public override void Connect(Link[] prior,NeuronBase[] next,Bias bias)
{
Utility.Verify(prior,x => x!=null && prior.Length>0,"Invalid prior layer");
Utility.Verify(next,x => x!=null && next.Length>0,"Invalid next layer");
Utility.Verify(() => next.Length % FeatureMapsCount==0,"FeatureMapsCount - next layer size conflict");
Utility.Verify(() => prior.Length % FeatureMapsCount==0,"FeatureMapsCount - prior layer size conflict");
Link[][] c=GetFeatureMaps(prior,FeatureMapsCount);
NeuronBase[][] s=GetFeatureMaps(next,FeatureMapsCount);
Link[][][] kernelCollection=GetKernelCollection(c,2,0); //always x/2
Dictionary<Link,List<Connection>> map=new Dictionary<Link,List<Connection>>();
for(var i=0;i<kernelCollection.Length;i++)
ConnectFeatureMaps(s[i],kernelCollection[i],bias,map);
foreach(Link link in prior)
link.Next= map[link].ToArray();
foreach(NeuronBase neuron in next)
neuron.Previous=map[neuron].ToArray();
if(bias!=null){
if(bias.Next==null)
bias.Next=new Connection[] { };
bias.Next=bias.Next.Concat(map[bias]).ToArray();
}
}
public NetworkStructure(Link[][] layers,Bias bias)
{
Utility.Verify(() => layers!=null && layers.All(x => x!=null&&x.Length>0),"layers");
Layers=layers;
Bias=bias;
Elements=Gather(layers,bias).ToArray();
}
IEnumerable<Connection> GetNeuronConnections(Link[] prior,NeuronBase neuron, Bias bias)
{
foreach(Link link in prior)
yield return new Connection(link,neuron);
if(bias!=null)
yield return new Connection(bias,neuron);
}
protected NetworkStructure GetLayers(int[] layers, Bias bias, IContinuousActivator func)
{
var l = layers.Length;
Link[][] structure = new Link[l][];
FullLayerConnector connector = new FullLayerConnector();
for (var i = 0; i < l - 1; i++)
{
if (i == 0)
{
structure[i] = Utility.Generate <Link>(layers[i]).ToArray();
}
int j = i + 1;
structure[j] = Utility.Generate <Neuron>(layers[j]).ToArray();
connector.Connect(structure[i], (NeuronBase[])structure[j], bias);
}
SetIdentity(structure);
foreach (Link[] layer in structure)
{
foreach (Link link in layer)
{
if (link is NeuronBase)
{
((NeuronBase)link).Func = func;
}
}
}
return(new NetworkStructure(structure, bias));
}
public NetworkStructure(Link[][] layers, Bias bias)
{
Utility.Verify(() => layers != null && layers.All(x => x != null && x.Length > 0), "layers");
Layers = layers;
Bias = bias;
Elements = Gather(layers, bias).ToArray();
}
public ConvolutionalNetwork Create(ConvolutionalTopology topology)
{
Utility.Verify(topology,t => t!=null,"Invalid network topology");
var l=topology.Layers.Length;
Link[][] layers=new Link[l][];
Bias bias=null;
if(topology.Bias.HasValue)
bias=new Bias(topology.Bias.Value);
for(var i=0; i<l-1;i++){
if(i==0)
layers[i]=Utility.Generate<Link>(topology.Layers[i]).ToArray();
int j=i+1;
int size=topology.Layers[j];
if(topology.Map[i] is SubSampling)
layers[j]= Utility.Generate<SubSamplingNeuron>(size).ToArray();
else
layers[j]= Utility.Generate<Neuron>(size).ToArray();
topology.Map[i].Connect(layers[i],(NeuronBase[])layers[j],bias);
}
SetIdentity(layers);
foreach(Link[] layer in layers)
foreach(Link link in layer)
if(link is NeuronBase)
((NeuronBase)link).Func=topology.Func;
return new ConvolutionalNetwork(new NetworkStructure(layers,bias));
}
protected NetworkStructure GetStructure(XElement root)
{
var unitElems = root.Elements("unit");
var weightElems = root.Elements("weight");
Dictionary <string, Tuple <Link, int> > unitsMap = GetUnits(unitElems).ToDictionary(pair => pair.Item1.GetDescription(), pair => pair);
Dictionary <string, Weight> weightsMap = GetWeights(weightElems).ToDictionary(pair => pair.Item1, pair => pair.Item2);
var connectionElems = root.Elements("connection");
Dictionary <Link, List <Connection> > mapNext = new Dictionary <Link, List <Connection> >();
Dictionary <Link, List <Connection> > mapPrevious = new Dictionary <Link, List <Connection> >();
foreach (XElement element in connectionElems)
{
Connection c = new Connection(weightsMap[element.Value], unitsMap[element.Attribute("from").Value].Item1, (NeuronBase)unitsMap[element.Attribute("to").Value].Item1);
if (!mapNext.ContainsKey(c.Previous))
{
mapNext.Add(c.Previous, new List <Connection>());
}
mapNext[c.Previous].Add(c);
if (!mapPrevious.ContainsKey(c.Next))
{
mapPrevious.Add(c.Next, new List <Connection>());
}
mapPrevious[c.Next].Add(c);
}
Link[] structure = unitsMap.Select(x => x.Value.Item1).ToArray();
foreach (Link link in structure)
{
if (mapNext.ContainsKey(link))
{
link.Next = mapNext[link].ToArray();
}
if (link is NeuronBase)
{
if (mapPrevious.ContainsKey(link))
{
((NeuronBase)link).Previous = mapPrevious[link].ToArray();
}
}
}
Bias bias = unitsMap.Select(x => x.Value.Item1).FirstOrDefault(x => x is Bias) as Bias;
int layersCount = int.Parse(root.Element("layers").Value);
List <List <Link> > layers = new List <List <Link> >(layersCount);
for (var i = 0; i < layersCount; i++)
{
layers.Add(new List <Link>());
}
foreach (var kvp in unitsMap)
{
int l = kvp.Value.Item2;
if (l != -1)
{
layers[l].Add(kvp.Value.Item1);
}
}
return(new NetworkStructure(layers.Select(x => x.ToArray()).ToArray(), bias));
}
public void TSSetUp()
{
l1= Utility.Generate<Link>(() => new Link(),36).ToArray();
l2=Utility.Generate<Neuron>(() => new Neuron(new Sigmoid()),4).ToArray();
Bias=new Bias(1);
Cma=new ConvolutionAuto(4,1,1,2);
Cma.Connect(l1,l2,Bias);
}
public void TSSetUp()
{
l1= Utility.Generate<Link>(() => new Link(),1176).ToArray();
l2=Utility.Generate<Neuron>(() => new Neuron(new Sigmoid()),1600).ToArray();
Func<double> wg= () => Weight++;
Bias=new Bias(1);
Cma=new Convolution(5,16,6,4,GetSchema());
Cma.Connect(l1,l2,Bias);
}
IEnumerable<Link> Gather(Link[][] layers,Bias bias)
{
foreach(Link[] layer in layers)
foreach(Link link in layer)
yield return link;
if(bias!=null)
yield return bias;
yield break;
}
public void TSSetUp()
{
l1= Utility.Generate<Link>(() => new Link(),4704).ToArray();
l2=Utility.Generate<Neuron>(() => new Neuron(new Sigmoid()),1176).ToArray();
Func<double> wg= () => Weight++;
Bias=new Bias(1);
SbSg=new SubSampling(6);
SbSg.Connect(l1,l2,Bias);
}
public override void Connect(Link[] prior, NeuronBase[] next, Bias bias)
{
Utility.Verify(prior, x => x != null && prior.Length > 0, "Invalid prior layer");
Utility.Verify(next, x => x != null && next.Length > 0, "Invalid next layer");
foreach (NeuronBase neuron in next)
{
neuron.Previous = GetNeuronConnections(prior, neuron, bias).ToArray();
}
int idx = 0;
foreach (Link link in prior)
{
link.Next = GetLinkConnections(idx++, next).ToArray();
}
if (bias != null)
{
var biasConnections = GetBiasConnections(bias, next);
if (bias.Next == null)
{
bias.Next = new Connection[] { }
}
;
bias.Next = bias.Next.Concat(biasConnections).ToArray();
}
}
IEnumerable <Connection> GetNeuronConnections(Link[] prior, NeuronBase neuron, Bias bias)
{
foreach (Link link in prior)
{
yield return(new Connection(link, neuron));
}
if (bias != null)
{
yield return(new Connection(bias, neuron));
}
}
IEnumerable <Connection> GetLinkConnections(int index, NeuronBase[] next)
{
foreach (NeuronBase neuron in next)
{
yield return(neuron.Previous[index]);
}
}
IEnumerable <Connection> GetBiasConnections(Bias bias, NeuronBase[] next)
{
foreach (NeuronBase neuron in next)
{
yield return(neuron.Previous[neuron.Previous.Length - 1]);
}
}
}
void ConnectFeatureMaps(NeuronBase[] next,Link[][] kernelCollection,Bias bias,Dictionary<Link,List<Connection>> map)
{
SharedWeight weight=new SharedWeight(next.Length);
SharedWeight biasWeight=new SharedWeight(next.Length);
for(var i=0;i<kernelCollection.Length;i++){
ConnectKernel(next[i],kernelCollection[i],weight,map);
if(bias!=null){
Map(bias,next[i],biasWeight,map);
}
}
}
public ConvolutionMapAuto(int kernelSize, int featureMapsCount, int overlap, Bias bias, Func <double> weightGenerator)
{
Utility.Verify(kernelSize, x => x > 0, "Invalid kernelSize value");
Utility.Verify(featureMapsCount, x => x > 0, "Invalid featureMapsCount value");
Utility.Verify(overlap, x => x >= 0, "Invalid overlap value");
Utility.Verify(weightGenerator, x => x != null, "Null weightGenerator");
KernelSize = kernelSize;
FeatureMapsCount = featureMapsCount;
Overlap = overlap;
WeightGenerator = weightGenerator;
Bias = bias;
}
void ConnectKernel(Link[] kernel,Node node,Weight[] weights, Bias bias,Dictionary<Link,int> ccd)
{
int l=kernel.Length;
if(bias!=null)
l++;
node.Previous=new Connection[l];
for(var i=0;i<kernel.Length;i++) {
kernel[i].Next=kernel[i].Next??new Connection[ccd[kernel[i]]];
Connect(kernel[i],node,weights[i]);
}
if(bias!=null)
Connect(bias,node,weights[kernel.Length]);
}
public ConvolutionMapAuto(int kernelSize, int featureMapsCount, int overlap, Bias bias, Func<double> weightGenerator)
{
Utility.Verify(kernelSize,x => x>0,"Invalid kernelSize value");
Utility.Verify(featureMapsCount,x => x>0,"Invalid featureMapsCount value");
Utility.Verify(overlap,x => x>=0,"Invalid overlap value");
Utility.Verify(weightGenerator,x => x!=null,"Null weightGenerator");
KernelSize=kernelSize;
FeatureMapsCount=featureMapsCount;
Overlap=overlap;
WeightGenerator=weightGenerator;
Bias=bias;
}
IEnumerable <Link> Gather(Link[][] layers, Bias bias)
{
foreach (Link[] layer in layers)
{
foreach (Link link in layer)
{
yield return(link);
}
}
if (bias != null)
{
yield return(bias);
}
yield break;
}
public override void Connect(Link[] prior,NeuronBase[] next,Bias bias)
{
Utility.Verify(prior,x => x!=null && prior.Length>0,"Invalid prior layer");
Utility.Verify(next,x => x!=null && next.Length>0,"Invalid next layer");
foreach(NeuronBase neuron in next)
neuron.Previous=GetNeuronConnections(prior,neuron,bias).ToArray();
int idx=0;
foreach(Link link in prior)
link.Next=GetLinkConnections(idx++,next).ToArray();
if(bias!=null){
var biasConnections=GetBiasConnections(bias,next);
if(bias.Next==null)
bias.Next=new Connection[] { };
bias.Next=bias.Next.Concat(biasConnections).ToArray();
}
}
public ConvolutionalNetwork Create(ConvolutionalTopology topology)
{
Utility.Verify(topology, t => t != null, "Invalid network topology");
var l = topology.Layers.Length;
Link[][] layers = new Link[l][];
Bias bias = null;
if (topology.Bias.HasValue)
{
bias = new Bias(topology.Bias.Value);
}
for (var i = 0; i < l - 1; i++)
{
if (i == 0)
{
layers[i] = Utility.Generate <Link>(topology.Layers[i]).ToArray();
}
int j = i + 1;
int size = topology.Layers[j];
if (topology.Map[i] is SubSampling)
{
layers[j] = Utility.Generate <SubSamplingNeuron>(size).ToArray();
}
else
{
layers[j] = Utility.Generate <Neuron>(size).ToArray();
}
topology.Map[i].Connect(layers[i], (NeuronBase[])layers[j], bias);
}
SetIdentity(layers);
foreach (Link[] layer in layers)
{
foreach (Link link in layer)
{
if (link is NeuronBase)
{
((NeuronBase)link).Func = topology.Func;
}
}
}
return(new ConvolutionalNetwork(new NetworkStructure(layers, bias)));
}
protected NetworkStructure GetLayers(int[] layers,Bias bias,IContinuousActivator func)
{
var l=layers.Length;
Link[][] structure=new Link[l][];
FullLayerConnector connector=new FullLayerConnector();
for(var i=0;i<l-1;i++) {
if(i==0)
structure[i]=Utility.Generate<Link>(layers[i]).ToArray();
int j=i+1;
structure[j]= Utility.Generate<Neuron>(layers[j]).ToArray();
connector.Connect(structure[i],(NeuronBase[])structure[j],bias);
}
SetIdentity(structure);
foreach(Link[] layer in structure)
foreach(Link link in layer)
if(link is NeuronBase)
((NeuronBase)link).Func=func;
return new NetworkStructure(structure,bias);
}
void ConnectKernel(Link[] kernel, Node node, Weight[] weights, Bias bias, Dictionary <Link, int> ccd)
{
int l = kernel.Length;
if (bias != null)
{
l++;
}
node.Previous = new Connection[l];
for (var i = 0; i < kernel.Length; i++)
{
kernel[i].Next = kernel[i].Next ?? new Connection[ccd[kernel[i]]];
Connect(kernel[i], node, weights[i]);
}
if (bias != null)
{
Connect(bias, node, weights[kernel.Length]);
}
}
public override void Connect(Link[] prior,NeuronBase[] next,Bias bias)
{
Utility.Verify(prior,x => x!=null && prior.Length>0,"Invalid prior layer");
Utility.Verify(next,x => x!=null && next.Length>0,"Invalid next layer");
Utility.Verify(() => next.Length % NextFeatureMapsCount==0,"NextFeatureMapsCount - next layer size conflict");
Utility.Verify(() => prior.Length % PriorFeatureMapsCount==0,"PriorFeatureMapsCount - prior layer size conflict");
Link[][] fmp=GetFeatureMaps(prior,PriorFeatureMapsCount);
Link[][] fmn=GetFeatureMaps(next,NextFeatureMapsCount);
Link[][][] kernelCollection=GetKernelCollection(fmp,KernelSize,Overlap);
Dictionary<Link,List<Connection>> map=new Dictionary<Link,List<Connection>>();
for(int f=0;f<fmn.Length;f++) {
Neuron[] nextFeatureMap=fmn[f].Select(x => (Neuron)x).ToArray();
for(int p=0;p<fmp.Length;p++) {
bool inSchema=Schema[p][f];
if(inSchema)
ConnectFeatureMaps(nextFeatureMap,kernelCollection[p],map);
}
if(bias!=null) {
Weight weight=new SharedWeight(nextFeatureMap.Length);
foreach(Neuron node in nextFeatureMap)
Map(bias,node,weight,map);
}
}
foreach(Link link in prior)
link.Next= map[link].ToArray();
foreach(Neuron node in next)
node.Previous=map[node].ToArray();
if(bias!=null){
if(bias.Next==null)
bias.Next=new Connection[] { };
bias.Next=bias.Next.Concat(map[bias]).ToArray();
}
}
public override void Connect(Link[] prior, NeuronBase[] next, Bias bias)
{
Utility.Verify(prior, x => x != null && prior.Length > 0, "Invalid prior layer");
Utility.Verify(next, x => x != null && next.Length > 0, "Invalid next layer");
Utility.Verify(() => next.Length % NextFeatureMapsCount == 0, "NextFeatureMapsCount - next layer size conflict");
Utility.Verify(() => prior.Length % PriorFeatureMapsCount == 0, "PriorFeatureMapsCount - prior layer size conflict");
Link[][] fmp = GetFeatureMaps(prior, PriorFeatureMapsCount);
Link[][] fmn = GetFeatureMaps(next, NextFeatureMapsCount);
Link[][][] kernelCollection = GetKernelCollection(fmp, KernelSize, Overlap);
Dictionary <Link, List <Connection> > map = new Dictionary <Link, List <Connection> >();
for (int f = 0; f < fmn.Length; f++)
{
Neuron[] nextFeatureMap = fmn[f].Select(x => (Neuron)x).ToArray();
for (int p = 0; p < fmp.Length; p++)
{
bool inSchema = Schema[p][f];
if (inSchema)
{
ConnectFeatureMaps(nextFeatureMap, kernelCollection[p], map);
}
}
if (bias != null)
{
Weight weight = new SharedWeight(nextFeatureMap.Length);
foreach (Neuron node in nextFeatureMap)
{
Map(bias, node, weight, map);
}
}
}
foreach (Link link in prior)
{
link.Next = map[link].ToArray();
}
foreach (Neuron node in next)
{
node.Previous = map[node].ToArray();
}
if (bias != null)
{
if (bias.Next == null)
{
bias.Next = new Connection[] { }
}
;
bias.Next = bias.Next.Concat(map[bias]).ToArray();
}
}
void ConnectFeatureMaps(Neuron[] next, Link[][] kernelCollection, Dictionary <Link, List <Connection> > map)
{
Weight[] weights = Utility.Generate(() => new SharedWeight(kernelCollection.Length), KernelSize * KernelSize).ToArray();
for (int i = 0; i < kernelCollection.Length; i++)
{
for (var j = 0; j < kernelCollection[i].Length; j++)
{
Link link = kernelCollection[i][j];
Map(link, next[i], weights[j], map);
}
}
}
void Map(Link a, Link b, Weight weight, Dictionary <Link, List <Connection> > map)
{
Connection connection = new Connection(weight, a, (NeuronBase)b);
if (!map.ContainsKey(a))
{
map[a] = new List <Connection>();
}
map[a].Add(connection);
if (!map.ContainsKey(b))
{
map[b] = new List <Connection>();
}
map[b].Add(connection);
}
}
IEnumerable<Connection> GetBiasConnections(Bias bias, NeuronBase[] next)
{
foreach(NeuronBase neuron in next)
yield return neuron.Previous[neuron.Previous.Length-1];
}
public void TSSetUp()
{
l1= Utility.Generate<Link>(() => new Link(),841).ToArray();
l2=Utility.Generate<Neuron>(() => new Neuron(new Sigmoid()),1014).ToArray();
Func<double> wg= () => Weight++;
Bias=new Bias(1);
Cma=new ConvolutionAuto(5,6,1,3);
Cma.Connect(l1,l2,Bias);
}
public override void Connect(Link[] prior, NeuronBase[] next, Bias bias)
{
Utility.Verify(prior, x => x != null && prior.Length > 0, "Invalid prior layer");
Utility.Verify(next, x => x != null && next.Length > 0, "Invalid next layer");
Utility.Verify(() => next.Length % FeatureMapsCount == 0, "FeatureMapsCount - next layer size conflict");
Utility.Verify(() => prior.Length % FeatureMapsCount == 0, "FeatureMapsCount - prior layer size conflict");
Link[][] c = GetFeatureMaps(prior, FeatureMapsCount);
NeuronBase[][] s = GetFeatureMaps(next, FeatureMapsCount);
Link[][][] kernelCollection = GetKernelCollection(c, 2, 0); //always x/2
Dictionary <Link, List <Connection> > map = new Dictionary <Link, List <Connection> >();
for (var i = 0; i < kernelCollection.Length; i++)
{
ConnectFeatureMaps(s[i], kernelCollection[i], bias, map);
}
foreach (Link link in prior)
{
link.Next = map[link].ToArray();
}
foreach (NeuronBase neuron in next)
{
neuron.Previous = map[neuron].ToArray();
}
if (bias != null)
{
if (bias.Next == null)
{
bias.Next = new Connection[] { }
}
;
bias.Next = bias.Next.Concat(map[bias]).ToArray();
}
}
void ConnectFeatureMaps(NeuronBase[] next, Link[][] kernelCollection, Bias bias, Dictionary <Link, List <Connection> > map)
{
SharedWeight weight = new SharedWeight(next.Length);
SharedWeight biasWeight = new SharedWeight(next.Length);
for (var i = 0; i < kernelCollection.Length; i++)
{
ConnectKernel(next[i], kernelCollection[i], weight, map);
if (bias != null)
{
Map(bias, next[i], biasWeight, map);
}
}
}
void ConnectKernel(NeuronBase neuron, Link[] kernel, SharedWeight weight, Dictionary <Link, List <Connection> > map)
{
foreach (Link link in kernel)
{
Map(link, neuron, weight, map);
}
}
void Map(Link a, Link b, Weight weight, Dictionary <Link, List <Connection> > map)
{
Connection connection = new Connection(weight, a, (NeuronBase)b);
if (!map.ContainsKey(a))
{
map[a] = new List <Connection>();
}
map[a].Add(connection);
if (!map.ContainsKey(b))
{
map[b] = new List <Connection>();
}
map[b].Add(connection);
}
}
public abstract void Connect(Link[] prior,NeuronBase[] next, Bias bias);
public MultilayerPerceptron Create(int[] layers, double?biasValue, IContinuousActivator func)
{
Bias bias = biasValue.HasValue? new Bias(biasValue.Value):null;
return(new MultilayerPerceptron(GetLayers(layers, bias, func)));
}
public abstract void Connect(Link[] prior, NeuronBase[] next, Bias bias);
