public BPResponse BP(BPRequest bpr)
{
Utility.Verify(() => bpr!=null,"BP bpr");
Stopwatch StopWatch=new Stopwatch();
StopWatch.Start();
bool trace=this.Inspector!=null;
if(trace)
Inspector.Trace("Backpropagation start, MaxEpochs={0}, MinEpochs={1}",bpr.MinEpochs.ToString(),bpr.MaxEpochs.ToString());
BPResponseEstimate estimate= BPResponseEstimate.Unknown;
NeuronBase[] output=Structure.Output;
List<BPEpochResponse> bper=new List<BPEpochResponse>(bpr.MaxEpochs);
for(int i=0;i<bpr.MaxEpochs;i++) {
bool succ=true;
if(trace)
Inspector.Trace("Backpropagation- start epoch {0}",(i+1).ToString());
TrainingData[] ts= bpr.GetTrainingData().ToArray();
List<double> errors=new List<double>(ts.Length);
for(int j=0; j<ts.Length; j++){
bool testPass=true;
if(ts[j].TestingSets!=null && ts[j].TestingSets.Any(x => GetPulseError(x,ts[j].Output)>bpr.MaxMSE))
testPass=false;
double mse=GetPulseError(ts[j].Input,ts[j].Output);
errors.Add(mse);
if(mse>bpr.MaxMSE || !testPass) {
for(var k=0;k!=output.Length;k++)
((Neuron)output[k]).BP(ts[j].Output[k]-output[k].LastOutputSignal,LearnFactor,Momentum);
succ=false;
}
}
double eps=GetTotalEps(errors.ToArray());
if(trace)
Inspector.Trace("End of epoch {0}, total Eps={1}",(i+1).ToString(),eps);
bper.Add(new BPEpochResponse(errors.ToArray(),eps));
if(i>=bpr.MinEpochs && succ) {
estimate=BPResponseEstimate.Success;
break;
}
if(i>=bpr.MaxEpochs-1){
if(!succ)
estimate=BPResponseEstimate.Failure;
break;
}
}
StopWatch.Stop();
BPResponse response=new BPResponse(estimate,bper.ToArray(),StopWatch.Elapsed);
return response;
}
//util
void StartLearning()
{
Stopped = false;
Thread.CurrentThread.Priority = ThreadPriority.Lowest;
Image[] bigimgs = Utility.Generate(i => Image.FromFile(@"mnist_train" + i.ToString() + ".jpg"), 10).ToArray();
int idx = 0, total = 0;
List <TrainingData> trainingList = new List <TrainingData>();
foreach (double[] input in GetAllImages(bigimgs))
{
if (Stopped)
{
return;
}
double[] output = Digitalizer.GetOutput(10, idx).ToArray();
Application.DoEvents();
trainingList.Add(new TrainingData(input, output));
GC.Collect();
GC.WaitForPendingFinalizers();
idx = (idx + 1) % 10;
if (idx == 0)
{
total++;
BPRequest request = new BPRequest(trainingList.ToArray(), 1)
{
ShuffleTrainingSet = true
};
BPResponse response = Network.BP(request);
double eps = response.Epochs[0].Epsilon;
double ms = response.BPTime.TotalMilliseconds;
trainingList.Clear();
}
}
StopLearning();
}
//util
void StartLearning()
{
Stopped=false;
button5.Text="Stop";
Log.Clear();
Thread.CurrentThread.Priority=ThreadPriority.Lowest;
Image[] bigimgs=Utility.Generate(i => Image.FromFile(@"mnist_train"+i.ToString()+".jpg"),10).ToArray();
int idx=0,total=0;
List<TrainingData> trainingList=new List<TrainingData>();
foreach(double[] input in GetAllImages(bigimgs)) {
if(Stopped)
return;
double[] output=Digitalizer.GetOutput(10,idx).ToArray();
DrawDigitalized(input,LearnDigitBox);
Application.DoEvents();
trainingList.Add(new TrainingData(input,output));
GC.Collect();
GC.WaitForPendingFinalizers();
idx= (idx+1)%10;
if(idx==0) {
total++;
BPRequest request=new BPRequest(trainingList.ToArray(),1) { ShuffleTrainingSet=true };
BPResponse response= Network.BP(request);
double eps=response.Epochs[0].Epsilon;
double ms=response.BPTime.TotalMilliseconds;
Log.Text+=string.Format("Process cycle {0}, error={1}, time[ms]={2}\r\n",total.ToString(),eps.ToString("0.000"),ms.ToString("0.000"));
trainingList.Clear();
}
}
StopLearning();
}
