NeuralNet
Object Pascal feed forward neural network component
- Active · last change 5-22-2025
- ★ 1 on GitHub
README
NeuralNet
Object Pascal feed forward neural network component
Simple 3 layer feed-forward neural network component
No dependencies
Can be used in conjunction with DWS compilers
demo1: https://lynkfs.design/QTX_Neural/
(compiled in QTX-IDE)
this demonstrates handwritten character recognition, based on the MNIST dataset
The net is trained with only 1000 examples to limit training time
(less than 30 sec on desktop, approx 4 min on mobile)
which means it is not perfect.
demo2: simple xor
//Init neural network
var MyNetwork : JW3NeuralNet := JW3NeuralNet.Create;
//
MyNetwork.AddLayer (2, 'Input', ''); // input layer. input layers have no activation function
MyNetwork.AddLayer (5, 'Hidden', 'Sigmoid'); // first hidden layer, with sigmoid activation
MyNetwork.AddLayer (2, 'Output', 'Sigmoid'); // output layer, with sigmoid activation
//
MyNetwork.SeedNetwork; //initialise with random values
// provide some input (XOR function)
for var iter := 0 to 1000 do
for var x := 0 to 1 do
for var y := 0 to 1 do
begin
var data: JW3TrainingRecord;
data.&inputs := [x, y];
if (x=0) and (y=0) then data.&outputs := [1.0,0.0];
if (x=0) and (y=1) then data.&outputs := [0.0,1.0];
if (x=1) and (y=0) then data.&outputs := [0.0,1.0];
if (x=1) and (y=1) then data.&outputs := [1.0,0.0];
MyNetwork.AddTrainingData(data.&inputs, data.&outputs);
end;
// train
MyNetwork.LearningRate := 0.2;
MyNetwork.TrainingSplit := 10; //split trainingset randomly: 10% for testing, 90% training
For var i := 1 to 5 do begin //every iteration is an epoch
//console.log('Epoch ' + inttostr(i) + '/5');
MyNetwork.Train;
end;
// test the 4 cases
console.log('0 xor 0 : ' + floattostr(MyNetwork.Test([0,0])) + ' - should be 0');
console.log('0 xor 1 : ' + floattostr(MyNetwork.Test([0,1])) + ' - should be 1');
console.log('1 xor 0 : ' + floattostr(MyNetwork.Test([1,0])) + ' - should be 1');
console.log('1 xor 1 : ' + floattostr(MyNetwork.Test([1,1])) + ' - should be 0');output
NeuralNet started
0 XOR 0: 0 - should be 0
1 XOR 0: 1 - should be 1
0 XOR 1: 1 - should be 1
1 XOR 1: 0 - should be 0
demo3: is it a bird, is it a plane, is it a rocket (input : wings, engine)
//Init neural network
MyNetwork := JW3NeuralNet.Create;
//
MyNetwork.AddLayer (2, 'Input', ''); // input layer. input layers have no activation function
MyNetwork.AddLayer (9, 'Hidden', 'Sigmoid'); // first hidden layer, with sigmoid activation
MyNetwork.AddLayer (3, 'Output', 'Sigmoid'); // output layer, with sigmoid activation
//
MyNetwork.SeedNetwork; //initialise with random values
//
//feed some examples. input: wings,engine output: bird,plane,rocket
for var iter := 1 to 10 do //just 10 examples
for var x := 0 to 1 do
for var y := 0 to 1 do
begin
var data: JW3TrainingRecord;
data.&inputs := [x, y];
if (x=1) and (y=0) then data.&outputs := [1.0,0.0,0.0]; //wings no engine : bird
if (x=1) and (y=1) then data.&outputs := [0.0,1.0,0.0]; //both wings and engine : plane
if (x=0) and (y=1) then data.&outputs := [0.0,0.0,1.0]; //no wings, engine : rocket
if (x=0) and (y=0) then //nothing : do nothing
else MyNetwork.AddTrainingData(data.&inputs, data.&outputs);
end;
//
//train network
MyNetwork.LearningRate := 0.2; //MyNetwork.TrainingSplit not necessary
For var i := 0 to 5 do //5 epochs
MyNetwork.Train;
//
//test
Button1.OnClick := procedure(sender: TObject)
begin
var i : integer := 0;
var j : integer := 0;
if CheckBox1.Checked Then i := 1;
if CheckBox2.Checked then j := 1;
if (i = 0) and (j = 0)
then Panel1.SetinnerHTML('unknown')
else begin
//run
var predictedIndex := MyNetwork.Test([i,j]);
case predictedIndex of
0 : Panel1.SetinnerHTML('bird');
1 : Panel1.SetinnerHTML('plane');
2 : Panel1.SetinnerHTML('rocket');
end;
end;
end;