public static void Run()
{
DataMaker dataMaker = new DataMaker(STEPS_PER_CYCLE, NUMBER_OF_CYCLES);
NdArray trainData = dataMaker.Make();
FunctionStack model = new FunctionStack("Test8",
new Linear(true, 1, 5, name: "Linear l1"),
new LSTM(true, 5, 5, name: "LSTM l2"),
new Linear(true, 5, 1, name: "Linear l3")
);
model.SetOptimizer(new Adam());
RILogManager.Default?.SendDebug("Training...");
for (int epoch = 0; epoch < TRAINING_EPOCHS; epoch++)
{
NdArray[] sequences = dataMaker.MakeMiniBatch(trainData, MINI_BATCH_SIZE, LENGTH_OF_SEQUENCE);
Real loss = ComputeLoss(model, sequences);
model.Update();
model.ResetState();
if (epoch != 0 && epoch % DISPLAY_EPOCH == 0)
{
RILogManager.Default?.SendDebug("[{0}]training loss:\t{1}", epoch, loss);
}
}
RILogManager.Default?.SendDebug("Testing...");
NdArray[] testSequences = dataMaker.MakeMiniBatch(trainData, MINI_BATCH_SIZE, LENGTH_OF_SEQUENCE);
int sample_index = 45;
predict(testSequences[sample_index], model, PREDICTION_LENGTH);
}
public static void Run()
{
//訓練回数
const int learningCount = 10000;
//訓練データ
Real[][] trainData =
{
new Real[] { 0, 0 },
new Real[] { 1, 0 },
new Real[] { 0, 1 },
new Real[] { 1, 1 }
};
//訓練データラベル
Real[][] trainLabel =
{
new Real[] { 0 },
new Real[] { 1 },
new Real[] { 1 },
new Real[] { 0 }
};
//ネットワークの構成を FunctionStack に書き連ねる
FunctionStack nn = new FunctionStack(
new Linear(2, 2, name: "l1 Linear"),
new ReLU(name: "l1 ReLU"),
new Linear(2, 1, name: "l2 Linear")
);
//optimizerを宣言(今回はAdam)
nn.SetOptimizer(new Adam());
//訓練ループ
Console.WriteLine("Training...");
for (int i = 0; i < learningCount; i++)
{
//今回はロス関数にMeanSquaredErrorを使う
Trainer.Train(nn, trainData[0], trainLabel[0], new MeanSquaredError(), false);
Trainer.Train(nn, trainData[1], trainLabel[1], new MeanSquaredError(), false);
Trainer.Train(nn, trainData[2], trainLabel[2], new MeanSquaredError(), false);
Trainer.Train(nn, trainData[3], trainLabel[3], new MeanSquaredError(), false);
//訓練後に毎回更新を実行しなければ、ミニバッチとして更新できる
nn.Update();
}
//訓練結果を表示
Console.WriteLine("Test Start...");
foreach (Real[] val in trainData)
{
NdArray result = nn.Predict(val)[0];
Console.WriteLine(val[0] + " xor " + val[1] + " = " + (result.Data[0] > 0.5 ? 1 : 0) + " " + result);
}
}
public static void Run()
{
//Number of exercises
const int learningCount = 10000;
//Training data
Real[][] trainData =
{
new Real[] { 0, 0 },
new Real[] { 1, 0 },
new Real[] { 0, 1 },
new Real[] { 1, 1 }
};
//Training data label
Real[][] trainLabel =
{
new Real[] { 0 },
new Real[] { 1 },
new Real[] { 1 },
new Real[] { 0 }
};
//Writing the network configuration in FunctionStack
FunctionStack nn = new FunctionStack(
new Linear(2, 2, name: "l1 Linear"),
new ReLU(name: "l1 ReLU"),
new Linear(2, 1, name: "l2 Linear")
);
//Declare optimizer (Adam in this time)
nn.SetOptimizer(new Adam());
//Training loop
Console.WriteLine("Training...");
for (int i = 0; i < learningCount; i++)
{
//This time use MeanSquaredError for loss function
Trainer.Train(nn, trainData[0], trainLabel[0], new MeanSquaredError(), false);
Trainer.Train(nn, trainData[1], trainLabel[1], new MeanSquaredError(), false);
Trainer.Train(nn, trainData[2], trainLabel[2], new MeanSquaredError(), false);
Trainer.Train(nn, trainData[3], trainLabel[3], new MeanSquaredError(), false);
//If you do not update every time after training, you can update it as a mini batch
nn.Update();
}
//Show training results
Console.WriteLine("Test Start...");
foreach (Real[] val in trainData)
{
NdArray result = nn.Predict(val)[0];
Console.WriteLine(val[0] + " xor " + val[1] + " = " + (result.Data[0] > 0.5 ? 1 : 0) + " " + result);
}
}
public static void Run()
{
const int learningCount = 10000;
Real[][] trainData =
{
new Real[] { 0, 0 },
new Real[] { 1, 0 },
new Real[] { 0, 1 },
new Real[] { 1, 1 }
};
Real[][] trainLabel =
{
new Real[] { 0 },
new Real[] { 1 },
new Real[] { 1 },
new Real[] { 0 }
};
FunctionStack nn = new FunctionStack("Test2",
new Linear(true, 2, 2, name: "l1 Linear"),
new ReLU(name: "l1 ReLU"),
new Linear(true, 2, 1, name: "l2 Linear"));
nn.SetOptimizer(new AdaGrad());
RILogManager.Default?.SendDebug("Training...");
for (int i = 0; i < learningCount; i++)
{
//use MeanSquaredError for loss function
Trainer.Train(nn, trainData[0], trainLabel[0], new MeanSquaredError(), false);
Trainer.Train(nn, trainData[1], trainLabel[1], new MeanSquaredError(), false);
Trainer.Train(nn, trainData[2], trainLabel[2], new MeanSquaredError(), false);
Trainer.Train(nn, trainData[3], trainLabel[3], new MeanSquaredError(), false);
//If you do not update every time after training, you can update it as a mini batch
nn.Update();
}
RILogManager.Default?.SendDebug("Test Start...");
foreach (Real[] val in trainData)
{
NdArray result = nn.Predict(true, val)[0];
RILogManager.Default?.SendDebug($"{val[0]} xor {val[1]} = {(result.Data[0] > 0.5 ? 1 : 0)} {result}");
}
}
//バッチで学習処理を行う
public static Real Train(FunctionStack functionStack, NdArray input, NdArray teach, LossFunction lossFunction, bool isUpdate = true)
{
//結果の誤差保存用
NdArray[] result = functionStack.Forward(input);
Real sumLoss = lossFunction.Evaluate(result, teach);
//Backwardのバッチを実行
functionStack.Backward(result);
//更新
if (isUpdate)
{
functionStack.Update();
}
return(sumLoss);
}
//Perform learning process in batch
public static Real Train(FunctionStack functionStack, NdArray input, NdArray teach, LossFunction lossFunction, bool isUpdate = true)
{
//For preserving error of result
NdArray[] result = functionStack.Forward(input);
Real sumLoss = lossFunction.Evaluate(result, teach);
//Run Backward's batch
functionStack.Backward(result);
//update
if (isUpdate)
{
functionStack.Update();
}
return(sumLoss);
}
////////////////////////////////////////////////////////////////////////////////////////////////////
/// <summary> Do a learning process with a batch. </summary>
///
/// <param name="functionStack"> Stack of functions. </param>
/// <param name="input"> The input data. </param>
/// <param name="teach"> The teaching data. </param>
/// <param name="lossFunction"> The loss function. </param>
/// <param name="isUpdate"> (Optional) True if this object is being updated. </param>
///
/// <returns> A Real. </returns>
////////////////////////////////////////////////////////////////////////////////////////////////////
public static Real Train([NotNull] FunctionStack functionStack, [CanBeNull] NdArray input, [CanBeNull] NdArray teach, [NotNull] LossFunction lossFunction, bool isUpdate = true,
bool verbose = true)
{
if (verbose)
{
RILogManager.Default?.EnterMethod("Training " + functionStack.Name);
}
if (verbose)
{
RILogManager.Default?.SendDebug("Forward propagation");
}
NdArray[] result = functionStack.Forward(verbose, input);
if (verbose)
{
RILogManager.Default?.SendDebug("Evaluating loss");
}
Real sumLoss = lossFunction.Evaluate(result, teach);
// Run Backward batch
if (verbose)
{
RILogManager.Default?.SendDebug("Backward propagation");
}
functionStack.Backward(verbose, result);
if (isUpdate)
{
if (verbose)
{
RILogManager.Default?.SendDebug("Updating stack");
}
functionStack.Update();
}
if (verbose)
{
RILogManager.Default?.ExitMethod("Training " + functionStack.Name);
RILogManager.Default?.ViewerSendWatch("Local Loss", sumLoss.ToString(), sumLoss);
}
return(sumLoss);
}
public static void Run()
{
DataMaker dataMaker = new DataMaker(STEPS_PER_CYCLE, NUMBER_OF_CYCLES);
NdArray trainData = dataMaker.Make();
//Network configuration is written in FunctionStack
FunctionStack model = new FunctionStack(
new Linear(1, 5, name: "Linear l1"),
new LSTM(5, 5, name: "LSTM l2"),
new Linear(5, 1, name: "Linear l3")
);
//Declare optimizer
model.SetOptimizer(new Adam());
//Training loop
Console.WriteLine("Training...");
for (int epoch = 0; epoch < TRAINING_EPOCHS; epoch++)
{
NdArray[] sequences = dataMaker.MakeMiniBatch(trainData, MINI_BATCH_SIZE, LENGTH_OF_SEQUENCE);
Real loss = ComputeLoss(model, sequences);
model.Update();
model.ResetState();
if (epoch != 0 && epoch % DISPLAY_EPOCH == 0)
{
Console.WriteLine("[{0}]training loss:\t{1}", epoch, loss);
}
}
Console.WriteLine("Testing...");
NdArray[] testSequences = dataMaker.MakeMiniBatch(trainData, MINI_BATCH_SIZE, LENGTH_OF_SEQUENCE);
int sample_index = 45;
predict(testSequences[sample_index], model, PREDICTION_LENGTH);
}
public static void Run()
{
// Write the configuration of the network you want to read into FunctionStack and adjust the parameters of each function
// Make sure to match name to the variable name of Chainer here
FunctionStack nn = new FunctionStack("Test16",
new Convolution2D(true, 1, 2, 3, name: "conv1", gpuEnable: true),// Do not forget the GPU flag if necessary
new ReLU(),
new MaxPooling(2, 2),
new Convolution2D(true, 2, 2, 2, name: "conv2", gpuEnable: true),
new ReLU(),
new MaxPooling(2, 2),
new Linear(true, 8, 2, name: "fl3"),
new ReLU(),
new Linear(true, 2, 2, name: "fl4")
);
/* Chainerでの宣言
* class NN(chainer.Chain):
* def __init__(self):
* super(NN, self).__init__(
* conv1 = L.Convolution2D(1,2,3),
* conv2 = L.Convolution2D(2,2,2),
* fl3 = L.Linear(8,2),
* fl4 = L.Linear(2,2)
* )
*
* def __call__(self, x):
* h_conv1 = F.relu(self.conv1(x))
* h_pool1 = F.max_pooling_2d(h_conv1, 2)
* h_conv2 = F.relu(self.conv2(h_pool1))
* h_pool2 = F.max_pooling_2d(h_conv2, 2)
* h_fc1 = F.relu(self.fl3(h_pool2))
* y = self.fl4(h_fc1)
* return y
*/
// Read parameters
ChainerModelDataLoader.ModelLoad(MODEL_FILE_PATH, nn);
// We will use the rest as usual
nn.SetOptimizer(new SGD());
NdArray x = new NdArray(new Real[, , ] {
{
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.2, 0.9, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.9, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.8, 0.5, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.3, 0.3, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.4, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.8, 0.4, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.8, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 }
}
});
Real[] t = { 0.0, 1.0 };
Trainer.Train(nn, x, t, new MeanSquaredError(), false);
Convolution2D l2 = (Convolution2D)nn.Functions[0];
RILogManager.Default?.SendDebug("gw1");
RILogManager.Default?.SendDebug(l2.Weight.ToString("Grad"));
RILogManager.Default?.SendDebug("gb1");
RILogManager.Default?.SendDebug(l2.Bias.ToString("Grad"));
// If Update is executed, grad is consumed, so output the value first
nn.Update();
RILogManager.Default?.SendDebug("w1");
RILogManager.Default?.SendDebug(l2.Weight.ToString());
RILogManager.Default?.SendDebug("b1");
RILogManager.Default?.SendDebug(l2.Bias.ToString());
}
public static void Run()
{
//各初期値を記述
Real[,,,] initial_W1 =
{
{ { { 1.0, 0.5, 0.0 }, { 0.5, 0.0, -0.5 }, { 0.0, -0.5, -1.0 } } },
{ { { 0.0, -0.1, 0.1 }, { -0.3, 0.4, 0.7 }, { 0.5, -0.2, 0.2 } } }
};
Real[] initial_b1 = { 0.5, 1.0 };
Real[,,,] initial_W2 =
{
{ { { -0.1, 0.6 }, { 0.3, -0.9 } }, { { 0.7, 0.9 }, { -0.2, -0.3 } } },
{ { { -0.6, -0.1 }, { 0.3, 0.3 } }, { { -0.5, 0.8 }, { 0.9, 0.1 } } }
};
Real[] initial_b2 = { 0.1, 0.9 };
Real[,] initial_W3 =
{
{ 0.5, 0.3, 0.4, 0.2, 0.6, 0.1, 0.4, 0.3 },
{ 0.6, 0.4, 0.9, 0.1, 0.5, 0.2, 0.3, 0.4 }
};
Real[] initial_b3 = { 0.01, 0.02 };
Real[,] initial_W4 = { { 0.8, 0.2 }, { 0.4, 0.6 } };
Real[] initial_b4 = { 0.02, 0.01 };
//入力データ
NdArray x = new NdArray(new Real[, , ] {
{
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.2, 0.9, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.9, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.8, 0.5, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.3, 0.3, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.4, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.8, 0.4, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.8, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 }
}
});
//教師信号
Real[] t = { 0.0, 1.0 };
//層の中身をチェックしたい場合は、層単体でインスタンスを持つ
Convolution2D l2 = new Convolution2D(1, 2, 3, initialW: initial_W1, initialb: initial_b1, name: "l2 Conv2D");
//ネットワークの構成を FunctionStack に書き連ねる
FunctionStack nn = new FunctionStack(
l2, //new Convolution2D(1, 2, 3, initialW: initial_W1, initialb: initial_b1),
new ReLU(name: "l2 ReLU"),
//new AveragePooling(2, 2, name: "l2 AVGPooling"),
new MaxPooling2D(2, 2, name: "l2 MaxPooling"),
new Convolution2D(2, 2, 2, initialW: initial_W2, initialb: initial_b2, name: "l3 Conv2D"),
new ReLU(name: "l3 ReLU"),
//new AveragePooling(2, 2, name: "l3 AVGPooling"),
new MaxPooling2D(2, 2, name: "l3 MaxPooling"),
new Linear(8, 2, initialW: initial_W3, initialb: initial_b3, name: "l4 Linear"),
new ReLU(name: "l4 ReLU"),
new Linear(2, 2, initialW: initial_W4, initialb: initial_b4, name: "l5 Linear")
);
nn.SetOptimizer(new SGD(0.1));
//訓練を実施
Trainer.Train(nn, x, t, new MeanSquaredError(), false);
//Updateを実行するとgradが消費されてしまうため値を先に出力
Console.WriteLine("gw1");
Console.WriteLine(l2.Weight.ToString("Grad"));
Console.WriteLine("gb1");
Console.WriteLine(l2.Bias.ToString("Grad"));
//更新
nn.Update();
Console.WriteLine("w1");
Console.WriteLine(l2.Weight);
Console.WriteLine("b1");
Console.WriteLine(l2.Bias);
}
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);
}
}
}
}
public static void Run()
{
//Prepare MNIST data
Console.WriteLine("MNIST Data Loading...");
MnistData mnistData = new MnistData();
Console.WriteLine("Training Start...");
//Writing the network configuration in FunctionStack
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 itself is also stacked as Function
FunctionStack nn = new FunctionStack
(
Layer1,
Layer2,
Layer3,
Layer4
);
FunctionStack cDNI1 = new FunctionStack(
new Linear(256 + 10, 1024, name: "cDNI1 Linear1"),
new BatchNormalization(1024, name: "cDNI1 Nrom1"),
new ReLU(name: "cDNI1 ReLU1"),
new Linear(1024, 256, initialW: new Real[1024, 256], name: "DNI1 Linear3")
);
FunctionStack cDNI2 = new FunctionStack(
new Linear(256 + 10, 1024, name: "cDNI2 Linear1"),
new BatchNormalization(1024, name: "cDNI2 Nrom1"),
new ReLU(name: "cDNI2 ReLU1"),
new Linear(1024, 256, initialW: new Real[1024, 256], name: "cDNI2 Linear3")
);
FunctionStack cDNI3 = new FunctionStack(
new Linear(256 + 10, 1024, name: "cDNI3 Linear1"),
new BatchNormalization(1024, name: "cDNI3 Nrom1"),
new ReLU(name: "cDNI3 ReLU1"),
new Linear(1024, 256, initialW: new Real[1024, 256], name: "cDNI3 Linear3")
);
//Declare optimizer
Layer1.SetOptimizer(new Adam(0.00003f));
Layer2.SetOptimizer(new Adam(0.00003f));
Layer3.SetOptimizer(new Adam(0.00003f));
Layer4.SetOptimizer(new Adam(0.00003f));
cDNI1.SetOptimizer(new Adam(0.00003f));
cDNI2.SetOptimizer(new Adam(0.00003f));
cDNI3.SetOptimizer(new Adam(0.00003f));
for (int epoch = 0; epoch < 10; epoch++)
{
Console.WriteLine("epoch " + (epoch + 1));
//Total error in the whole
Real totalLoss = 0;
Real cDNI1totalLoss = 0;
Real cDNI2totalLoss = 0;
Real cDNI3totalLoss = 0;
long totalLossCount = 0;
long cDNI1totalLossCount = 0;
long cDNI2totalLossCount = 0;
long cDNI3totalLossCount = 0;
//How many times to run the batch
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
//Get data randomly from training data
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT);
//Run first tier
NdArray[] layer1ForwardResult = Layer1.Forward(datasetX.Data);
ResultDataSet layer1ResultDataSet = new ResultDataSet(layer1ForwardResult, datasetX.Label);
//Get the inclination of the first layer
NdArray[] cDNI1Result = cDNI1.Forward(layer1ResultDataSet.GetTrainData());
//Apply the inclination of the first layer
layer1ForwardResult[0].Grad = cDNI1Result[0].Data.ToArray();
//Update first layer
Layer1.Backward(layer1ForwardResult);
layer1ForwardResult[0].ParentFunc = null;
Layer1.Update();
//Run Layer 2
NdArray[] layer2ForwardResult = Layer2.Forward(layer1ResultDataSet.Result);
ResultDataSet layer2ResultDataSet = new ResultDataSet(layer2ForwardResult, layer1ResultDataSet.Label);
//Get inclination of second layer
NdArray[] cDNI2Result = cDNI2.Forward(layer2ResultDataSet.GetTrainData());
//Apply the inclination of the second layer
layer2ForwardResult[0].Grad = cDNI2Result[0].Data.ToArray();
//Update 2nd tier
Layer2.Backward(layer2ForwardResult);
layer2ForwardResult[0].ParentFunc = null;
//Perform learning of first layer cDNI
Real cDNI1loss = new MeanSquaredError().Evaluate(cDNI1Result, new NdArray(layer1ResultDataSet.Result[0].Grad, cDNI1Result[0].Shape, cDNI1Result[0].BatchCount));
Layer2.Update();
cDNI1.Backward(cDNI1Result);
cDNI1.Update();
cDNI1totalLoss += cDNI1loss;
cDNI1totalLossCount++;
//Run Third Tier
NdArray[] layer3ForwardResult = Layer3.Forward(layer2ResultDataSet.Result);
ResultDataSet layer3ResultDataSet = new ResultDataSet(layer3ForwardResult, layer2ResultDataSet.Label);
//Get the inclination of the third layer
NdArray[] cDNI3Result = cDNI3.Forward(layer3ResultDataSet.GetTrainData());
//Apply the inclination of the third layer
layer3ForwardResult[0].Grad = cDNI3Result[0].Data.ToArray();
//Update third layer
Layer3.Backward(layer3ForwardResult);
layer3ForwardResult[0].ParentFunc = null;
//Perform learning of cDNI for layer 2
Real cDNI2loss = new MeanSquaredError().Evaluate(cDNI2Result, new NdArray(layer2ResultDataSet.Result[0].Grad, cDNI2Result[0].Shape, cDNI2Result[0].BatchCount));
Layer3.Update();
cDNI2.Backward(cDNI2Result);
cDNI2.Update();
cDNI2totalLoss += cDNI2loss;
cDNI2totalLossCount++;
//Run Layer 4
NdArray[] layer4ForwardResult = Layer4.Forward(layer3ResultDataSet.Result);
//Get inclination of the fourth layer
Real sumLoss = new SoftmaxCrossEntropy().Evaluate(layer4ForwardResult, layer3ResultDataSet.Label);
//Update fourth layer
Layer4.Backward(layer4ForwardResult);
layer4ForwardResult[0].ParentFunc = null;
totalLoss += sumLoss;
totalLossCount++;
//Perform learning of cDNI for the third layer
Real cDNI3loss = new MeanSquaredError().Evaluate(cDNI3Result, new NdArray(layer3ResultDataSet.Result[0].Grad, cDNI3Result[0].Shape, cDNI3Result[0].BatchCount));
Layer4.Update();
cDNI3.Backward(cDNI3Result);
cDNI3.Update();
cDNI3totalLoss += cDNI3loss;
cDNI3totalLossCount++;
Console.WriteLine("\nbatch count " + i + "/" + TRAIN_DATA_COUNT);
//Result output
Console.WriteLine("total loss " + totalLoss / totalLossCount);
Console.WriteLine("local loss " + sumLoss);
Console.WriteLine("\ncDNI1 total loss " + cDNI1totalLoss / cDNI1totalLossCount);
Console.WriteLine("cDNI2 total loss " + cDNI2totalLoss / cDNI2totalLossCount);
Console.WriteLine("cDNI3 total loss " + cDNI3totalLoss / cDNI3totalLossCount);
Console.WriteLine("\ncDNI1 local loss " + cDNI1loss);
Console.WriteLine("cDNI2 local loss " + cDNI2loss);
Console.WriteLine("cDNI3 local loss " + cDNI3loss);
//Test the accuracy if you move the batch 20 times
if (i % 20 == 0)
{
Console.WriteLine("\nTesting...");
//Get data randomly from test data
TestDataSet datasetY = mnistData.GetRandomYSet(TEST_DATA_COUNT);
//Run test
Real accuracy = Trainer.Accuracy(nn, datasetY.Data, datasetY.Label);
Console.WriteLine("accuracy " + accuracy);
}
}
}
}
public static void Main()
{
// platformIdは、OpenCL・GPUの導入の記事に書いてある方法でご確認ください
// https://jinbeizame.hateblo.jp/entry/kelpnet_opencl_gpu
Weaver.Initialize(ComputeDeviceTypes.Gpu, platformId: 1, deviceIndex: 0);
// ネットからVGGの学習済みモデルをダウンロード
string modelFilePath = InternetFileDownloader.Donwload(DOWNLOAD_URL, MODEL_FILE);
// 学習済みモデルをFunctionのリストとして保存
List <Function> vgg16Net = CaffemodelDataLoader.ModelLoad(modelFilePath);
// VGGの出力層とその活性化関数を削除
vgg16Net.RemoveAt(vgg16Net.Count() - 1);
vgg16Net.RemoveAt(vgg16Net.Count() - 1);
// VGGの各FunctionのgpuEnableをtrueに
for (int i = 0; i < vgg16Net.Count - 1; i++)
{
// GPUに対応している層であれば、GPU対応へ
if (vgg16Net[i] is Convolution2D || vgg16Net[i] is Linear || vgg16Net[i] is MaxPooling)
{
((IParallelizable)vgg16Net[i]).SetGpuEnable(true);
}
}
// VGGをリストからFunctionStackに変換
FunctionStack vgg = new FunctionStack(vgg16Net.ToArray());
// 層を圧縮
vgg.Compress();
// 新しく出力層とその活性化関数を用意
FunctionStack nn = new FunctionStack(
new Linear(4096, 1, gpuEnable: true),
new Sigmoid()
);
// 最適化手法としてAdamをセット
nn.SetOptimizer(new Adam());
Console.WriteLine("DataSet Loading...");
// 訓練・テストデータ用のNdArrayを用意
// データセットは以下のURLからダウンロードを行い、
// VGGTransfer /bin/Debug/Data にtrainフォルダを置いてください。
// https://www.kaggle.com/c/dogs-vs-cats/data
NdArray[] trainData = new NdArray[TRAIN_DATA_LENGTH * 2];
NdArray[] trainLabel = new NdArray[TRAIN_DATA_LENGTH * 2];
NdArray[] testData = new NdArray[TEST_DATA_LENGTH * 2];
NdArray[] testLabel = new NdArray[TEST_DATA_LENGTH * 2];
for (int i = 0; i < TRAIN_DATA_LENGTH + TEST_DATA_LENGTH; i++)
{
// 犬・猫の画像読み込み
Bitmap baseCatImage = new Bitmap("Data/train/cat." + i + ".jpg");
Bitmap baseDogImage = new Bitmap("Data/train/dog." + i + ".jpg");
// 変換後の画像を格納するBitmapを定義
Bitmap catImage = new Bitmap(224, 224, PixelFormat.Format24bppRgb);
Bitmap dogImage = new Bitmap(224, 224, PixelFormat.Format24bppRgb);
// Graphicsオブジェクトに変換
Graphics gCat = Graphics.FromImage(catImage);
Graphics gDog = Graphics.FromImage(dogImage);
// Graphicsオブジェクト（の中のcatImageに）baseImageを変換して描画
gCat.DrawImage(baseCatImage, 0, 0, 224, 224);
gDog.DrawImage(baseDogImage, 0, 0, 224, 224);
// Graphicsオブジェクトを破棄し、メモリを解放
gCat.Dispose();
gDog.Dispose();
// 訓練・テストデータにデータを格納
// 先にテストデータの枚数分テストデータに保存し、その後訓練データを保存する
// 画素値の値域は0 ~ 255のため、255で割ることで0 ~ 1に正規化
if (i < TEST_DATA_LENGTH)
{
// ImageをNdArrayに変換したものをvggに入力し、出力した特徴量を入力データとして保存
testData[i * 2] = vgg.Predict(NdArrayConverter.Image2NdArray(catImage, false, true) / 255.0)[0];
testLabel[i * 2] = new NdArray(new Real[] { 0 });
testData[i * 2 + 1] = vgg.Predict(NdArrayConverter.Image2NdArray(dogImage, false, true) / 255.0)[0];
testLabel[i * 2 + 1] = new NdArray(new Real[] { 1 });
}
else
{
trainData[(i - TEST_DATA_LENGTH) * 2] = vgg.Predict(NdArrayConverter.Image2NdArray(catImage, false, true) / 255.0)[0];
trainLabel[(i - TEST_DATA_LENGTH) * 2] = new NdArray(new Real[] { 0 }); //new Real [] { 0 };
trainData[(i - TEST_DATA_LENGTH) * 2] = vgg.Predict(NdArrayConverter.Image2NdArray(dogImage, false, true) / 255.0)[0];
trainLabel[(i - TEST_DATA_LENGTH) * 2] = new NdArray(new Real[] { 1 }); // = new Real [] { 1 };
}
}
Console.WriteLine("Training Start...");
// ミニバッチ用のNdArrayを定義
NdArray batchData = new NdArray(new[] { 4096 }, BATCH_SIZE);
NdArray batchLabel = new NdArray(new[] { 1 }, BATCH_SIZE);
// 誤差関数を定義（今回は二値分類なので二乗誤差関数(MSE)）
LossFunction lossFunction = new MeanSquaredError();
// エポックを回す
for (int epoch = 0; epoch < 10; epoch++)
{
// １エポックで訓練データ // バッチサイズ の回数分学習
for (int step = 0; step < TRAIN_DATA_COUNT; step++)
{
// ミニバッチを用意
for (int i = 0; i < BATCH_SIZE; i++)
{
// 0 ~ 訓練データサイズ-1 の中からランダムで整数を取得
int index = Mother.Dice.Next(trainData.Length);
// trainData(NdArray[])を、batchData(NdArray)の形にコピー
Array.Copy(trainData[index].Data, 0, batchData.Data, i * batchData.Length, batchData.Length);
batchLabel.Data[i] = trainLabel[index].Data[0];
}
// 学習（順伝播、誤差の計算、逆伝播、更新）
NdArray[] output = nn.Forward(batchData);
Real loss = lossFunction.Evaluate(output, batchLabel);
nn.Backward(output);
nn.Update();
}
// 認識率（accuracy）の計算
// テストデータの回数データを回す
Real accuracy = 0;
for (int i = 0; i < TEST_DATA_LENGTH * 2; i++)
{
NdArray[] output = nn.Predict(testData[i]);
// 出力outputと正解の誤差が0.5以下（正解が0のときにoutput<0.5、正解が1のときにoutput>0.5）
// の際に正確に認識したとする
if (Math.Abs(output[0].Data[0] - trainLabel[i].Data[0]) < 0.5)
{
accuracy += 1;
}
accuracy /= TEST_DATA_LENGTH * 2.0;
Console.WriteLine("Epoch:" + epoch + "accuracy:" + accuracy);
}
}
}
public static void Run()
{
//読み込みたいネットワークの構成を FunctionStack に書き連ね、各 Function のパラメータを合わせる
//ここで必ず name を Chainer の変数名に合わせておくこと
FunctionStack nn = new FunctionStack(
new Convolution2D(1, 2, 3, name: "conv1", gpuEnable: true),//必要であればGPUフラグも忘れずに
new ReLU(),
new MaxPooling(2, 2),
new Convolution2D(2, 2, 2, name: "conv2", gpuEnable: true),
new ReLU(),
new MaxPooling(2, 2),
new Linear(8, 2, name: "fl3"),
new ReLU(),
new Linear(2, 2, name: "fl4")
);
/* Chainerでの宣言
* class NN(chainer.Chain):
* def __init__(self):
* super(NN, self).__init__(
* conv1 = L.Convolution2D(1,2,3),
* conv2 = L.Convolution2D(2,2,2),
* fl3 = L.Linear(8,2),
* fl4 = L.Linear(2,2)
* )
*
* def __call__(self, x):
* h_conv1 = F.relu(self.conv1(x))
* h_pool1 = F.max_pooling_2d(h_conv1, 2)
* h_conv2 = F.relu(self.conv2(h_pool1))
* h_pool2 = F.max_pooling_2d(h_conv2, 2)
* h_fc1 = F.relu(self.fl3(h_pool2))
* y = self.fl4(h_fc1)
* return y
*/
//パラメータを読み込み
ChainerModelDataLoader.ModelLoad(MODEL_FILE_PATH, nn);
//あとは通常通り使用する
nn.SetOptimizer(new SGD());
//入力データ
NdArray x = new NdArray(new Real[, , ] {
{
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.2, 0.9, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.9, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.8, 0.5, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.3, 0.3, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.4, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.8, 0.4, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.8, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 }
}
});
//教師信号
Real[] t = { 0.0, 1.0 };
//訓練を実施
Trainer.Train(nn, x, t, new MeanSquaredError(), false);
//結果表示用に退避
Convolution2D l2 = (Convolution2D)nn.Functions[0];
//Updateを実行するとgradが消費されてしまうため値を先に出力
Console.WriteLine("gw1");
Console.WriteLine(l2.Weight.ToString("Grad"));
Console.WriteLine("gb1");
Console.WriteLine(l2.Bias.ToString("Grad"));
//更新
nn.Update();
Console.WriteLine("w1");
Console.WriteLine(l2.Weight);
Console.WriteLine("b1");
Console.WriteLine(l2.Bias);
}
public static void Run()
{
// Prepare MNIST data
RILogManager.Default?.SendDebug("MNIST Data Loading...");
MnistData mnistData = new MnistData(28);
RILogManager.Default?.SendDebug("Training Start...");
// Write the network configuration in FunctionStack
FunctionStack Layer1 = new FunctionStack("Test12 Layer 1",
new Linear(true, 28 * 28, 256, name: "l1 Linear"),
new BatchNormalization(true, 256, name: "l1 Norm"),
new ReLU(name: "l1 ReLU")
);
FunctionStack Layer2 = new FunctionStack("Test12 Layer 2",
new Linear(true, 256, 256, name: "l2 Linear"),
new BatchNormalization(true, 256, name: "l2 Norm"),
new ReLU(name: "l2 ReLU")
);
FunctionStack Layer3 = new FunctionStack("Test12 Layer 3",
new Linear(true, 256, 256, name: "l3 Linear"),
new BatchNormalization(true, 256, name: "l3 Norm"),
new ReLU(name: "l3 ReLU")
);
FunctionStack Layer4 = new FunctionStack("Test12 Layer 4",
new Linear(true, 256, 10, name: "l4 Linear")
);
// Function stack itself is also stacked as Function
FunctionStack nn = new FunctionStack
("Test12",
Layer1,
Layer2,
Layer3,
Layer4
);
FunctionStack cDNI1 = new FunctionStack("Test12 DNI 1",
new Linear(true, 256 + 10, 1024, name: "cDNI1 Linear1"),
new BatchNormalization(true, 1024, name: "cDNI1 Norm1"),
new ReLU(name: "cDNI1 ReLU1"),
new Linear(true, 1024, 256, initialW: new Real[1024, 256], name: "DNI1 Linear3")
);
FunctionStack cDNI2 = new FunctionStack("Test12 DNI 2",
new Linear(true, 256 + 10, 1024, name: "cDNI2 Linear1"),
new BatchNormalization(true, 1024, name: "cDNI2 Norm1"),
new ReLU(name: "cDNI2 ReLU1"),
new Linear(true, 1024, 256, initialW: new Real[1024, 256], name: "cDNI2 Linear3")
);
FunctionStack cDNI3 = new FunctionStack("Test12 DNI 3",
new Linear(true, 256 + 10, 1024, name: "cDNI3 Linear1"),
new BatchNormalization(true, 1024, name: "cDNI3 Norm1"),
new ReLU(name: "cDNI3 ReLU1"),
new Linear(true, 1024, 256, initialW: new Real[1024, 256], name: "cDNI3 Linear3")
);
Layer1.SetOptimizer(new Adam("Adam", 0.00003f));
Layer2.SetOptimizer(new Adam("Adam", 0.00003f));
Layer3.SetOptimizer(new Adam("Adam", 0.00003f));
Layer4.SetOptimizer(new Adam("Adam", 0.00003f));
cDNI1.SetOptimizer(new Adam("Adam", 0.00003f));
cDNI2.SetOptimizer(new Adam("Adam", 0.00003f));
cDNI3.SetOptimizer(new Adam("Adam", 0.00003f));
// Describe each function stack;
RILogManager.Default?.SendDebug(Layer1.Describe());
RILogManager.Default?.SendDebug(Layer2.Describe());
RILogManager.Default?.SendDebug(Layer3.Describe());
RILogManager.Default?.SendDebug(Layer4.Describe());
RILogManager.Default?.SendDebug(cDNI1.Describe());
RILogManager.Default?.SendDebug(cDNI2.Describe());
RILogManager.Default?.SendDebug(cDNI3.Describe());
for (int epoch = 0; epoch < 10; epoch++)
{
// Total error in the whole
Real totalLoss = 0;
Real cDNI1totalLoss = 0;
Real cDNI2totalLoss = 0;
Real cDNI3totalLoss = 0;
long totalLossCount = 0;
long cDNI1totalLossCount = 0;
long cDNI2totalLossCount = 0;
long cDNI3totalLossCount = 0;
// how many times to run the batch
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
RILogManager.Default?.SendDebug("epoch: " + (epoch + 1) + " of 10, batch iteration: " + i + " of " + TRAIN_DATA_COUNT);
RILogManager.Default?.ViewerSendWatch("Epoch", epoch + 1);
RILogManager.Default?.ViewerSendWatch("Batch Iteration", i);
// Get data randomly from the training data
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT, 28, 28);
// Run first tier
NdArray[] layer1ForwardResult = Layer1.Forward(true, datasetX.Data);
ResultDataSet layer1ResultDataSet = new ResultDataSet(layer1ForwardResult, datasetX.Label);
// Obtain the slope of the first layer
NdArray[] cDNI1Result = cDNI1.Forward(true, layer1ResultDataSet.GetTrainData());
// Apply the slope of the first layer
layer1ForwardResult[0].Grad = cDNI1Result[0].Data.ToArray();
//Update first layer
Layer1.Backward(true, layer1ForwardResult);
layer1ForwardResult[0].ParentFunc = null;
Layer1.Update();
// Run Layer 2
NdArray[] layer2ForwardResult = Layer2.Forward(true, layer1ResultDataSet.Result);
ResultDataSet layer2ResultDataSet = new ResultDataSet(layer2ForwardResult, layer1ResultDataSet.Label);
// Get the inclination of the second layer
NdArray[] cDNI2Result = cDNI2.Forward(true, layer2ResultDataSet.GetTrainData());
// Apply the slope of the second layer
layer2ForwardResult[0].Grad = cDNI2Result[0].Data.ToArray();
//Update layer 2
Layer2.Backward(true, layer2ForwardResult);
layer2ForwardResult[0].ParentFunc = null;
//Perform learning of first layer cDNI
Real cDNI1loss = new MeanSquaredError().Evaluate(cDNI1Result, new NdArray(layer1ResultDataSet.Result[0].Grad, cDNI1Result[0].Shape, cDNI1Result[0].BatchCount));
Layer2.Update();
cDNI1.Backward(true, cDNI1Result);
cDNI1.Update();
cDNI1totalLoss += cDNI1loss;
cDNI1totalLossCount++;
//Run Third Tier
NdArray[] layer3ForwardResult = Layer3.Forward(true, layer2ResultDataSet.Result);
ResultDataSet layer3ResultDataSet = new ResultDataSet(layer3ForwardResult, layer2ResultDataSet.Label);
//Get the inclination of the third layer
NdArray[] cDNI3Result = cDNI3.Forward(true, layer3ResultDataSet.GetTrainData());
//Apply the inclination of the third layer
layer3ForwardResult[0].Grad = cDNI3Result[0].Data.ToArray();
//Update third layer
Layer3.Backward(true, layer3ForwardResult);
layer3ForwardResult[0].ParentFunc = null;
//Perform learning of cDNI for layer 2
Real cDNI2loss = new MeanSquaredError().Evaluate(cDNI2Result, new NdArray(layer2ResultDataSet.Result[0].Grad, cDNI2Result[0].Shape, cDNI2Result[0].BatchCount));
Layer3.Update();
cDNI2.Backward(true, cDNI2Result);
cDNI2.Update();
cDNI2totalLoss += cDNI2loss;
cDNI2totalLossCount++;
NdArray[] layer4ForwardResult = Layer4.Forward(true, layer3ResultDataSet.Result);
Real sumLoss = new SoftmaxCrossEntropy().Evaluate(layer4ForwardResult, layer3ResultDataSet.Label);
Layer4.Backward(true, layer4ForwardResult);
layer4ForwardResult[0].ParentFunc = null;
totalLoss += sumLoss;
totalLossCount++;
Real cDNI3loss = new MeanSquaredError().Evaluate(cDNI3Result, new NdArray(layer3ResultDataSet.Result[0].Grad, cDNI3Result[0].Shape, cDNI3Result[0].BatchCount));
Layer4.Update();
cDNI3.Backward(true, cDNI3Result);
cDNI3.Update();
cDNI3totalLoss += cDNI3loss;
cDNI3totalLossCount++;
RILogManager.Default?.SendDebug("\nbatch count " + i + "/" + TRAIN_DATA_COUNT);
RILogManager.Default?.SendDebug("total loss " + totalLoss / totalLossCount);
RILogManager.Default?.SendDebug("local loss " + sumLoss);
RILogManager.Default?.SendDebug("\ncDNI1 total loss " + cDNI1totalLoss / cDNI1totalLossCount);
RILogManager.Default?.SendDebug("cDNI2 total loss " + cDNI2totalLoss / cDNI2totalLossCount);
RILogManager.Default?.SendDebug("cDNI3 total loss " + cDNI3totalLoss / cDNI3totalLossCount);
RILogManager.Default?.SendDebug("\ncDNI1 local loss " + cDNI1loss);
RILogManager.Default?.SendDebug("cDNI2 local loss " + cDNI2loss);
RILogManager.Default?.SendDebug("cDNI3 local loss " + cDNI3loss);
if (i % 20 == 0)
{
RILogManager.Default?.SendDebug("\nTesting...");
TestDataSet datasetY = mnistData.GetRandomYSet(TEST_DATA_COUNT, 28);
Real accuracy = Trainer.Accuracy(nn, datasetY.Data, datasetY.Label);
RILogManager.Default?.SendDebug("accuracy " + accuracy);
}
}
}
}
public static void Run()
{
// Describe each initial value
Real[,,,] initial_W1 =
{
{ { { 1.0, 0.5, 0.0 }, { 0.5, 0.0, -0.5 }, { 0.0, -0.5, -1.0 } } },
{ { { 0.0, -0.1, 0.1 }, { -0.3, 0.4, 0.7 }, { 0.5, -0.2, 0.2 } } }
};
Real[] initial_b1 = { 0.5, 1.0 };
Real[,,,] initial_W2 =
{
{ { { -0.1, 0.6 }, { 0.3, -0.9 } }, { { 0.7, 0.9 }, { -0.2, -0.3 } } },
{ { { -0.6, -0.1 }, { 0.3, 0.3 } }, { { -0.5, 0.8 }, { 0.9, 0.1 } } }
};
Real[] initial_b2 = { 0.1, 0.9 };
Real[,] initial_W3 =
{
{ 0.5, 0.3, 0.4, 0.2, 0.6, 0.1, 0.4, 0.3 },
{ 0.6, 0.4, 0.9, 0.1, 0.5, 0.2, 0.3, 0.4 }
};
Real[] initial_b3 = { 0.01, 0.02 };
Real[,] initial_W4 = { { 0.8, 0.2 }, { 0.4, 0.6 } };
Real[] initial_b4 = { 0.02, 0.01 };
//Input data
NdArray x = new NdArray(new Real[, , ] {
{
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.2, 0.9, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.9, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.8, 0.5, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.3, 0.3, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.4, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.8, 0.4, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.8, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 }
}
});
//teacher signal
Real[] t = { 0.0, 1.0 };
// If you want to check the contents of a layer, have an instance as a single layer
Convolution2D l2 = new Convolution2D(true, 1, 2, 3, initialW: initial_W1, initialb: initial_b1, name: "l2 Conv2D");
// Write the network configuration in FunctionStack
FunctionStack nn = new FunctionStack("Test5",
l2, //new Convolution2D(1, 2, 3, initialW: initial_W1, initialb: initial_b1),
new ReLU(name: "l2 ReLU"),
//new AveragePooling(2, 2, name: "l2 AVGPooling"),
new MaxPooling(2, 2, name: "l2 MaxPooling"),
new Convolution2D(true, 2, 2, 2, initialW: initial_W2, initialb: initial_b2, name: "l3 Conv2D"),
new ReLU(name: "l3 ReLU"),
//new AveragePooling(2, 2, name: "l3 AVGPooling"),
new MaxPooling(2, 2, name: "l3 MaxPooling"),
new Linear(true, 8, 2, initialW: initial_W3, initialb: initial_b3, name: "l4 Linear"),
new ReLU(name: "l4 ReLU"),
new Linear(true, 2, 2, initialW: initial_W4, initialb: initial_b4, name: "l5 Linear")
);
// If you omit the optimizer declaration, the default SGD(0.1) will be used
// nn.SetOptimizer(new SGD());
// Training conducted
Trainer.Train(nn, x, t, new MeanSquaredError(), false);
// If Update is executed, grad is consumed, so output the value first
RILogManager.Default?.SendDebug("gw1");
RILogManager.Default?.SendDebug(l2.Weight.ToString("Grad"));
RILogManager.Default?.SendDebug("gb1");
RILogManager.Default?.SendDebug(l2.Bias.ToString("Grad"));
//update
nn.Update();
RILogManager.Default?.SendDebug("w1");
RILogManager.Default?.SendDebug(l2.Weight.ToString());
RILogManager.Default?.SendDebug("b1");
RILogManager.Default?.SendDebug(l2.Bias.ToString());
}
public static void Run()
{
_outputStream = File.Create(LogPath);
_logWriter = new HistogramLogWriter(_outputStream);
_logWriter.Write(DateTime.Now);
var recorder = HistogramFactory
.With64BitBucketSize()
?.WithValuesFrom(1)
?.WithValuesUpTo(2345678912345)
?.WithPrecisionOf(3)
?.WithThreadSafeWrites()
?.WithThreadSafeReads()
?.Create();
var accumulatingHistogram = new LongHistogram(2345678912345, 3);
var size = accumulatingHistogram.GetEstimatedFootprintInBytes();
RILogManager.Default?.SendDebug("Histogram size = {0} bytes ({1:F2} MB)", size, size / 1024.0 / 1024.0);
RILogManager.Default?.SendDebug("Recorded latencies [in system clock ticks]");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.None, useCsvFormat: true);
Console.WriteLine();
RILogManager.Default?.SendDebug("Recorded latencies [in usec]");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.TimeStampToMicroseconds, useCsvFormat: true);
Console.WriteLine();
RILogManager.Default?.SendDebug("Recorded latencies [in msec]");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.TimeStampToMilliseconds, useCsvFormat: true);
Console.WriteLine();
RILogManager.Default?.SendDebug("Recorded latencies [in sec]");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.TimeStampToSeconds, useCsvFormat: true);
DocumentResults(accumulatingHistogram, recorder);
RILogManager.Default?.SendDebug("Build Vocabulary.");
DocumentResults(accumulatingHistogram, recorder);
Vocabulary vocabulary = new Vocabulary();
DocumentResults(accumulatingHistogram, recorder);
string trainPath = InternetFileDownloader.Download(DOWNLOAD_URL + TRAIN_FILE, TRAIN_FILE);
DocumentResults(accumulatingHistogram, recorder);
string validPath = InternetFileDownloader.Download(DOWNLOAD_URL + VALID_FILE, VALID_FILE);
DocumentResults(accumulatingHistogram, recorder);
string testPath = InternetFileDownloader.Download(DOWNLOAD_URL + TEST_FILE, TEST_FILE);
DocumentResults(accumulatingHistogram, recorder);
int[] trainData = vocabulary.LoadData(trainPath);
DocumentResults(accumulatingHistogram, recorder);
int[] validData = vocabulary.LoadData(validPath);
DocumentResults(accumulatingHistogram, recorder);
int[] testData = vocabulary.LoadData(testPath);
DocumentResults(accumulatingHistogram, recorder);
int nVocab = vocabulary.Length;
RILogManager.Default?.SendDebug("Network Initializing.");
FunctionStack model = new FunctionStack("Test10",
new EmbedID(nVocab, N_UNITS, name: "l1 EmbedID"),
new Dropout(),
new LSTM(true, N_UNITS, N_UNITS, name: "l2 LSTM"),
new Dropout(),
new LSTM(true, N_UNITS, N_UNITS, name: "l3 LSTM"),
new Dropout(),
new Linear(true, N_UNITS, nVocab, name: "l4 Linear")
);
DocumentResults(accumulatingHistogram, recorder);
// Do not cease at the given threshold, correct the rate by taking the rate from L2Norm of all parameters
GradientClipping gradientClipping = new GradientClipping(threshold: GRAD_CLIP);
SGD sgd = new SGD(learningRate: 1);
model.SetOptimizer(gradientClipping, sgd);
DocumentResults(accumulatingHistogram, recorder);
Real wholeLen = trainData.Length;
int jump = (int)Math.Floor(wholeLen / BATCH_SIZE);
int epoch = 0;
Stack <NdArray[]> backNdArrays = new Stack <NdArray[]>();
RILogManager.Default?.SendDebug("Train Start.");
double dVal;
NdArray x = new NdArray(new[] { 1 }, BATCH_SIZE, (Function)null);
NdArray t = new NdArray(new[] { 1 }, BATCH_SIZE, (Function)null);
for (int i = 0; i < jump * N_EPOCH; i++)
{
for (int j = 0; j < BATCH_SIZE; j++)
{
x.Data[j] = trainData[(int)((jump * j + i) % wholeLen)];
t.Data[j] = trainData[(int)((jump * j + i + 1) % wholeLen)];
}
NdArray[] result = model.Forward(true, x);
Real sumLoss = new SoftmaxCrossEntropy().Evaluate(result, t);
backNdArrays.Push(result);
RILogManager.Default?.SendDebug("[{0}/{1}] Loss: {2}", i + 1, jump, sumLoss);
//Run truncated BPTT
if ((i + 1) % BPROP_LEN == 0)
{
for (int j = 0; backNdArrays.Count > 0; j++)
{
RILogManager.Default?.SendDebug("backward" + backNdArrays.Count);
model.Backward(true, backNdArrays.Pop());
}
model.Update();
model.ResetState();
}
if ((i + 1) % jump == 0)
{
epoch++;
RILogManager.Default?.SendDebug("evaluate");
dVal = Evaluate(model, validData);
RILogManager.Default?.SendDebug($"validation perplexity: {dVal}");
if (epoch >= 6)
{
sgd.LearningRate /= 1.2;
RILogManager.Default?.SendDebug("learning rate =" + sgd.LearningRate);
}
}
DocumentResults(accumulatingHistogram, recorder);
}
RILogManager.Default?.SendDebug("test start");
dVal = Evaluate(model, testData);
RILogManager.Default?.SendDebug("test perplexity:" + dVal);
DocumentResults(accumulatingHistogram, recorder);
_logWriter.Dispose();
_outputStream.Dispose();
RILogManager.Default?.SendDebug("Log contents");
RILogManager.Default?.SendDebug(File.ReadAllText(LogPath));
Console.WriteLine();
RILogManager.Default?.SendDebug("Percentile distribution (values reported in milliseconds)");
accumulatingHistogram.OutputPercentileDistribution(Console.Out, outputValueUnitScalingRatio: OutputScalingFactor.TimeStampToMilliseconds, useCsvFormat: true);
RILogManager.Default?.SendDebug("Mean: " + BytesToString(accumulatingHistogram.GetMean()) + ", StdDev: " +
BytesToString(accumulatingHistogram.GetStdDeviation()));
}
public static void Run()
{
Console.WriteLine("Build Vocabulary.");
Vocabulary vocabulary = new Vocabulary();
string trainPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TRAIN_FILE, TRAIN_FILE);
string validPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + VALID_FILE, VALID_FILE);
string testPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TEST_FILE, TEST_FILE);
int[] trainData = vocabulary.LoadData(trainPath);
int[] validData = vocabulary.LoadData(validPath);
int[] testData = vocabulary.LoadData(testPath);
int nVocab = vocabulary.Length;
Console.WriteLine("Network Initilizing.");
FunctionStack model = new FunctionStack(
new EmbedID(nVocab, N_UNITS, name: "l1 EmbedID"),
new Dropout(),
new LSTM(N_UNITS, N_UNITS, name: "l2 LSTM"),
new Dropout(),
new LSTM(N_UNITS, N_UNITS, name: "l3 LSTM"),
new Dropout(),
new Linear(N_UNITS, nVocab, name: "l4 Linear")
);
//与えられたthresholdで頭打ちではなく、全パラメータのL2Normからレートを取り補正を行う
GradientClipping gradientClipping = new GradientClipping(threshold: GRAD_CLIP);
SGD sgd = new SGD(learningRate: 1);
model.SetOptimizer(gradientClipping, sgd);
Real wholeLen = trainData.Length;
int jump = (int)Math.Floor(wholeLen / BATCH_SIZE);
int epoch = 0;
Stack <NdArray[]> backNdArrays = new Stack <NdArray[]>();
Console.WriteLine("Train Start.");
for (int i = 0; i < jump * N_EPOCH; i++)
{
NdArray x = new NdArray(new[] { 1 }, BATCH_SIZE);
NdArray t = new NdArray(new[] { 1 }, BATCH_SIZE);
for (int j = 0; j < BATCH_SIZE; j++)
{
x.Data[j] = trainData[(int)((jump * j + i) % wholeLen)];
t.Data[j] = trainData[(int)((jump * j + i + 1) % wholeLen)];
}
NdArray[] result = model.Forward(x);
Real sumLoss = new SoftmaxCrossEntropy().Evaluate(result, t);
backNdArrays.Push(result);
Console.WriteLine("[{0}/{1}] Loss: {2}", i + 1, jump, sumLoss);
//Run truncated BPTT
if ((i + 1) % BPROP_LEN == 0)
{
for (int j = 0; backNdArrays.Count > 0; j++)
{
Console.WriteLine("backward" + backNdArrays.Count);
model.Backward(backNdArrays.Pop());
}
model.Update();
model.ResetState();
}
if ((i + 1) % jump == 0)
{
epoch++;
Console.WriteLine("evaluate");
Console.WriteLine("validation perplexity: {0}", Evaluate(model, validData));
if (epoch >= 6)
{
sgd.LearningRate /= 1.2;
Console.WriteLine("learning rate =" + sgd.LearningRate);
}
}
}
Console.WriteLine("test start");
Console.WriteLine("test perplexity:" + Evaluate(model, testData));
}
public static void Run()
{
//Write the configuration of the network you want to read into FunctionStack and adjust the parameters of each function
//Make sure to match name to the variable name of Chainer
FunctionStack nn = new FunctionStack(
new Convolution2D(1, 2, 3, name: "conv1", gpuEnable: true),//Do not forget the GPU flag if necessary
new ReLU(),
new MaxPooling(2, 2),
new Convolution2D(2, 2, 2, name: "conv2", gpuEnable: true),
new ReLU(),
new MaxPooling(2, 2),
new Linear(8, 2, name: "fl3"),
new ReLU(),
new Linear(2, 2, name: "fl4")
);
/* Declaration in Chainer
* class NN (chainer.Chain):
* def __init __ (self):
* super (NN, self).__ init __ (
* conv 1 = L. Convolution 2 D (1, 2, 3),
* conv 2 = L. Convolution 2 D (2, 2, 2),
* fl3 = L. Linear (8, 2),
* fl4 = L. Linear (2, 2)
* )
*
* def __call __ (self, x):
* h_conv 1 = F.relu (self.conv 1 (x))
* h_pool 1 = F.max_pooling - 2 d (h_conv 1, 2)
* h_conv 2 = F.relu (self.conv 2 (h_pool 1))
* h_pool 2 = F.max_pooling - 2 d (h_conv 2, 2)
* h_fc1 = F.relu (self.fl3 (h_pool2))
* y = self.fl 4 (h_fc 1)
* return y
* */
//Read parameters
ChainerModelDataLoader.ModelLoad(MODEL_FILE_PATH, nn);
//Use it as usual
nn.SetOptimizer(new SGD());
//Input data
NdArray x = new NdArray(new Real[, , ] {
{
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.2, 0.9, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.9, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.8, 0.5, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.3, 0.3, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.1, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.4, 0.8, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.8, 0.4, 0.1, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.2, 0.8, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.8, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.1, 0.7, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0 },
{ 0.0, 0.0, 0.0, 0.0, 0.0, 0.3, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 }
}
});
//Teacher signal
Real[] t = { 0.0, 1.0 };
//Training conducted
Trainer.Train(nn, x, t, new MeanSquaredError(), false);
//Evacuate for results display
Convolution2D l2 = (Convolution2D)nn.Functions[0];
//When updating is executed grad will be consumed, so output the value first
Console.WriteLine("gw1");
Console.WriteLine(l2.Weight.ToString("Grad"));
Console.WriteLine("gb1");
Console.WriteLine(l2.Bias.ToString("Grad"));
//update
nn.Update();
Console.WriteLine("w1");
Console.WriteLine(l2.Weight);
Console.WriteLine("b1");
Console.WriteLine(l2.Bias);
}
public static void Run()
{
Console.WriteLine("Build Vocabulary.");
Vocabulary vocabulary = new Vocabulary();
string trainPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TRAIN_FILE, TRAIN_FILE, TRAIN_FILE_HASH);
string testPath = InternetFileDownloader.Donwload(DOWNLOAD_URL + TEST_FILE, TEST_FILE, TEST_FILE_HASH);
int[] trainData = vocabulary.LoadData(trainPath);
int[] testData = vocabulary.LoadData(testPath);
int nVocab = vocabulary.Length;
Console.WriteLine("Done.");
Console.WriteLine("Network Initilizing.");
FunctionStack model = new FunctionStack(
new EmbedID(nVocab, N_UNITS, name: "l1 EmbedID"),
new Linear(N_UNITS, N_UNITS, name: "l2 Linear"),
new TanhActivation("l2 Tanh"),
new Linear(N_UNITS, nVocab, name: "l3 Linear"),
new Softmax("l3 Sonftmax")
);
model.SetOptimizer(new Adam());
List <int> s = new List <int>();
Console.WriteLine("Train Start.");
SoftmaxCrossEntropy softmaxCrossEntropy = new SoftmaxCrossEntropy();
for (int epoch = 0; epoch < TRAINING_EPOCHS; epoch++)
{
for (int pos = 0; pos < trainData.Length; pos++)
{
NdArray h = new NdArray(new Real[N_UNITS]);
int id = trainData[pos];
s.Add(id);
if (id == vocabulary.EosID)
{
Real accumloss = 0;
Stack <NdArray> tmp = new Stack <NdArray>();
for (int i = 0; i < s.Count; i++)
{
int tx = i == s.Count - 1 ? vocabulary.EosID : s[i + 1];
//l1 EmbedID
NdArray l1 = model.Functions[0].Forward(s[i])[0];
//l2 Linear
NdArray l2 = model.Functions[1].Forward(h)[0];
//Add
NdArray xK = l1 + l2;
//l2 Tanh
h = model.Functions[2].Forward(xK)[0];
//l3 Linear
NdArray h2 = model.Functions[3].Forward(h)[0];
Real loss = softmaxCrossEntropy.Evaluate(h2, tx);
tmp.Push(h2);
accumloss += loss;
}
Console.WriteLine(accumloss);
for (int i = 0; i < s.Count; i++)
{
model.Backward(tmp.Pop());
}
model.Update();
s.Clear();
}
if (pos % 100 == 0)
{
Console.WriteLine(pos + "/" + trainData.Length + " finished");
}
}
}
Console.WriteLine("Test Start.");
Real sum = 0;
int wnum = 0;
List <int> ts = new List <int>();
bool unkWord = false;
for (int pos = 0; pos < 1000; pos++)
{
int id = testData[pos];
ts.Add(id);
if (id > trainData.Length)
{
unkWord = true;
}
if (id == vocabulary.EosID)
{
if (!unkWord)
{
Console.WriteLine("pos" + pos);
Console.WriteLine("tsLen" + ts.Count);
Console.WriteLine("sum" + sum);
Console.WriteLine("wnum" + wnum);
sum += CalPs(model, ts);
wnum += ts.Count - 1;
}
else
{
unkWord = false;
}
ts.Clear();
}
}
Console.WriteLine(Math.Pow(2.0, sum / wnum));
}
public static void Run()
{
// Prepare MNIST data
RILogManager.Default?.SendDebug("MNIST Data Loading...");
MnistData mnistData = new MnistData(28);
RILogManager.Default?.SendDebug("Training Start...");
// Write the network configuration in FunctionStack
FunctionStack Layer1 = new FunctionStack("Test11 Layer 1",
new Linear(true, 28 * 28, 256, name: "l1 Linear"),
new BatchNormalization(true, 256, name: "l1 Norm"),
new ReLU(name: "l1 ReLU")
);
FunctionStack Layer2 = new FunctionStack("Test11 Layer 2",
new Linear(true, 256, 256, name: "l2 Linear"),
new BatchNormalization(true, 256, name: "l2 Norm"),
new ReLU(name: "l2 ReLU")
);
FunctionStack Layer3 = new FunctionStack("Test11 Layer 3",
new Linear(true, 256, 256, name: "l3 Linear"),
new BatchNormalization(true, 256, name: "l3 Norm"),
new ReLU(name: "l3 ReLU")
);
FunctionStack Layer4 = new FunctionStack("Test11 Layer 4",
new Linear(true, 256, 10, name: "l4 Linear")
);
// Function stack itself is also stacked as Function
FunctionStack nn = new FunctionStack
("Test11",
Layer1,
Layer2,
Layer3,
Layer4
);
FunctionStack DNI1 = new FunctionStack("Test11 DNI1",
new Linear(true, 256, 1024, name: "DNI1 Linear1"),
new BatchNormalization(true, 1024, name: "DNI1 Norm1"),
new ReLU(name: "DNI1 ReLU1"),
new Linear(true, 1024, 1024, name: "DNI1 Linear2"),
new BatchNormalization(true, 1024, name: "DNI1 Norm2"),
new ReLU(name: "DNI1 ReLU2"),
new Linear(true, 1024, 256, initialW: new Real[1024, 256], name: "DNI1 Linear3")
);
FunctionStack DNI2 = new FunctionStack("Test11 DNI2",
new Linear(true, 256, 1024, name: "DNI2 Linear1"),
new BatchNormalization(true, 1024, name: "DNI2 Norm1"),
new ReLU(name: "DNI2 ReLU1"),
new Linear(true, 1024, 1024, name: "DNI2 Linear2"),
new BatchNormalization(true, 1024, name: "DNI2 Norm2"),
new ReLU(name: "DNI2 ReLU2"),
new Linear(true, 1024, 256, initialW: new Real[1024, 256], name: "DNI2 Linear3")
);
FunctionStack DNI3 = new FunctionStack("Test11 DNI3",
new Linear(true, 256, 1024, name: "DNI3 Linear1"),
new BatchNormalization(true, 1024, name: "DNI3 Norm1"),
new ReLU(name: "DNI3 ReLU1"),
new Linear(true, 1024, 1024, name: "DNI3 Linear2"),
new BatchNormalization(true, 1024, name: "DNI3 Norm2"),
new ReLU(name: "DNI3 ReLU2"),
new Linear(true, 1024, 256, initialW: new Real[1024, 256], name: "DNI3 Linear3")
);
//optimizer
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());
// Three generations learning
for (int epoch = 0; epoch < 20; epoch++)
{
RILogManager.Default?.SendDebug("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;
// how many times to run the batch
for (int i = 1; i < TRAIN_DATA_COUNT + 1; i++)
{
// Get data randomly from the training data
TestDataSet datasetX = mnistData.GetRandomXSet(BATCH_DATA_COUNT, 28, 28);
// Run first tier
NdArray[] layer1ForwardResult = Layer1.Forward(true, datasetX.Data);
// Obtain the slope of the first layer
NdArray[] DNI1Result = DNI1.Forward(true, layer1ForwardResult);
// Apply the slope of the first layer
layer1ForwardResult[0].Grad = DNI1Result[0].Data.ToArray();
// Update first layer
Layer1.Backward(true, layer1ForwardResult);
layer1ForwardResult[0].ParentFunc = null; // Backward was executed and cut off calculation graph
Layer1.Update();
// Run Layer 2
NdArray[] layer2ForwardResult = Layer2.Forward(true, layer1ForwardResult);
// Get the inclination of the second layer
NdArray[] DNI2Result = DNI2.Forward(true, layer2ForwardResult);
// Apply the slope of the second layer
layer2ForwardResult[0].Grad = DNI2Result[0].Data.ToArray();
// Update layer 2
Layer2.Backward(true, layer2ForwardResult);
layer2ForwardResult[0].ParentFunc = null;
// Learn DNI for first tier
Real DNI1loss = new MeanSquaredError().Evaluate(DNI1Result, new NdArray(layer1ForwardResult[0].Grad, DNI1Result[0].Shape, DNI1Result[0].BatchCount));
Layer2.Update();
DNI1.Backward(true, DNI1Result);
DNI1.Update();
DNI1totalLoss += DNI1loss;
DNI1totalLossCount++;
// run layer 3
NdArray[] layer3ForwardResult = Layer3.Forward(true, layer2ForwardResult);
// Get the inclination of the third layer
NdArray[] DNI3Result = DNI3.Forward(true, layer3ForwardResult);
// Apply the slope of the third layer
layer3ForwardResult[0].Grad = DNI3Result[0].Data.ToArray();
// Update layer 3
Layer3.Backward(true, layer3ForwardResult);
layer3ForwardResult[0].ParentFunc = null;
// Run DNI learning for layer 2
Real DNI2loss = new MeanSquaredError().Evaluate(DNI2Result, new NdArray(layer2ForwardResult[0].Grad, DNI2Result[0].Shape, DNI2Result[0].BatchCount));
Layer3.Update();
DNI2.Backward(true, DNI2Result);
DNI2.Update();
DNI2totalLoss += DNI2loss;
DNI2totalLossCount++;
// run layer 4
NdArray[] layer4ForwardResult = Layer4.Forward(true, layer3ForwardResult);
// Obtain the slope of the fourth layer
Real sumLoss = new SoftmaxCrossEntropy().Evaluate(layer4ForwardResult, datasetX.Label);
// Update fourth layer
Layer4.Backward(true, layer4ForwardResult);
layer4ForwardResult[0].ParentFunc = null;
totalLoss += sumLoss;
totalLossCount++;
// Run DNI learning for layer 3
Real DNI3loss = new MeanSquaredError().Evaluate(DNI3Result, new NdArray(layer3ForwardResult[0].Grad, DNI3Result[0].Shape, DNI3Result[0].BatchCount));
Layer4.Update();
DNI3.Backward(true, DNI3Result);
DNI3.Update();
DNI3totalLoss += DNI3loss;
DNI3totalLossCount++;
RILogManager.Default?.SendDebug("batch count " + i + "/" + TRAIN_DATA_COUNT);
RILogManager.Default?.SendDebug("total loss " + totalLoss / totalLossCount);
RILogManager.Default?.SendDebug("local loss " + sumLoss);
RILogManager.Default?.SendDebug("DNI1 total loss " + DNI1totalLoss / DNI1totalLossCount);
RILogManager.Default?.SendDebug("DNI2 total loss " + DNI2totalLoss / DNI2totalLossCount);
RILogManager.Default?.SendDebug("DNI3 total loss " + DNI3totalLoss / DNI3totalLossCount);
RILogManager.Default?.SendDebug("DNI1 local loss " + DNI1loss);
RILogManager.Default?.SendDebug("DNI2 local loss " + DNI2loss);
RILogManager.Default?.SendDebug("DNI3 local loss " + DNI3loss);
// Test the accuracy if you move the batch 20 times
if (i % 20 == 0)
{
RILogManager.Default?.SendDebug("Testing...");
// Get data randomly from test data
TestDataSet datasetY = mnistData.GetRandomYSet(TEST_DATA_COUNT, 28);
// Run test
Real accuracy = Trainer.Accuracy(nn, datasetY.Data, datasetY.Label);
RILogManager.Default?.SendDebug("accuracy " + accuracy);
}
}
}
}