private void настроитьПредобучениеСлояToolStripMenuItem_Click(object sender, EventArgs e)
{
if (listBox2.SelectedIndex != -1)
{
pretrain = new FullConPretrain(ref NetworkData.network);
pretrain.SetupPretrain(listBox2.SelectedIndex);
pretrained_layer = listBox2.SelectedIndex;
timer2.Start();
}
}
public PretrainMaster()
{
InitializeComponent();
Network network = NetworkData.network;
for (int i = 0; i < network.layers.Count; i++)
{
listBox1.Items.Add(network.layers[i].GetType().Name + " {" + network.layers[i].output_size[0] + ","
+ network.layers[i].output_size[1] + "," + network.layers[i].output_size[2] + "}");
}
test_arr = NetworkData.train.GetRandom(ref network).Key;
base_arr = test_arr;
testImg = ImageDataConverter.GetImage(test_arr.ConvertToRGD());
//pictureBox1.Image = testImg;
pretrain = new FullConPretrain(ref NetworkData.network);
}
private void timer2_Tick(object sender, EventArgs e)
{
if (train)
{
Update();
pretrain_index++;
if (pretrain_index % 20 == 0)
{
NetworkData.SaveNetwork("autosave_pretrain");
pretrain.SaveAutoencoderToFile(pretrained_layer + "_autosave_autoencoder.neural");
if (NetworkData.val != null)
{
var val_pair = pretrain.Validation(NetworkData.val);
if (min_err > val_pair.Key)
{
listBox1.Items.Insert(0, "New minimum! Save at min.neural");
NetworkData.SaveNetwork("min");
pretrain.SaveAutoencoderToFile(pretrained_layer + "_autoencoder_min.neural");
}
min_err = Math.Min(min_err, val_pair.Key);
listBox1.Items.Insert(0, pretrained_layer + " :Val_err: " + (float)val_pair.Key);
listBox1.Items.Insert(0, pretrained_layer + " :Val_win: " + (float)val_pair.Value);
if (val_pair.Key > min_err)
{
bad_val_count++;
listBox1.Items.Insert(0, "Bad validation " + bad_val_count);
if (bad_val_count > 12)
{
listBox1.Items.Insert(0, "Stop train. Bad validation");
try
{
listBox1.Items.Insert(0, "Загружаю наилучшую сеть по валидации");
NetworkData.LoadNetwork("min.neural");
}
catch
{
listBox1.Items.Insert(0, "Ошибка загрузки!");
}
if (pretrained_layer < NetworkData.network.layers.Count - 2)
{
listBox1.Items.Insert(0, "Начинаю предобучение следующего слоя");
min_err = double.MaxValue;
pretrained_layer++;
while (!NetworkData.network.layers[pretrained_layer].ITrained && pretrained_layer < NetworkData.network.layers.Count - 1)
{
pretrained_layer++;
}
if (pretrained_layer < NetworkData.network.layers.Count - 1)
{
listBox1.Items.Insert(0, "Начинаю предобучение следующего слоя");
NetworkData.network.LoadJSON(File.ReadAllText("min.neural"));
NetworkData.network.CompileOnlyError();
pretrain = new FullConPretrain(ref NetworkData.network);
pretrain.SetupPretrain(pretrained_layer);
pret_task = Task <double> .Run(() => pretrain.TrainStart(NetworkData.optimizer, NetworkData.train, pretrin_batch));
}
}
else
{
pret_task = null;
return;
}
}
}
else
{
bad_val_count = 0;
}
}
}
//var res = NetworkData.network.GetOutput(test_arr);
var err = pretrain.TrainContinue(NetworkData.optimizer, NetworkData.train, pretrin_batch);
//listBox1.Items.Insert(0, test_res[0, 0, 0] + " " + test_res[0, 0, 1]);
//listBox1.Items.Insert(0, (float)res[0, 0, 0] + " " + (float)res[0, 0, 1]);
listBox1.Items.Insert(0, pretrained_layer + " : " + (float)err);
if (listBox1.Items.Count > 1000)
{
listBox1.Items.RemoveAt(listBox1.Items.Count - 1);
listBox1.Items.RemoveAt(listBox1.Items.Count - 1);
listBox1.Items.RemoveAt(listBox1.Items.Count - 1);
listBox1.Items.RemoveAt(listBox1.Items.Count - 1);
}
Update();
}
}
