public static void Run()
{
//MNISTのデータを用意する
Console.WriteLine("MNIST Data Loading...");
MnistData mnistData = new MnistData();
Console.WriteLine("Training Start...");
//ネットワークの構成を FunctionStack に書き連ねる
FunctionStack nn = new FunctionStack(
new Linear(28 * 28, 1024, name: "l1 Linear"),
new Sigmoid(name: "l1 Sigmoid"),
new Linear(1024, 10, name: "l2 Linear")
);
//optimizerを宣言
nn.SetOptimizer(new MomentumSGD());
//三世代学習
for (int epoch = 0; epoch < 3; epoch++)
{
Console.WriteLine("epoch " + (epoch + 1));
//全体での誤差を集計
Real totalLoss = 0;
long totalLossCount = 0;
//何回バッチを実行するか
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
//訓練データからランダムにデータを取得
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT);
//バッチ学習を並列実行する
Real sumLoss = Trainer.Train(nn, datasetX.Data, datasetX.Label, new SoftmaxCrossEntropy());
totalLoss = sumLoss;
totalLossCount++;
//20回バッチを動かしたら精度をテストする
if (i % 20 == 0)
{
Console.WriteLine("\nbatch count " + i + "/" + TRAIN_DATA_COUNT);
//結果出力
Console.WriteLine("total loss " + totalLoss / totalLossCount);
Console.WriteLine("local loss " + sumLoss);
Console.WriteLine("\nTesting...");
//テストデータからランダムにデータを取得
TestDataSet datasetY = mnistData.GetRandomYSet(TEST_DATA_COUNT);
//テストを実行
Real accuracy = Trainer.Accuracy(nn, datasetY.Data, datasetY.Label);
Console.WriteLine("accuracy " + accuracy);
}
}
}
}
public static void Run()
{
Stopwatch sw = new Stopwatch();
//MNISTのデータを用意する
Console.WriteLine("MNIST Data Loading...");
MnistData mnistData = new MnistData();
//ネットワークの構成を FunctionStack に書き連ねる
FunctionStack nn = new FunctionStack(
new Convolution2D(1, 32, 5, pad: 2, name: "l1 Conv2D", activation: new ReLU(name: "l1 ReLU"), gpuEnable: true),
new MaxPooling(2, 2, name: "l1 MaxPooling", gpuEnable: true),
new Convolution2D(32, 64, 5, pad: 2, name: "l2 Conv2D", activation: new ReLU(name: "l2 ReLU"), gpuEnable: true),
new MaxPooling(2, 2, name: "l2 MaxPooling", gpuEnable: true),
new Linear(7 * 7 * 64, 1024, name: "l3 Linear", activation: new ReLU(name: "l3 ReLU"), gpuEnable: true),
new Dropout(name: "l3 DropOut"),
new Linear(1024, 10, name: "l4 Linear", gpuEnable: true)
);
//optimizerを宣言
nn.SetOptimizer(new Adam());
Console.WriteLine("Training Start...");
//三世代学習
for (int epoch = 1; epoch < 3; epoch++)
{
Console.WriteLine("epoch " + epoch);
//全体での誤差を集計
Real totalLoss = 0;
long totalLossCount = 0;
//何回バッチを実行するか
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
sw.Restart();
Console.WriteLine("\nbatch count " + i + "/" + TRAIN_DATA_COUNT);
//訓練データからランダムにデータを取得
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT);
//バッチ学習を並列実行する
Real sumLoss = Trainer.Train(nn, datasetX.Data, datasetX.Label, new SoftmaxCrossEntropy());
totalLoss += sumLoss;
totalLossCount++;
//結果出力
Console.WriteLine("total loss " + totalLoss / totalLossCount);
Console.WriteLine("local loss " + sumLoss);
sw.Stop();
Console.WriteLine("time" + sw.Elapsed.TotalMilliseconds);
//20回バッチを動かしたら精度をテストする
if (i % 20 == 0)
{
Console.WriteLine("\nTesting...");
//テストデータからランダムにデータを取得
TestDataSet datasetY = mnistData.GetRandomYSet(TEACH_DATA_COUNT);
//テストを実行
Real accuracy = Trainer.Accuracy(nn, datasetY.Data, datasetY.Label);
Console.WriteLine("accuracy " + accuracy);
}
}
}
}
const int N = 30; //参考先リンクと同様の1000でも動作するがCPUでは遅いので
public static void Run()
{
//MNISTのデータを用意する
Console.WriteLine("MNIST Data Loading...");
MnistData mnistData = new MnistData();
Console.WriteLine("Training Start...");
//ネットワークの構成を FunctionStack に書き連ねる
FunctionStack nn = new FunctionStack(
new Linear(28 * 28, N, name: "l1 Linear"), // L1
new BatchNormalization(N, name: "l1 BatchNorm"),
new ReLU(name: "l1 ReLU"),
new Linear(N, N, name: "l2 Linear"), // L2
new BatchNormalization(N, name: "l2 BatchNorm"),
new ReLU(name: "l2 ReLU"),
new Linear(N, N, name: "l3 Linear"), // L3
new BatchNormalization(N, name: "l3 BatchNorm"),
new ReLU(name: "l3 ReLU"),
new Linear(N, N, name: "l4 Linear"), // L4
new BatchNormalization(N, name: "l4 BatchNorm"),
new ReLU(name: "l4 ReLU"),
new Linear(N, N, name: "l5 Linear"), // L5
new BatchNormalization(N, name: "l5 BatchNorm"),
new ReLU(name: "l5 ReLU"),
new Linear(N, N, name: "l6 Linear"), // L6
new BatchNormalization(N, name: "l6 BatchNorm"),
new ReLU(name: "l6 ReLU"),
new Linear(N, N, name: "l7 Linear"), // L7
new BatchNormalization(N, name: "l7 BatchNorm"),
new ReLU(name: "l7 ReLU"),
new Linear(N, N, name: "l8 Linear"), // L8
new BatchNormalization(N, name: "l8 BatchNorm"),
new ReLU(name: "l8 ReLU"),
new Linear(N, N, name: "l9 Linear"), // L9
new BatchNormalization(N, name: "l9 BatchNorm"),
new ReLU(name: "l9 ReLU"),
new Linear(N, N, name: "l10 Linear"), // L10
new BatchNormalization(N, name: "l10 BatchNorm"),
new ReLU(name: "l10 ReLU"),
new Linear(N, N, name: "l11 Linear"), // L11
new BatchNormalization(N, name: "l11 BatchNorm"),
new ReLU(name: "l11 ReLU"),
new Linear(N, N, name: "l12 Linear"), // L12
new BatchNormalization(N, name: "l12 BatchNorm"),
new ReLU(name: "l12 ReLU"),
new Linear(N, N, name: "l13 Linear"), // L13
new BatchNormalization(N, name: "l13 BatchNorm"),
new ReLU(name: "l13 ReLU"),
new Linear(N, N, name: "l14 Linear"), // L14
new BatchNormalization(N, name: "l14 BatchNorm"),
new ReLU(name: "l14 ReLU"),
new Linear(N, 10, name: "l15 Linear") // L15
);
//この構成では学習が進まない
//FunctionStack nn = new FunctionStack(
// new Linear(28 * 28, N), // L1
// new ReLU(),
// new Linear(N, N), // L2
// new ReLU(),
//
// (中略)
//
// new Linear(N, N), // L14
// new ReLU(),
// new Linear(N, 10) // L15
//);
//optimizerを宣言
nn.SetOptimizer(new AdaGrad());
//三世代学習
for (int epoch = 0; epoch < 3; epoch++)
{
Console.WriteLine("epoch " + (epoch + 1));
//全体での誤差を集計
//List<Real> totalLoss = new List<Real>();
Real totalLoss = 0;
long totalLossCounter = 0;
//何回バッチを実行するか
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
//訓練データからランダムにデータを取得
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT);
//学習を実行
Real sumLoss = Trainer.Train(nn, datasetX.Data, datasetX.Label, new SoftmaxCrossEntropy());
totalLoss += sumLoss;
totalLossCounter++;
//20回バッチを動かしたら精度をテストする
if (i % 20 == 0)
{
//結果出力
Console.WriteLine("\nbatch count " + i + "/" + TRAIN_DATA_COUNT);
Console.WriteLine("total loss " + totalLoss / totalLossCounter);
Console.WriteLine("local loss " + sumLoss);
Console.WriteLine("");
Console.WriteLine("Testing...");
//テストデータからランダムにデータを取得
TestDataSet datasetY = mnistData.GetRandomYSet(TEST_DATA_COUNT);
//テストを実行
Real accuracy = Trainer.Accuracy(nn, datasetY.Data, datasetY.Label);
Console.WriteLine("accuracy " + accuracy);
}
}
}
}
public static void Run()
{
//MNISTのデータを用意する
Console.WriteLine("MNIST Data Loading...");
MnistData mnistData = new MnistData();
Console.WriteLine("Training Start...");
//ネットワークの構成を FunctionStack に書き連ねる
FunctionStack Layer1 = new FunctionStack(
new Linear(28 * 28, 256, name: "l1 Linear"),
new BatchNormalization(256, name: "l1 Norm"),
new ReLU(name: "l1 ReLU")
);
FunctionStack Layer2 = new FunctionStack(
new Linear(256, 256, name: "l2 Linear"),
new BatchNormalization(256, name: "l2 Norm"),
new ReLU(name: "l2 ReLU")
);
FunctionStack Layer3 = new FunctionStack(
new Linear(256, 256, name: "l3 Linear"),
new BatchNormalization(256, name: "l3 Norm"),
new ReLU(name: "l3 ReLU")
);
FunctionStack Layer4 = new FunctionStack(
new Linear(256, 10, name: "l4 Linear")
);
//FunctionStack自身もFunctionとして積み上げられる
FunctionStack nn = new FunctionStack
(
Layer1,
Layer2,
Layer3,
Layer4
);
FunctionStack DNI1 = new FunctionStack(
new Linear(256, 1024, name: "DNI1 Linear1"),
new BatchNormalization(1024, name: "DNI1 Nrom1"),
new ReLU(name: "DNI1 ReLU1"),
new Linear(1024, 1024, name: "DNI1 Linear2"),
new BatchNormalization(1024, name: "DNI1 Nrom2"),
new ReLU(name: "DNI1 ReLU2"),
new Linear(1024, 256, initialW: new Real[1024, 256], name: "DNI1 Linear3")
);
FunctionStack DNI2 = new FunctionStack(
new Linear(256, 1024, name: "DNI2 Linear1"),
new BatchNormalization(1024, name: "DNI2 Nrom1"),
new ReLU(name: "DNI2 ReLU1"),
new Linear(1024, 1024, name: "DNI2 Linear2"),
new BatchNormalization(1024, name: "DNI2 Nrom2"),
new ReLU(name: "DNI2 ReLU2"),
new Linear(1024, 256, initialW: new Real[1024, 256], name: "DNI2 Linear3")
);
FunctionStack DNI3 = new FunctionStack(
new Linear(256, 1024, name: "DNI3 Linear1"),
new BatchNormalization(1024, name: "DNI3 Nrom1"),
new ReLU(name: "DNI3 ReLU1"),
new Linear(1024, 1024, name: "DNI3 Linear2"),
new BatchNormalization(1024, name: "DNI3 Nrom2"),
new ReLU(name: "DNI3 ReLU2"),
new Linear(1024, 256, initialW: new Real[1024, 256], name: "DNI3 Linear3")
);
//optimizerを宣言
Layer1.SetOptimizer(new Adam());
Layer2.SetOptimizer(new Adam());
Layer3.SetOptimizer(new Adam());
Layer4.SetOptimizer(new Adam());
DNI1.SetOptimizer(new Adam());
DNI2.SetOptimizer(new Adam());
DNI3.SetOptimizer(new Adam());
//三世代学習
for (int epoch = 0; epoch < 20; epoch++)
{
Console.WriteLine("epoch " + (epoch + 1));
Real totalLoss = 0;
Real DNI1totalLoss = 0;
Real DNI2totalLoss = 0;
Real DNI3totalLoss = 0;
long totalLossCount = 0;
long DNI1totalLossCount = 0;
long DNI2totalLossCount = 0;
long DNI3totalLossCount = 0;
//何回バッチを実行するか
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
//訓練データからランダムにデータを取得
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT);
//第一層を実行
NdArray[] layer1ForwardResult = Layer1.Forward(datasetX.Data);
//第一層の傾きを取得
NdArray[] DNI1Result = DNI1.Forward(layer1ForwardResult);
//第一層の傾きを適用
layer1ForwardResult[0].Grad = DNI1Result[0].Data.ToArray();
//第一層を更新
Layer1.Backward(layer1ForwardResult);
layer1ForwardResult[0].ParentFunc = null; //Backwardを実行したので計算グラフを切っておく
Layer1.Update();
//第二層を実行
NdArray[] layer2ForwardResult = Layer2.Forward(layer1ForwardResult);
//第二層の傾きを取得
NdArray[] DNI2Result = DNI2.Forward(layer2ForwardResult);
//第二層の傾きを適用
layer2ForwardResult[0].Grad = DNI2Result[0].Data.ToArray();
//第二層を更新
Layer2.Backward(layer2ForwardResult);
layer2ForwardResult[0].ParentFunc = null;
//第一層用のDNIの学習を実行
Real DNI1loss = new MeanSquaredError().Evaluate(DNI1Result, new NdArray(layer1ForwardResult[0].Grad, DNI1Result[0].Shape, DNI1Result[0].BatchCount));
Layer2.Update();
DNI1.Backward(DNI1Result);
DNI1.Update();
DNI1totalLoss += DNI1loss;
DNI1totalLossCount++;
//第三層を実行
NdArray[] layer3ForwardResult = Layer3.Forward(layer2ForwardResult);
//第三層の傾きを取得
NdArray[] DNI3Result = DNI3.Forward(layer3ForwardResult);
//第三層の傾きを適用
layer3ForwardResult[0].Grad = DNI3Result[0].Data.ToArray();
//第三層を更新
Layer3.Backward(layer3ForwardResult);
layer3ForwardResult[0].ParentFunc = null;
//第二層用のDNIの学習を実行
Real DNI2loss = new MeanSquaredError().Evaluate(DNI2Result, new NdArray(layer2ForwardResult[0].Grad, DNI2Result[0].Shape, DNI2Result[0].BatchCount));
Layer3.Update();
DNI2.Backward(DNI2Result);
DNI2.Update();
DNI2totalLoss += DNI2loss;
DNI2totalLossCount++;
//第四層を実行
NdArray[] layer4ForwardResult = Layer4.Forward(layer3ForwardResult);
//第四層の傾きを取得
Real sumLoss = new SoftmaxCrossEntropy().Evaluate(layer4ForwardResult, datasetX.Label);
//第四層を更新
Layer4.Backward(layer4ForwardResult);
layer4ForwardResult[0].ParentFunc = null;
totalLoss += sumLoss;
totalLossCount++;
//第三層用のDNIの学習を実行
Real DNI3loss = new MeanSquaredError().Evaluate(DNI3Result, new NdArray(layer3ForwardResult[0].Grad, DNI3Result[0].Shape, DNI3Result[0].BatchCount));
Layer4.Update();
DNI3.Backward(DNI3Result);
DNI3.Update();
DNI3totalLoss += DNI3loss;
DNI3totalLossCount++;
Console.WriteLine("\nbatch count " + i + "/" + TRAIN_DATA_COUNT);
//結果出力
Console.WriteLine("total loss " + totalLoss / totalLossCount);
Console.WriteLine("local loss " + sumLoss);
Console.WriteLine("\nDNI1 total loss " + DNI1totalLoss / DNI1totalLossCount);
Console.WriteLine("DNI2 total loss " + DNI2totalLoss / DNI2totalLossCount);
Console.WriteLine("DNI3 total loss " + DNI3totalLoss / DNI3totalLossCount);
Console.WriteLine("\nDNI1 local loss " + DNI1loss);
Console.WriteLine("DNI2 local loss " + DNI2loss);
Console.WriteLine("DNI3 local loss " + DNI3loss);
//20回バッチを動かしたら精度をテストする
if (i % 20 == 0)
{
Console.WriteLine("\nTesting...");
//テストデータからランダムにデータを取得
TestDataSet datasetY = mnistData.GetRandomYSet(TEST_DATA_COUNT);
//テストを実行
Real accuracy = Trainer.Accuracy(nn, datasetY.Data, datasetY.Label);
Console.WriteLine("accuracy " + accuracy);
}
}
}
}