private static void Train()
{
var load = Tools.cifarTrainingData();
var dd = new DataToSave()
{
Labels = load.Item2, Data = load.Item1
};
string[] classes_txt = { "airplane", "car", "bird", "cat", "deer", "dog", "frog", "horse", "ship", "truck" };
var image_dimension = 32;
var image_channels = 3;
var random_flip = true;
var random_position = true;
Network net = Tools.CreateNewNetwork();
Trainer trainer = new Trainer();
trainer.Net = net;
trainer.method = TrainingMethod.adagrad;
trainer.batch_size = 200;
trainer.l2_decay = 0.0001f;
trainer.momentum = 0.9f;
trainer.learning_rate = 0.001f;
Random random = new Random();
var data = dd.Data;
var rp = Tools.randperm(data.Length);
float[] norm = new float[] { 123.68f, 116.779f, 103.939f };
for (int ep = 0; ep < 40; ep++)
{
int all = 0;
int good = 0;
for (int we = 0; we < data.Length; we++)
{
int i = rp[we];
var y = (int)dd.Labels[i];
var p = data[i];
var x = new Vol(32, 32, 3, 0.0f);
var W = image_dimension * image_dimension;
var ii = 0;
for (var dc = 0; dc < image_channels; dc++)
{
for (var xc = 0; xc < image_dimension; xc++)
{
for (var yc = 0; yc < image_dimension; yc++)
{
x.Set(yc, xc, dc, p[ii] - norm[dc]);
ii++;
}
}
}
if (random_position)
{
var dx = Math.Floor(random.NextDouble() * 5 - 2);
var dy = Math.Floor(random.NextDouble() * 5 - 2);
x = Vol.Augment(x, image_dimension, (int)dx, (int)dy, false); //maybe change position
}
if (random_flip)
{
x = Vol.Augment(x, image_dimension, 0, 0, random.NextDouble() < 0.5); //maybe flip horizontally
}
// train on it with network
var stats = trainer.Train(x, y);
var lossx = stats.cost_loss;
var lossw = stats.l2_decay_loss;
// keep track of stats such as the average training error and loss
var yhat = net.GetPrediction();
var train_acc = yhat == y ? 1 : 0;
all++;
good += train_acc;
var acc = (float)good * 100.0 / (float)all;
Console.WriteLine("It :{2} , Loss :{0} , Acc:{1}", lossx, acc, we);
}
FileStream fs2 = new FileStream("network.dat", FileMode.OpenOrCreate, FileAccess.ReadWrite);
BinaryFormatter bf2 = new BinaryFormatter();
bf2.Serialize(fs2, net);
fs2.Close();
}
Console.Read();
}
private static void Test()
{
var load = Tools.cifarTestData();
var dd = new DataToSave()
{
Labels = load.Item2, Data = load.Item1
};
string[] classes_txt = { "airplane", "car", "bird", "cat", "deer", "dog", "frog", "horse", "ship", "truck" };
var image_dimension = 32;
var image_channels = 3;
var random_flip = true;
var random_position = true;
FileStream fs = new FileStream("network.dat", FileMode.Open, FileAccess.Read);
BinaryFormatter bf = new BinaryFormatter();
var net = bf.Deserialize(fs) as Network;
fs.Close();
Trainer trainer = new Trainer();
trainer.Net = net;
trainer.method = TrainingMethod.adadelta;
trainer.batch_size = 4;
trainer.l2_decay = 0;
trainer.momentum = 0.99f;
trainer.learning_rate = 0.01f;
Random random = new Random();
var data = dd.Data;
var rp = Tools.randperm(data.Length);
int all = 0;
int good = 0;
for (int we = 0; we < data.Length; we++)
{
int i = we;
var y = (int)dd.Labels[i];
var p = data[i];
var x = new Vol(32, 32, 3, 0.0f);
var W = image_dimension * image_dimension;
var ii = 0;
for (var dc = 0; dc < image_channels; dc++)
{
for (var xc = 0; xc < image_dimension; xc++)
{
for (var yc = 0; yc < image_dimension; yc++)
{
x.Set(yc, xc, dc, p[ii] / 255.0f - 0.5f);// ;
ii++;
}
}
}
if (random_position)
{
var dx = Math.Floor(random.NextDouble() * 5 - 2);
var dy = Math.Floor(random.NextDouble() * 5 - 2);
x = Vol.Augment(x, image_dimension, (int)dx, (int)dy, false); //maybe change position
}
if (random_flip)
{
x = Vol.Augment(x, image_dimension, 0, 0, random.NextDouble() < 0.5); //maybe flip horizontally
}
// train on it with network
var stats = net.Forward(x, false);
// keep track of stats such as the average training error and loss
var yhat = net.GetPrediction();
var train_acc = yhat == y ? 1 : 0;
all++;
good += train_acc;
var acc = (float)good * 100.0 / (float)all;
Console.WriteLine("It :{2} , Loss :{0} , Acc:{1}", 0, acc, we);
}
Console.Read();
}