/// <summary>
/// Constructs a AcyclicNeuralNet with the provided neural net definition parameters.
/// </summary>
/// <param name="digraph">Network structure definition</param>
/// <param name="weightArr">Connection weights array.</param>
/// <param name="activationFn">Node activation function.</param>
public NeuralNetAcyclic(
DirectedGraphAcyclic digraph,
double[] weightArr,
VecFnSegment <double> activationFn)
{
// Store refs to network structure data.
_srcIdArr = digraph.ConnectionIdArrays._sourceIdArr;
_tgtIdArr = digraph.ConnectionIdArrays._targetIdArr;
_weightArr = weightArr;
_layerInfoArr = digraph.LayerArray;
// Store network activation function.
_activationFn = activationFn;
// Store input/output node counts.
_inputCount = digraph.InputCount;
_outputCount = digraph.OutputCount;
// Create working array for node activation signals.
_activationArr = new double[digraph.TotalNodeCount];
// Wrap a sub-range of the _activationArr that holds the activation values for the input nodes.
this.InputVector = new VectorSegment <double>(_activationArr, 0, _inputCount);
// Wrap the output nodes. Nodes have been sorted by depth within the network therefore the output
// nodes can no longer be guaranteed to be in a contiguous segment at a fixed location. As such their
// positions are indicated by outputNodeIdxArr, and so we package up this array with the node signal
// array to abstract away the indirection described by outputNodeIdxArr.
this.OutputVector = new MappingVector <double>(_activationArr, digraph.OutputNodeIdxArr);
}
/// <summary>
/// Constructs a AcyclicNeuralNet with the provided neural net definition parameters.
/// </summary>
/// <param name="digraph">Network structure definition.</param>
/// <param name="weightArr">Connection weights array.</param>
/// <param name="activationFn">Node activation function.</param>
public NeuralNetAcyclicSafe(
DirectedGraphAcyclic digraph,
double[] weightArr,
VecFn <double> activationFn)
{
Debug.Assert(digraph.ConnectionIds.GetSourceIdSpan().Length == weightArr.Length);
// Store refs to network structure data.
_connIds = digraph.ConnectionIds;
_weightArr = weightArr;
_layerInfoArr = digraph.LayerArray;
// Store network activation function.
_activationFn = activationFn;
// Store input/output node counts.
_inputCount = digraph.InputCount;
_outputCount = digraph.OutputCount;
// Create working array for node activation signals.
_activationArr = new double[digraph.TotalNodeCount];
// Map the inputs vector to the corresponding segment of node activation values.
this.Inputs = new Memory <double>(_activationArr, 0, _inputCount);
// Get an array to act a as a contiguous run of output signals.
_outputArr = new double[_outputCount];
this.Outputs = _outputArr;
// Store the indexes into _activationArr that give the output signals.
_outputNodeIdxArr = digraph.OutputNodeIdxArr;
}
/// <summary>
/// Create a NeatGenome with the given meta data, connection genes and supplementary data.
/// </summary>
/// <param name="id">Genome ID.</param>
/// <param name="birthGeneration">Birth generation.</param>
/// <param name="connGenes">Connection genes.</param>
/// <param name="hiddenNodeIdArr">An array of the hidden node IDs in the genome, in ascending order.</param>
/// <returns>A new NeatGenome instance.</returns>
public NeatGenome <T> Create(
int id, int birthGeneration,
ConnectionGenes <T> connGenes,
int[] hiddenNodeIdArr)
{
int inputCount = _metaNeatGenome.InputNodeCount;
// Create a mapping from node IDs to node indexes.
Dictionary <int, int> nodeIdxById = BuildNodeIndexById(hiddenNodeIdArr);
// Create a DictionaryNodeIdMap.
DictionaryNodeIdMap nodeIndexByIdMap = new DictionaryNodeIdMap(inputCount, nodeIdxById);
// Create a digraph from the genome.
DirectedGraph digraph = NeatGenomeBuilderUtils.CreateDirectedGraph(
_metaNeatGenome, connGenes, nodeIndexByIdMap);
// Calc the depth of each node in the digraph.
GraphDepthInfo depthInfo = _graphDepthAnalysis.CalculateNodeDepths(digraph);
// Create a weighted acyclic digraph.
// Note. This also outputs connectionIndexMap. For each connection in the acyclic graph this gives
// the index of the same connection in the genome; this is because connections are re-ordered based
// on node depth in the acyclic graph.
DirectedGraphAcyclic acyclicDigraph = DirectedGraphAcyclicBuilderUtils.CreateDirectedGraphAcyclic(
digraph,
depthInfo,
out int[] newIdByOldId,
out int[] connectionIndexMap,
ref _timesortWorkArr,
ref _timesortWorkVArr);
// TODO: Write unit tests to cover this!
// Update nodeIdxById with the new depth based node index allocations.
// Notes.
// The current nodeIndexByIdMap maps node IDs (also know as innovation IDs in NEAT) to a compact
// ID space in which any gaps have been removed, i.e. a compacted set of IDs that can be used as indexes,
// i.e. if there are N nodes in total then the highest node ID will be N-1.
//
// Here we map the new compact IDs to an alternative ID space that is also compact, but ensures that nodeIDs
// reflect the depth of a node in the acyclic graph.
UpdateNodeIndexById(nodeIdxById, hiddenNodeIdArr, newIdByOldId);
// Create the neat genome.
return(new NeatGenome <T>(
_metaNeatGenome, id, birthGeneration,
connGenes,
hiddenNodeIdArr,
nodeIndexByIdMap,
acyclicDigraph,
connectionIndexMap));
}
/// <summary>
/// Create an array that gives the node layer for each node, keyed by node index.
/// </summary>
/// <param name="digraphAcyclic">The directed acyclic graph.</param>
/// <returns>A new integer array.</returns>
private static int[] BuildNodeLayerByIdx_Acyclic(DirectedGraphAcyclic digraphAcyclic)
{
// Alloc the array, and populate with the already determined node depth values as represented
// by the layer info provided by DirectedGraphAcyclic.
int nodeCount = digraphAcyclic.TotalNodeCount;
int[] nodeLayerByIdx = new int[nodeCount];
// Use a restricted scope for the loop variables.
{
// The first layer is layer 1; layer zero will be used later to represent input nodes only.
int layerIdx = 1;
int nodeIdx = 0;
foreach (LayerInfo layerInfo in digraphAcyclic.LayerArray)
{
for (; nodeIdx < layerInfo.EndNodeIdx; nodeIdx++)
{
nodeLayerByIdx[nodeIdx] = layerIdx;
}
layerIdx++;
}
}
// Assign input nodes to their own layer (layer zero).
for (int i = 0; i < digraphAcyclic.InputCount; i++)
{
nodeLayerByIdx[i] = 0;
}
// Assign output nodes to their own layer.
int outputLayerIdx = digraphAcyclic.LayerArray.Length + 1;
foreach (int outputNodeIdx in digraphAcyclic.OutputNodeIdxArr)
{
nodeLayerByIdx[outputNodeIdx] = outputLayerIdx;
}
// Remove empty layers (if any), by adjusting the depth values in nodeLayerByIdx.
RemoveEmptyLayers(nodeLayerByIdx, outputLayerIdx + 1);
// Return the constructed node layer lookup table.
return(nodeLayerByIdx);
}
/// <summary>
/// Constructs a AcyclicNeuralNet with the provided neural net definition parameters.
/// </summary>
/// <param name="digraph">Network structure definition.</param>
/// <param name="weightArr">Connection weights array.</param>
/// <param name="activationFn">Node activation function.</param>
public NeuralNetAcyclic(
DirectedGraphAcyclic digraph,
double[] weightArr,
VecFn <double> activationFn)
{
Debug.Assert(digraph.ConnectionIds.GetSourceIdSpan().Length == weightArr.Length);
// Store refs to network structure data.
_connIds = digraph.ConnectionIds;
_weightArr = weightArr;
_layerInfoArr = digraph.LayerArray;
// Store network activation function.
_activationFn = activationFn;
// Store input/output node counts.
_inputCount = digraph.InputCount;
_outputCount = digraph.OutputCount;
_totalNodeCount = digraph.TotalNodeCount;
// Get a working array for node activations signals and a separate output signal segment on the end.
// And map the memory segments onto the array.
_workingArr = ArrayPool <double> .Shared.Rent(_totalNodeCount + _outputCount);
_activationMem = _workingArr.AsMemory(0, _totalNodeCount);
_outputMem = _workingArr.AsMemory(_totalNodeCount, _outputCount);
// Map the inputs vector to the corresponding segment of node activation values.
this.Inputs = _activationMem.Slice(0, _inputCount);
// Use the already defined outputs memory segment.
this.Outputs = _outputMem;
// Store the indexes into _activationArr that give the output signals.
_outputNodeIdxArr = digraph.OutputNodeIdxArr;
}