public override void Update(int iteration, BaseLayer layer)
{
if (DecayRate > 0)
{
LearningRate = LearningRate * (1 / 1 + DecayRate * iteration);
}
float t = iteration + 1;
float lr_t = Convert.ToSingle(LearningRate / Math.Sqrt(1f - Math.Pow(Beta1, t)));
foreach (var item in layer.Params)
{
var param = item.Value;
if (!ms.ContainsKey(param.Name))
{
ms[param.Name] = Tensor.Constant(0, Global.Device, DType.Float32, param.Data.Shape);
us[param.Name] = Tensor.Constant(0, Global.Device, DType.Float32, param.Data.Shape);
var m_t = (Beta1 * ms[param.Name]) + (1 - Beta1) * param.Grad;
var u_t = TOps.Maximum((Beta2 * us[param.Name]), Abs(param.Grad));
param.Data = param.Data - lr_t * m_t / (u_t + EPSILON);
ms[param.Name] = m_t;
us[param.Name] = u_t;
param.ApplyConstraint();
}
}
}
public void Exponential(NDArray result, int?seed, float lambda)
{
using (var cpuCopy = new NDArray(cpuAllocator, result.ElementType, result.Shape))
{
cpuRandom.Exponential(cpuCopy, seed, lambda);
TOps.Copy(result, cpuCopy);
}
}
public void Normal(NDArray result, int?seed, float mean, float stdv)
{
using (var cpuCopy = new NDArray(cpuAllocator, result.ElementType, result.Shape))
{
cpuRandom.Normal(cpuCopy, seed, mean, stdv);
TOps.Copy(result, cpuCopy);
}
}
public void Uniform(NDArray result, int?seed, float min, float max)
{
using (var cpuCopy = new NDArray(cpuAllocator, result.ElementType, result.Shape))
{
cpuRandom.Uniform(cpuCopy, seed, min, max);
TOps.Copy(result, cpuCopy);
}
}
public void Bernoulli(NDArray result, int?seed, float p)
{
using (var cpuCopy = new NDArray(cpuAllocator, result.ElementType, result.Shape))
{
cpuRandom.Bernoulli(cpuCopy, seed, p);
TOps.Copy(result, cpuCopy);
}
}
public void Cauchy(NDArray result, int?seed, float median, float sigma)
{
using (var cpuCopy = new NDArray(cpuAllocator, result.ElementType, result.Shape))
{
cpuRandom.Cauchy(cpuCopy, seed, median, sigma);
TOps.Copy(result, cpuCopy);
}
}
public void Geometric(Tensor result, int?seed, float p)
{
using (var cpuCopy = new Tensor(cpuAllocator, result.ElementType, result.Shape))
{
cpuRandom.Geometric(cpuCopy, seed, p);
TOps.Copy(result, cpuCopy);
}
}
public void LogNormal(Tensor result, int?seed, float mean, float stdv)
{
using (var cpuCopy = new Tensor(cpuAllocator, result.ElementType, result.Shape))
{
cpuRandom.LogNormal(cpuCopy, seed, mean, stdv);
TOps.Copy(result, cpuCopy);
}
}
public void Fit(IFrameIter train, int epochs, int batchSize, IFrameIter val = null)
{
DateTime start = DateTime.Now;
List <float> train_losses = new List <float>();
List <float> train_metrics = new List <float>();
List <float> val_losses = new List <float>();
List <float> val_metrics = new List <float>();
train.SetBatchSize(batchSize);
for (int iteration = 1; iteration <= epochs; iteration++)
{
train.Reset();
while (train.Next())
{
var(x, y) = train.GetBatch();
using (Variable pred = Forward(x))
using (Tensor lossVal = LossFn.Call(pred.Data, y))
using (Tensor grad = LossFn.CalcGrad(pred.Data, y))
using (Tensor reg_loss = ApplyRegularizer(lossVal))
{
//var metricVal = MetricFn.Call(pred.Data, y);
train_losses.Add(reg_loss.TVar().ToScalar().Evaluate());
//train_metrics.Add(metricVal.ToScalar().Evaluate());
Backward(grad);
ApplyDeltaRegularizer();
foreach (var layer in Layers)
{
OptimizerFn.Update(iteration, layer);
}
}
x.Dispose();
y.Dispose();
}
if (val != null)
{
while (val.Next())
{
var(x, y) = val.GetBatch();
var pred = Forward(x);
var lossVal = LossFn.Call(pred.Data, y);
var metricVal = MetricFn.Call(pred.Data, y);
val_losses.Add(TOps.MeanF(lossVal));
val_metrics.Add(TOps.MeanF(metricVal));
}
}
Console.WriteLine("Epoch: {0}, Loss: {1}", iteration, train_losses.Average());
}
}
public override void Update(int iteration, BaseLayer layer)
{
if (DecayRate > 0)
{
LearningRate = LearningRate * (1 / 1 + DecayRate * iteration);
}
float t = iteration + 1;
float lr_t = Convert.ToSingle(LearningRate * Math.Sqrt(1f - Math.Pow(Beta2, t)) / (1f - Math.Pow(Beta1, t)));
foreach (var item in layer.Params)
{
var param = item.Value;
if (!ms.ContainsKey(param.Name))
{
ms[param.Name] = Tensor.Constant(0, Global.Device, DType.Float32, param.Data.Shape);
vs[param.Name] = Tensor.Constant(0, Global.Device, DType.Float32, param.Data.Shape);
if (AmsGrad)
{
vhats[param.Name] = Tensor.Constant(0, Global.Device, DType.Float32, param.Data.Shape);
}
else
{
vhats[param.Name] = Tensor.Constant(0, Global.Device, DType.Float32, 1);
}
var m_t = (Beta1 * ms[param.Name]) + (1 - Beta1) * param.Grad;
var v_t = (Beta2 * vs[param.Name]) + (1 - Beta2) * Square(param.Grad);
if (AmsGrad)
{
Tensor vhat_t = TOps.Maximum(vhats[param.Name], v_t);
param.Data = param.Data - lr_t * m_t / (Sqrt(vhat_t) + EPSILON);
vhats[param.Name] = vhat_t;
}
else
{
param.Data = param.Data - lr_t * m_t / (Sqrt(v_t) + EPSILON);
}
ms[param.Name] = m_t;
vs[param.Name] = v_t;
param.ApplyConstraint();
}
}
}
static void Main(string[] args)
{
Global.UseGpu();
Tensor x = Tensor.FromArray(Global.Device, new float[] { 1, 2, 3, 4, 5, 6, 7, 8, 9 });
x = x.Reshape(3, 3);
var result = TOps.Diag(x);
result.Print();
string datasetFolder = @"C:\dataset\MNIST";
bool useDenseModel = false;
var((trainX, trainY), (valX, valY)) = MNISTParser.LoadDataSet(datasetFolder, trainCount: 60000, testCount: 10000, flatten: useDenseModel);
Console.WriteLine("Train and Test data loaded");
DataFrameIter trainIter = new DataFrameIter(trainX, trainY);
DataFrameIter valIter = new DataFrameIter(valX, valY);
Sequential model = null;
if (useDenseModel)
{
model = BuildFCModel();
}
else
{
model = BuildConvModel();
}
model.Compile(OptimizerType.Adam, LossType.CategorialCrossEntropy, MetricType.Accuracy);
Console.WriteLine("Model compiled.. initiating training");
model.EpochEnd += Model_EpochEnd;
model.Train(trainIter, 10, 32, valIter);
Console.ReadLine();
}
