private void Test(Parameter parameter)
{
try
{
IList <Matrix <C> > trainingImages;
IList <T> trainingLabels;
IList <Matrix <C> > testingImages;
IList <T> testingLabels;
Logger.Info("Start load train images");
Load(parameter.Dataset, "train", out trainingImages, out trainingLabels);
Logger.Info($"Load train images: {trainingImages.Count}");
Logger.Info("Start load test images");
Load(parameter.Dataset, "test", out testingImages, out testingLabels);
Logger.Info($"Load test images: {testingImages.Count}");
Logger.Info("");
// So with that out of the way, we can make a network instance.
var networkId = SetupNetwork();
using (var net = LossMulticlassLog.Deserialize(parameter.Model, networkId))
{
this.SetEvalMode(networkId, net);
var validationParameter = new ValidationParameter <T, C>
{
BaseName = Path.GetFileNameWithoutExtension(parameter.Model),
Trainer = net,
TrainingImages = trainingImages,
TrainingLabels = trainingLabels,
TestingImages = testingImages,
TestingLabels = testingLabels,
UseConsole = true,
SaveToXml = false,
OutputDiffLog = true,
Output = Path.GetDirectoryName(parameter.Model)
};
Validation(validationParameter, out _, out _);
}
}
catch (Exception e)
{
Logger.Error(e.Message);
}
}
private void Validation(ValidationParameter <T, C> parameter,
out double trainAccuracy,
out double testAccuracy)
{
trainAccuracy = 0;
testAccuracy = 0;
var net = parameter.Trainer;
var trainingImages = parameter.TrainingImages;
var trainingLabels = parameter.TrainingLabels;
var testingImages = parameter.TestingImages;
var testingLabels = parameter.TestingLabels;
var saveToXml = parameter.SaveToXml;
var baseName = parameter.BaseName;
var useConsole = parameter.UseConsole;
using (var predictedLabels = net.Operator(trainingImages))
{
var numRight = 0;
var numWrong = 0;
// And then let's see if it classified them correctly.
for (var i = 0; i < trainingImages.Count; ++i)
{
if (this.Compare(predictedLabels[i], trainingLabels[i]))
{
++numRight;
}
else
{
++numWrong;
}
}
if (useConsole)
{
Logger.Info($"training num_right: {numRight}");
Logger.Info($"training num_wrong: {numWrong}");
Logger.Info($"training accuracy: {numRight / (double)(numRight + numWrong)}");
}
trainAccuracy = numRight / (double)(numRight + numWrong);
using (var predictedLabels2 = net.Operator(testingImages))
{
numRight = 0;
numWrong = 0;
for (var i = 0; i < testingImages.Count; ++i)
{
if (this.Compare(predictedLabels2[i], testingLabels[i]))
{
++numRight;
}
else
{
++numWrong;
}
}
if (useConsole)
{
Logger.Info($"testing num_right: {numRight}");
Logger.Info($"testing num_wrong: {numWrong}");
Logger.Info($"testing accuracy: {numRight / (double)(numRight + numWrong)}");
}
testAccuracy = numRight / (double)(numRight + numWrong);
// Finally, you can also save network parameters to XML files if you want to do
// something with the network in another tool. For example, you could use dlib's
// tools/convert_dlib_nets_to_caffe to convert the network to a caffe model.
if (saveToXml)
{
Dlib.NetToXml(net, $"{baseName}.xml");
}
}
}
}
private void Train(Parameter parameter)
{
try
{
IList <Matrix <C> > trainingImages;
IList <T> trainingLabels;
IList <Matrix <C> > testingImages;
IList <T> testingLabels;
Logger.Info("Start load train images");
Load(parameter.Dataset, "train", out trainingImages, out trainingLabels);
Logger.Info($"Load train images: {trainingImages.Count}");
Logger.Info("Start load test images");
Load(parameter.Dataset, "test", out testingImages, out testingLabels);
Logger.Info($"Load test images: {testingImages.Count}");
Logger.Info("");
// So with that out of the way, we can make a network instance.
var networkId = SetupNetwork();
using (var net = new LossMulticlassLog(networkId))
using (var solver = new Adam())
using (var trainer = new DnnTrainer <LossMulticlassLog>(net, solver))
{
var learningRate = parameter.LearningRate;
var minLearningRate = parameter.MinLearningRate;
var miniBatchSize = parameter.MiniBatchSize;
var baseName = parameter.BaseName;
var epoch = parameter.Epoch;
var validation = parameter.Validation;
trainer.SetLearningRate(learningRate);
trainer.SetMinLearningRate(minLearningRate);
trainer.SetMiniBatchSize(miniBatchSize);
trainer.BeVerbose();
trainer.SetSynchronizationFile(baseName, 180);
// create array box
var trainingImagesCount = trainingImages.Count;
var trainingLabelsCount = trainingLabels.Count;
var maxIteration = (int)Math.Ceiling(trainingImagesCount / (float)miniBatchSize);
var imageBatches = new Matrix <C> [maxIteration][];
var labelBatches = new uint[maxIteration][];
for (var i = 0; i < maxIteration; i++)
{
if (miniBatchSize <= trainingImagesCount - i * miniBatchSize)
{
imageBatches[i] = new Matrix <C> [miniBatchSize];
labelBatches[i] = new uint[miniBatchSize];
}
else
{
imageBatches[i] = new Matrix <C> [trainingImagesCount % miniBatchSize];
labelBatches[i] = new uint[trainingLabelsCount % miniBatchSize];
}
}
using (var fs = new FileStream($"{baseName}.log", FileMode.Create, FileAccess.Write, FileShare.Write))
using (var sw = new StreamWriter(fs, Encoding.UTF8))
for (var e = 0; e < epoch; e++)
{
var randomArray = Enumerable.Range(0, trainingImagesCount).OrderBy(i => Guid.NewGuid()).ToArray();
var index = 0;
for (var i = 0; i < imageBatches.Length; i++)
{
var currentImages = imageBatches[i];
var currentLabels = labelBatches[i];
for (var j = 0; j < imageBatches[i].Length; j++)
{
var rIndex = randomArray[index];
currentImages[j] = trainingImages[rIndex];
currentLabels[j] = this.Cast(trainingLabels[rIndex]);
index++;
}
}
for (var i = 0; i < maxIteration; i++)
{
LossMulticlassLog.TrainOneStep(trainer, imageBatches[i], labelBatches[i]);
}
var lr = trainer.GetLearningRate();
var loss = trainer.GetAverageLoss();
var trainLog = $"Epoch: {e}, learning Rate: {lr}, average loss: {loss}";
Logger.Info(trainLog);
sw.WriteLine(trainLog);
if (e >= 0 && e % validation == 0)
{
var validationParameter = new ValidationParameter <T, C>
{
BaseName = parameter.BaseName,
Output = parameter.Output,
Trainer = net,
TrainingImages = trainingImages,
TrainingLabels = trainingLabels,
TestingImages = testingImages,
TestingLabels = testingLabels,
UseConsole = true,
SaveToXml = true,
OutputDiffLog = true
};
Validation(validationParameter, out var trainAccuracy, out var testAccuracy);
var validationLog = $"Epoch: {e}, train accuracy: {trainAccuracy}, test accuracy: {testAccuracy}";
Logger.Info(validationLog);
sw.WriteLine(validationLog);
var name = this.GetBaseName(parameter.Epoch,
parameter.LearningRate,
parameter.MinLearningRate,
parameter.MiniBatchSize);
UpdateBestModelFile(net, testAccuracy, parameter.Output, name, "test");
UpdateBestModelFile(net, trainAccuracy, parameter.Output, name, "train");
}
if (lr < minLearningRate)
{
Logger.Info($"Stop training: {lr} < {minLearningRate}");
break;
}
}
// wait for training threads to stop
trainer.GetNet();
Logger.Info("done training");
net.Clean();
LossMulticlassLog.Serialize(net, $"{baseName}.tmp");
// Now let's run the training images through the network. This statement runs all the
// images through it and asks the loss layer to convert the network's raw output into
// labels. In our case, these labels are the numbers between 0 and 9.
var validationParameter2 = new ValidationParameter <T, C>
{
BaseName = parameter.BaseName,
Output = parameter.Output,
Trainer = net,
TrainingImages = trainingImages,
TrainingLabels = trainingLabels,
TestingImages = testingImages,
TestingLabels = testingLabels,
UseConsole = true,
SaveToXml = true,
OutputDiffLog = true
};
Validation(validationParameter2, out _, out _);
// clean up tmp files
Clean(parameter.Output);
}
}
catch (Exception e)
{
Logger.Error(e.Message);
}
}
