static void Train(Iterator train_iterator, MnistLoader loader, VariableField x, VariableField t, StoreField accnode, StoreField lossnode, Flow trainflow, Parameters parameters)
{
Stopwatch sw = new Stopwatch();
sw.Start();
while (train_iterator.Epoch < 50)
{
(float[] images, float[] labels) = loader.GetTrain(train_iterator.Next());
x.ValueTensor.State = images;
t.ValueTensor.State = labels;
trainflow.Execute(enable_processing_time: true);
parameters.Update();
if (train_iterator.Iteration % 100 != 0)
{
continue;
}
float train_acc = accnode.State[0];
float train_loss = lossnode.State[0];
Console.WriteLine($"[{train_iterator.Iteration}] train acc: {train_acc:F3} train loss: {train_loss:E3}");
}
sw.Stop();
Console.WriteLine($"{sw.ElapsedMilliseconds} msec, {sw.ElapsedMilliseconds / train_iterator.Epoch} msec/epoch");
}
static void Test(Iterator test_iterator, MnistLoader loader, VariableField x, VariableField t, Flow testflow, StoreField accnode)
{
List <float> test_acc_list = new List <float>();
Stopwatch sw = new Stopwatch();
sw.Start();
while (test_iterator.Epoch < 1)
{
(float[] images, float[] labels) = loader.GetTest(test_iterator.Next());
x.ValueTensor.State = images;
t.ValueTensor.State = labels;
testflow.Execute();
float test_acc = accnode.State[0];
test_acc_list.Add(test_acc);
}
sw.Stop();
Console.WriteLine($"{sw.ElapsedMilliseconds} msec");
Console.WriteLine($"[Result] acc: {test_acc_list.Average():F5}");
}
private static void Main(string[] args)
{
Console.Clear();
var watch = Stopwatch.StartNew();
ProcessingDevice.Device = DeviceType.CPU_Parallel;
watch.Stop();
Console.WriteLine($"Device Time: {watch.ElapsedMilliseconds}ms");
var model = new DenseModel();
model.AddLayer(BuildedModels.DenseLeakRelu(784, 100, 1e-4f, 2e-1f, EnumOptimizerFunction.RmsProp));
model.AddLayer(BuildedModels.DenseLeakRelu(100, 30, 1e-4f, 2e-1f, EnumOptimizerFunction.RmsProp));
model.AddLayer(BuildedModels.DenseSoftMax(30, 10, 1e-4f, 2e-1f, EnumOptimizerFunction.RmsProp));
model.SetLossFunction(new CrossEntropyLossFunction());
watch = Stopwatch.StartNew();
watch.Stop();
Console.WriteLine($"Sinapse Time: {watch.ElapsedMilliseconds}ms");
MnistLoader.DataPath = path;
var trainingValues = MnistLoader.OpenMnist();
int cont = 0;
int sizeTrain = trainingValues.Count;
var err = 100f;
var e = 0f;
while (true)
{
watch = Stopwatch.StartNew();
e = 0f;
int ct = 0;
for (int i = 0; i < sizeTrain; i++)
{
var index = i;
var inputs = new FloatArray(ArrayMethods.ByteToArray(trainingValues[index].pixels, 28, 28));
var target = new FloatArray(ArrayMethods.ByteToArray(trainingValues[index].label, 10));
// Learning
e += model.Learn(inputs, target);
// Sample
if (ct % 1000 == 0)
{
Write(trainingValues[index].ToString());
Write(ArrayMethods.PrintArray(model.Output(inputs), 10));
Write("\n");
}
ct++;
}
err = 0.999f * err + 0.001f * e;
cont++;
watch.Stop();
var time = watch.ElapsedMilliseconds;
Console.Title =
$"Error: {err} --- TSPS (Training Sample per Second): {Math.Ceiling(1000d / ((double)time / (double)sizeTrain))}";
}
}
static void Main()
{
const string dirpath_dataset = "mnist_dataset";
const string dirpath_result = "result";
const int classes = 10;
Console.WriteLine("Download mnist...");
MnistDownloader.Download(dirpath_dataset);
Console.WriteLine("Setup loader...");
Random random = new Random(1234);
MnistLoader loader = new MnistLoader(dirpath_dataset, num_batches: 1000);
Iterator train_iterator = new ShuffleIterator(loader.NumBatches, loader.CountTrainDatas, random);
Iterator test_iterator = new ShuffleIterator(loader.NumBatches, loader.CountTestDatas, random);
Console.WriteLine("Create input tensor...");
VariableField x = new Tensor(loader.BatchShape);
VariableField t = new Tensor(Shape.Vector(loader.NumBatches));
Console.WriteLine("Build model...");
Field y = CNN.Forward(x, classes);
Field acc = Accuracy(y, t);
Field err = Sum(SoftmaxCrossEntropy(y, OneHotVector(t, classes)), axes: new int[] { Axis.Map0D.Channels });
StoreField accnode = acc.Save(), lossnode = Average(err).Save();
Console.WriteLine("Build optimize flow...");
(Flow trainflow, Parameters parameters) = Flow.Optimize(err);
Console.WriteLine("Initialize params...");
parameters
.Where((parameter) => parameter.Category == ParameterCategory.Kernel)
.InitializeTensor((tensor) => new HeNormal(tensor, random));
parameters
.Where((parameter) => parameter.Category == ParameterCategory.Bias)
.InitializeTensor((tensor) => new Zero(tensor));
Console.WriteLine("Set params updater...");
parameters.AddUpdater((parameter) => new Nadam(parameter, alpha: 0.01f));
parameters.AddUpdater((parameter) => new Ridge(parameter, decay: 1e-4f));
Console.WriteLine("Training...");
Train(train_iterator, loader, x, t, accnode, lossnode, trainflow, parameters);
Console.WriteLine("Build inference flow...");
Flow testflow = Flow.Inference(accnode);
Console.WriteLine("Testing...");
Test(test_iterator, loader, x, t, testflow, accnode);
Console.WriteLine("Saving snapshot...");
Snapshot snapshot = parameters.Save();
SnapshotSaver saver = new ZippedBinaryShapshotSaver();
if (!Directory.Exists(dirpath_result))
{
Directory.CreateDirectory(dirpath_result);
}
saver.Save($"{dirpath_result}/mnist.tss", snapshot);
Console.WriteLine("END");
Console.Read();
}