode,min_bin);
LeakingOdeSystem sys(dyn);
GeomParameter par_geom(sys);
GeomDelayAlg alg(par_geom);
Rate rate_ext = 800.0;
RateAlgorithm<MPILib::DelayedConnection> alg_ext(rate_ext);
Network network;
NodeId id_rate = network.addNode(alg_ext,EXCITATORY_DIRECT);
NodeId id_alg = network.addNode(alg, EXCITATORY_DIRECT);
MPILib::DelayedConnection con(1,0.03,0.0);
network.makeFirstInputOfSecond(id_rate,id_alg,con);
MPILib::report::handler::RootReportHandler handler("singlepoptest", true, true);
handler.addNodeToCanvas(id_alg);
const SimulationRunParameter
par_run
(
handler,
10000000,
0.0,
0.5,
1e-3,
1e-4,
"singlepoptest.log"
);
network.configureSimulation(par_run);
network.evolve();
return 0;
}
int main(){
Rate rate_ext = 0;
RateFunctor<double> input(Inp);
RateFunctor<double> openlock(Opl);
RateAlgorithm<double> alg_ext(rate_ext);
WilsonCowanParameter par_wc;
par_wc._f_bias = 0;
par_wc._f_noise = 1.0;
par_wc._rate_maximum = 50;
par_wc._time_membrane = 50e-3;
WilsonCowanAlgorithm alg(par_wc);
Network network;
NodeId id_input = network.addNode(input, EXCITATORY_DIRECT);
NodeId id_gate = network.addNode(alg, EXCITATORY_DIRECT);
NodeId id_output = network.addNode(alg, EXCITATORY_DIRECT);
NodeId id_lock = network.addNode(alg, INHIBITORY_DIRECT);
NodeId id_openlock = network.addNode(openlock ,INHIBITORY_DIRECT);
double weight = 1.0;
network.makeFirstInputOfSecond(id_input, id_gate, weight);
network.makeFirstInputOfSecond(id_input, id_lock, weight);
network.makeFirstInputOfSecond(id_gate, id_output, weight);
network.makeFirstInputOfSecond(id_lock, id_gate, -1*weight);
network.makeFirstInputOfSecond(id_openlock, id_lock, -2*weight);
MPILib::CanvasParameter par_canvas;
par_canvas._state_min = -0.020;
par_canvas._state_max = 0.020;
par_canvas._t_min = 0.0;
par_canvas._t_max = 2.0;
par_canvas._f_min = 0.0;
par_canvas._f_max = 50.0;
par_canvas._dense_min = 0.0;
par_canvas._dense_max = 200.0;
MPILib::report::handler::RootReportHandler handler("singlepoptest", true, true, par_canvas );
handler.addNodeToCanvas(id_input);
handler.addNodeToCanvas(id_gate);
handler.addNodeToCanvas(id_output);
handler.addNodeToCanvas(id_lock);
handler.addNodeToCanvas(id_openlock);
const SimulationRunParameter
par_run
(
handler,
10000000,
0.0,
2.0,
1e-3,
1e-3,
"wilsom.log"
);
network.configureSimulation(par_run);
network.evolve();
return 0;
}
int main(){
cout << "Demonstrating Quadratic-Integrate-and-Fire under jump response" << endl;
Number n_bins = 1000;
Potential V_min = -10.0;
NeuronParameter
par_neuron
(
10.0,
-10.0,
-10.0,
0.0,
10e-3
);
OdeParameter
par_ode
(
n_bins,
V_min,
par_neuron,
InitialDensityParameter(0.0,0.0)
);
GeomLib::QifParameter
par_qif
(
0.5, //
0.5 // default gamma sys
);
DiffusionParameter par_diffusion(0.03,0.03);
CurrentCompensationParameter par_current(0.0,0.0);
SpikingQifNeuralDynamics dyn(par_ode,par_qif);
QifOdeSystem sys(dyn);
GeomDelayAlg alg_qif(GeomParameter(sys, par_diffusion, par_current));
// Reproduce mu = 0.6, sigma = 0.7
Rate rate_ext = 6000;
Efficacy h = -8e-3;
RateAlgorithm<MPILib::DelayedConnection> alg_ext(rate_ext);
Network network;
NodeId id_rate = network.addNode(alg_ext, MPILib::INHIBITORY_GAUSSIAN);
NodeId id_alg = network.addNode(alg_qif, MPILib::EXCITATORY_GAUSSIAN);
MPILib::DelayedConnection con(1,h,0.0);
network.makeFirstInputOfSecond(id_rate,id_alg,con);
MPILib::CanvasParameter par_canvas;
par_canvas._state_min = -10.0;
par_canvas._state_max = 10.0;
par_canvas._t_min = 0.0;
par_canvas._t_max = 0.5;
par_canvas._f_min = 0.0;
par_canvas._f_max = 10.0;
par_canvas._dense_min = 0.0;
par_canvas._dense_max = 5.0;
MPILib::report::handler::RootReportHandler handler("twopopcanvas.root", true, true, par_canvas);
handler.addNodeToCanvas(id_alg);
const SimulationRunParameter
par_run
(
handler,
10000000,
0.0,
0.5,
2e-3,
1e-4,
"singlepoptest.log"
);
network.configureSimulation(par_run);
network.evolve();
return 0;
}
