private void testNeuron_Click(object sender, EventArgs e)
{
Port[] neuronPorts = new Port[2];
neuronPorts[0] = new Port(Port.portDirection.IN, "in_0", Utilities.VHDLDataType.SIGNED_FIXED_POINT, 3, -4);
neuronPorts[1] = new Port(Port.portDirection.IN, "in_1", Utilities.VHDLDataType.SIGNED_FIXED_POINT, 3, -4);
AsyncNeuron n = new AsyncNeuron(neuronPorts, 4, 4, 4, AsyncNeuron.NeuronActivationType.SIGMOID_POLY_APPROX, "neuron_intfrac");
if (!n.writeVHDL("neuron_intfrac.vhd"))
{
MessageBox.Show("Problem in writing neuron_intfrac.vhd");
return;
}
MessageBox.Show("Files successfully written.");
return;
}
/// <summary>
/// Determines if two neurons are equal
/// </summary>
/// <param name="obj">Object to compare to this neuron</param>
/// <returns>True if the objects are equal, false otherwise</returns>
public override bool Equals(object obj)
{
if (obj == null)
{
return(false);
}
AsyncNeuron n_obj = obj as AsyncNeuron;
if (n_obj == null)
{
return(false);
}
var n_obj_inputPorts = n_obj.getInputPorts();
var n_obj_internalSignals = n_obj.getInternalSignals();
var n_obj_outputs = n_obj.getOutputPorts();
var n_obj_activation = n_obj.activationType;
if (n_obj.activationType != this.activationType)
{
return(false);
}
if (n_obj_inputPorts.Length != this.neuronInputs.Length || n_obj_internalSignals.Length != this.getInternalSignals().Length)
{
return(false);
}
/*TODO: Handle zero edge cases*/
if (n_obj_inputPorts[0].top != this.neuronInputs[0].top || n_obj_inputPorts[0].bottom != this.neuronInputs[0].bottom || n_obj_inputPorts[0].type != this.neuronInputs[0].type)
{
return(false);
}
if (n_obj_outputs[0].top != this.neuronOutput.top || n_obj_outputs[0].bottom != this.neuronOutput.bottom || n_obj_outputs[0].type != this.neuronOutput.type)
{
return(false);
}
if (n_obj.numOutputIntBits != this.numOutputIntBits || n_obj.numOutputFracBits != this.numOutputFracBits)
{
return(false);
}
return(true);
}
/// <summary>
/// Writes the .vhd files necessary to compile this entity.
/// All other necessary entities (i.e. neurons, thresholding functions, etc.) will also be written
/// when this function returns
/// </summary>
/// <param name="file">The file path in which to write the files (do NOT include "...name.vhd"</param>
/// <returns>true if the files were written successfully, false otherwise</returns>
public override bool writeVHDL(string file)
{
bool successfulWrite = false;
string entityFile = file + this.name + ".vhd";
using (StreamWriter sw = new StreamWriter(entityFile))
{
/*Compile all of the entities into one list*/
List <Entity> uniqueEntities = new List <Entity>();
uniqueEntities.Add(this.wm);
for (int i = 0; i < this.uniqueNeuronEntities.Count; i++)
{
uniqueEntities.Add(this.uniqueNeuronEntities[i]);
}
/*Write header*/
if (!((this.writeVHDLIncludes(sw)) && (this.writeVHDLEntity(sw)) && (this.writeArchitectureStatement(sw)) && (this.writeVHDLDependencies(sw, uniqueEntities)) && (this.writeVHDLSignals(sw))))
{
return(false);
}
sw.WriteLine(String.Format("\t{0} <= {1};", this.ready.portName, this.ready_signal.name));
/*Write the memory signals*/
sw.WriteLine(String.Format("\twm: {0} port map", this.wm.name));
sw.WriteLine(String.Format("\t("));
sw.WriteLine(String.Format("\t\t{0} => std_logic_vector({1}),", this.wm.addr.portName, this.addrCounter.name));
sw.WriteLine(String.Format("\t\t{0} => {1},", this.wm.clk.portName, this.clk.portName));
sw.WriteLine(String.Format("\t\t{0} => {1}", this.wm.data.portName, this.loadValConv.name));
sw.WriteLine(String.Format("\t);"));
sw.WriteLine(String.Format(""));
sw.WriteLine(String.Format("\t{0} <= to_sfixed({1}, {0}'high, {0}'low);", this.loadVal.name, this.loadValConv.name));
sw.WriteLine("");
/*Write the neural network connections*/
int currentNeuron = 0;
int previousLayerStartingPoint = 0;
for (int i = 1; i < this.neuronCounts.Length; i++)
{
for (int j = 0; j < this.neuronCounts[i]; j++)
{
AsyncNeuron n = this.neuronEntities[currentNeuron];
sw.WriteLine(String.Format("\tNeuron_{0}_{1} : {2} port map", i, j, n.name));
sw.WriteLine(String.Format("\t("));
/*INPUT STATEMENTS*/
/*First row - use inputs*/
if (i == 1)
{
for (int k = 0; k < this.nnInputPorts.Length; k++)
{
sw.WriteLine(String.Format("\t\t{0} => {1},", n.neuronInputs[k].portName, this.nnInputPorts[k].portName));
}
}
/*Later row - use previous rows outputs*/
else
{
for (int k = 0; k < this.neuronEntities[currentNeuron].neuronInputs.Length; k++)
{
sw.WriteLine(String.Format("\t\t{0} => {1},", n.neuronInputs[k].portName, this.neuron_outputs[previousLayerStartingPoint + k].name));
}
}
sw.WriteLine(String.Format("\t\t{0} => {1}({2}),", n.loadEnablePort.portName, this.load_sig.name, currentNeuron)); //Load signal
sw.WriteLine(String.Format("\t\t{0} => {1},", n.loadValuePort.portName, this.loadVal.name)); //LoadVal signal
sw.WriteLine(String.Format("\t\t{0} => {1}({2} downto 0),", n.loadOffsetPort.portName, this.loadOffset.name, n.loadOffsetPort.top)); //LoadOff signal
sw.WriteLine(String.Format("\t\t{0} => {1},", n.finalLoadPort.portName, this.finalLoadSignals[currentNeuron].name)); //FinalLoad signal
sw.WriteLine(String.Format("\t\t{0} => {1},", n.loadClkPort.portName, this.clk.portName)); //Clk signal
sw.WriteLine(String.Format("\t\t{0} => {1}", n.neuronOutput.portName, this.neuron_outputs[currentNeuron].name)); //Output signal
sw.WriteLine("\t);");
sw.WriteLine("");
currentNeuron++;
}
if (i != 1)
{
previousLayerStartingPoint += this.neuronCounts[i - 1];
}
}
/*Load process*/
sw.WriteLine(String.Format("\tprocess ({0})", this.clk.portName));
sw.WriteLine(String.Format("\tbegin"));
sw.WriteLine(String.Format("\t\tif (rising_edge({0}) and ( ({1} = '1') or ({2} = '0') ) ) then", this.clk.portName, this.reset.portName, this.ready_signal.name));
sw.WriteLine(String.Format("\t\t\tif ({0} = '1') then", this.reset.portName));
sw.WriteLine(String.Format("\t\t\t\t{0} <= '0';", this.ready_signal.name));
sw.WriteLine(String.Format("\t\t\t\t{0} <= (others => '0');", this.loadOffset.name));
sw.WriteLine(String.Format("\t\t\t\t{0} <= (others => '0');", this.addrCounter.name));
sw.WriteLine(String.Format("\t\t\t\t{0} <= (0 => '1', others => '0');", this.load_sig.name));
for (int i = 0; i < this.finalLoadSignals.Length; i++)
{
sw.WriteLine(String.Format("\t\t\telsif ({0} = '1') then", this.finalLoadSignals[i].name));
if (i != this.finalLoadSignals.Length - 1)
{
sw.WriteLine(String.Format("\t\t\t\t{0} <= ({1} => '1', others => '0');", this.load_sig.name, i + 1));
}
else
{
sw.WriteLine(String.Format("\t\t\t\t{0} <= (others => '0');", this.load_sig.name));
}
sw.WriteLine(String.Format("\t\t\t\t{0} <= (others => '0');", this.loadOffset.name));
}
sw.WriteLine(String.Format("\t\t\telse"));
sw.WriteLine(String.Format("\t\t\t\t{0} <= {0} + 1;", this.loadOffset.name));
sw.WriteLine(String.Format("\t\t\t\t{0} <= {0} + 1;", this.addrCounter.name));
sw.WriteLine(String.Format("\t\t\tend if;"));
sw.WriteLine(String.Format(""));
sw.WriteLine(String.Format("\t\t\tif ({0} = {1}) then", this.addrCounter.name, wm.weights_str.Count));
sw.WriteLine(String.Format("\t\t\t\t{0} <= '1';", this.ready_signal.name));
sw.WriteLine(String.Format("\t\t\tend if;"));
sw.WriteLine(String.Format("\t\tend if;"));
sw.WriteLine(String.Format("\tend process;"));
sw.WriteLine("");
/*Write the comparitors for classification problems*/
if (this.isClassifier)
{
int currentOutputNeuronIndex = this.neuronEntities.Length - this.neuronCounts[this.neuronCounts.Length - 1];
int outputIndex = 0;
while (currentOutputNeuronIndex < this.neuronEntities.Length)
{
Signal current_out = this.neuron_outputs[currentOutputNeuronIndex];
sw.WriteLine(String.Format("\tprocess ({0})", current_out.name));
sw.WriteLine(String.Format("\tbegin"));
sw.WriteLine(String.Format("\t\tif ({0} > to_sfixed({1}, {0})) then", current_out.name, this.classifierThresholds[outputIndex]));
sw.WriteLine(String.Format("\t\t\t{0} <= '1';", this.nnOutputPorts[outputIndex].portName));
sw.WriteLine(String.Format("\t\telse"));
sw.WriteLine(String.Format("\t\t\t{0} <= '0';", this.nnOutputPorts[outputIndex].portName));
sw.WriteLine(String.Format("\t\tend if;"));
sw.WriteLine(String.Format("\tend process;"));
currentOutputNeuronIndex++;
outputIndex++;
}
}
/*Write the raw signal outputs for regression problems*/
else
{
int currentOutputNeuronIndex = this.neuronEntities.Length - this.neuronCounts[this.neuronCounts.Length - 1];
int outputIndex = 0;
while (currentOutputNeuronIndex < this.neuronEntities.Length)
{
Signal current_out = this.neuron_outputs[currentOutputNeuronIndex];
sw.WriteLine(String.Format("\t{0} <= {1}", this.nnOutputPorts[outputIndex].portName, current_out.name));
sw.WriteLine("");
currentOutputNeuronIndex++;
outputIndex++;
}
}
/*Write dependent entities*/
foreach (Entity e in uniqueEntities)
{
if (!e.writeVHDL(file))
{
return(false);
}
}
successfulWrite = this.writeVHDLFooter(sw);
}
return(successfulWrite);
}
/// <summary>
/// The constructor for a neural network object. Calling this method's write() function will write the entire network.
/// </summary>
/// <param name="_neuronCounts">The number of neurons in each layer, excluding the bias node.</param>
/// <param name="_biasValues">The bias value feeding forward into the next layer. </param>
/// <param name="_activationTypes">The activation types for the neurons in each layer</param>
/// <param name="_numIntBits">The number of integer bits to use for the inputs for each neuron (Sigmoid activated neurons automatically get this set to zero)</param>
/// <param name="_numFracBits">The number of fractional bits to use for the inputs for each neuron</param>
/// <param name="_numWeightUpperBits">The number of integer bits used for the weights for each neuron.</param>
/// <param name="_isClassifier">If true, a comparitor will be instantiatated at the output to each node, comparing the output layer nodes to the threshold value.</param>
/// <param name="_classifierThresholds">The threshold values for the output comparitors</param>
/// <param name="_weights">The list of weights read in for the neurons from WeightReader.readWeightsFromFile()</param>
/// <remark> For example, to intialize a classification neural network with three inputs, two hidden sigmoid-activated nodes, and one linear output node, use the following line:</remark>
/// <remarks> NeuralNetwork nn = new NeuralNetwork([3, 2, 1], [-1, -1], [..SIGMOID_POLY_APPROX, ..LINEAR], 4, 4, 4, true, [0.5], _weights);</remarks>
public AsyncNeuralNetwork(int[] _neuronCounts, double[] _biasValues, AsyncNeuron.NeuronActivationType[] _activationTypes, int _numIntBits, int _numFracBits, int _numWeightUpperBits, bool _isClassifier, double[] _classifierThresholds, List <double> _weights)
{
/*Copy member variables*/
this.neuronCounts = _neuronCounts;
this.biasValues = _biasValues;
this.activationTypes = _activationTypes;
this.numIntBits = _numIntBits;
this.numFracBits = _numFracBits;
this.numWeightUpperBits = _numWeightUpperBits;
this.classifierThresholds = _classifierThresholds;
this.isClassifier = _isClassifier;
/*Some useful variables for computing the signals*/
int numNeurons = 0;
int maxWeights = neuronCounts[0];
for (int i = 1; i < neuronCounts.Length; i++)
{
numNeurons += neuronCounts[i];
if (maxWeights < neuronCounts[i])
{
maxWeights = neuronCounts[i];
}
}
int offsetWidth = Utilities.getNumUnsignedBits(maxWeights);
int addrCounterWidth = Utilities.getNumUnsignedBits(_weights.Count);
/*Initialize memory signals and entities*/
this.wm = new WeightMemory(_numWeightUpperBits, _numFracBits, _weights);
this.load_sig = new Signal("load_sig", Utilities.VHDLDataType.STD_LOGIC_VECTOR, new String('0', numNeurons), numNeurons - 1, 0);
this.loadVal = new Signal("loadVal", Utilities.VHDLDataType.SIGNED_FIXED_POINT, null, numWeightUpperBits, -1 * numFracBits); //NOT -1 => signed
this.loadOffset = new Signal("loadOff", Utilities.VHDLDataType.UNSIGNED, new String('0', offsetWidth), offsetWidth - 1, 0);
this.addrCounter = new Signal("addrCounter", Utilities.VHDLDataType.UNSIGNED, new String('0', addrCounterWidth), addrCounterWidth - 1, 0);
this.ready_signal = new Signal("ready_sig", Utilities.VHDLDataType.STD_LOGIC, null, 0, 0);
this.finalLoadSignals = new Signal[numNeurons];
int arrCounter = 0;
for (int i = 1; i < neuronCounts.Length; i++)
{
for (int j = 0; j < neuronCounts[i]; j++)
{
this.finalLoadSignals[arrCounter] = new Signal(String.Format("finalLoad_{0}_{1}", i - 1, j), Utilities.VHDLDataType.STD_LOGIC, null, 0, 0);
arrCounter++;
}
}
this.loadValConv = new Signal("loadValConv", Utilities.VHDLDataType.STD_LOGIC_VECTOR, null, _numWeightUpperBits + _numFracBits, 0);
/*Compute the bus widths between neurons*/
int maxNeuronInput = Convert.ToInt32(Math.Pow(2, this.numIntBits)) - 1;
int currentNeuronInput = Convert.ToInt32(Math.Pow(2, this.numIntBits)) - 1;
this.intBusWidths = new int[numNeurons + this.neuronCounts[0]];
double[] maxValues = new double[numNeurons + this.neuronCounts[0]];
int neuronStart = 0;
int weightCounter = 0;
for (neuronStart = 0; neuronStart < this.neuronCounts[0]; neuronStart++)
{
maxValues[neuronStart] = (double)maxNeuronInput;
this.intBusWidths[neuronStart] = Utilities.getNumUnsignedBits(Convert.ToInt32(Math.Ceiling((double)maxNeuronInput)));
}
neuronStart = 0;
for (int i = 1; i < this.neuronCounts.Length; i++)
{
int[] currentNeuronMaxIntWidths = new int[this.neuronCounts[i]];
if (this.activationTypes[i - 1] == AsyncNeuron.NeuronActivationType.SIGMOID_POLY_APPROX)
{
for (int j = 0; j < this.neuronCounts[i]; j++)
{
maxValues[neuronStart + this.neuronCounts[i - 1]] = 0.9999;
this.intBusWidths[neuronStart + this.neuronCounts[i - 1]] = 0;
neuronStart++;
}
}
else if (this.activationTypes[i - 1] == AsyncNeuron.NeuronActivationType.NONE)
{
throw new Exception("YOU DUN F****D UP HERE TOO"); //debugging
}
else if (this.activationTypes[i - 1] == AsyncNeuron.NeuronActivationType.LINEAR)
{
for (int j = 0; j < this.neuronCounts[i]; j++)
{
double currentMax = 0;
int neuronIndex = neuronStart;
for (int k = 0; k < this.neuronCounts[i - 1]; k++)
{
currentMax += _weights[weightCounter] * maxValues[neuronIndex];
weightCounter++;
neuronIndex++;
}
maxValues[j + neuronStart + this.neuronCounts[i - 1]] = currentMax;
this.intBusWidths[j + neuronStart + this.neuronCounts[i - 1]] = Utilities.getNumUnsignedBits(Convert.ToInt32(Math.Ceiling(currentMax)));
}
neuronStart += this.neuronCounts[i - 1];
}
}
if (this.isClassifier)
{
for (int i = this.intBusWidths.Length - 1; (this.intBusWidths.Length - i) <= this.neuronCounts[this.neuronCounts.Length - 1]; i--)
{
this.intBusWidths[i] = 1;
}
}
/*Initialize neurons and their connectivities*/
this.neuronEntities = new AsyncNeuron[numNeurons];
this.neuron_outputs = new Signal[numNeurons];
this.uniqueNeuronEntities = new List <AsyncNeuron>();
var currentPorts = new List <Port>();
this.nnInputPorts = new Port[this.neuronCounts[0]];
neuronStart = 0;
for (int i = 0; i < this.neuronCounts[0]; i++)
{
this.nnInputPorts[i] = new Port(Port.portDirection.IN, String.Format("nnIn_{0}", i), Utilities.VHDLDataType.SIGNED_FIXED_POINT, this.intBusWidths[neuronStart], this.numFracBits * -1);
neuronStart++;
}
for (int i = 0; i < this.nnInputPorts.Length; i++)
{
Port P = this.nnInputPorts[i].copy();
P.rename(String.Format("in_{0}", i));
currentPorts.Add(P);
}
arrCounter = 0;
for (int i = 1; i < this.neuronCounts.Length; i++)
{
var lastPorts = currentPorts.ToArray();
currentPorts = new List <Port>();
if (this.activationTypes[i - 1] == AsyncNeuron.NeuronActivationType.SIGMOID_POLY_APPROX)
{
for (int j = 0; j < this.neuronCounts[i]; j++)
{
AsyncNeuron n = new AsyncNeuron(lastPorts, this.intBusWidths[neuronStart], this.numFracBits, this.numWeightUpperBits, AsyncNeuron.NeuronActivationType.SIGMOID_POLY_APPROX, String.Format("sig_poly_approx_neuron_{0}_{1}", i, 0));
neuronStart++;
if (j == 0)
{
/*Check for uniqueness*/
if (!this.uniqueNeuronEntities.Contains(n))
{
this.uniqueNeuronEntities.Add(n);
}
}
this.neuronEntities[arrCounter] = n;
Signal currentNeuronOutput = n.neuronOutput.toSignal(String.Format("out_{0}_{1}", i, j));
this.neuron_outputs[arrCounter] = currentNeuronOutput;
Port tempOut = n.neuronOutput.copy();
tempOut.rename(String.Format("in_{0}", j));
tempOut.setDirection(Port.portDirection.IN);
currentPorts.Add(tempOut);
arrCounter++;
}
}
else if (this.activationTypes[i - 1] == AsyncNeuron.NeuronActivationType.LINEAR)
{
for (int j = 0; j < this.neuronCounts[i]; j++)
{
AsyncNeuron n = new AsyncNeuron(lastPorts, this.intBusWidths[neuronStart], this.numFracBits, this.numWeightUpperBits, AsyncNeuron.NeuronActivationType.LINEAR, String.Format("linear_neuron_{0}_{1}", i, 0));
neuronStart++;
if (j == 0)
{
if (!this.uniqueNeuronEntities.Contains(n))
{
this.uniqueNeuronEntities.Add(n);
}
}
this.neuronEntities[arrCounter] = n;
Signal currentNeuronOutput = n.neuronOutput.toSignal(String.Format("out_{0}_{1}", i, j));
this.neuron_outputs[arrCounter] = currentNeuronOutput;
Port tempOut = n.neuronOutput.copy();
tempOut.rename(String.Format("in_{0}", j));
tempOut.setDirection(Port.portDirection.IN);
currentPorts.Add(tempOut);
arrCounter++;
}
}
}
/*Initialize outputs*/
this.nnOutputPorts = new Port[this.neuronCounts[this.neuronCounts.Length - 1]];
if (this.isClassifier)
{
for (int i = 0; i < this.neuronCounts[this.neuronCounts.Length - 1]; i++)
{
this.nnOutputPorts[i] = new Port(Port.portDirection.OUT, String.Format("out_{0}", i), Utilities.VHDLDataType.STD_LOGIC, 0, 0);
}
}
else
{
neuronStart -= this.neuronCounts[this.neuronCounts.Length - 1];
for (int i = 0; i < this.neuronCounts[this.neuronCounts.Length - 1]; i++)
{
this.nnOutputPorts[i] = new Port(Port.portDirection.OUT, String.Format("out{0}", i), Utilities.VHDLDataType.SIGNED_FIXED_POINT, this.intBusWidths[neuronStart], this.numFracBits);
neuronStart++;
}
}
/*Initialize other signals*/
this.ready = new Port(Port.portDirection.OUT, "ready", Utilities.VHDLDataType.STD_LOGIC, 0, 0);
this.reset = new Port(Port.portDirection.IN, "reset", Utilities.VHDLDataType.STD_LOGIC, 0, 0);
this.clk = new Port(Port.portDirection.IN, "clk", Utilities.VHDLDataType.STD_LOGIC, 0, 0);
return;
}
