public void Train(IntPtr d_source,
IntPtr d_target,
float learningRate,
int threadID,
out float[] prediction,
out float[] loss)
{
if (!ForTraining)
{
throw new Exception("Network was loaded in prediction mode, but asked to train.");
}
lock (TFHelper.DeviceSync[DeviceID])
{
GPU.CopyDeviceToHostPinned(d_source, TensorSource[threadID].Data, BatchSize * (int)BoxDimensions.Elements());
GPU.CopyDeviceToHostPinned(d_target, TensorTarget[threadID].Data, BatchSize * (int)BoxDimensions.Elements());
Marshal.Copy(new[] { learningRate }, 0, TensorLearningRate[threadID].Data, 1);
var Output = RunnerTraining[threadID].Run();
Marshal.Copy(Output[0].Data, ResultPredicted[threadID], 0, BatchSize * (int)BoxDimensions.Elements());
Marshal.Copy(Output[1].Data, ResultLoss[threadID], 0, 1);
prediction = ResultPredicted[threadID];
loss = ResultLoss[threadID];
foreach (var tensor in Output)
{
tensor.Dispose();
}
}
}
public void Encode(IntPtr d_source,
IntPtr d_weight,
int threadID,
//out float[] prediction,
out float[] bottleneck)
{
//if (!ForTraining)
// throw new Exception("Network was loaded in prediction mode, but asked to train.");
GPU.CopyDeviceToHostPinned(d_source, TensorSource[threadID].Data, BatchSize * (int)BoxDimensions.ElementsFFT() * 2);
GPU.CopyDeviceToHostPinned(d_weight, TensorWeightSource[threadID].Data, BatchSize * (int)BoxDimensions.ElementsFFT());
Marshal.Copy(new[] { 0f }, 0, TensorDropoutRate[threadID].Data, 1);
var Output = RunnerEncode[threadID].Run();
//Marshal.Copy(Output[0].Data, ResultPredicted[threadID], 0, BatchSize * (int)BoxDimensions.ElementsFFT() * 2);
Marshal.Copy(Output[0].Data, ResultBottleneck[threadID], 0, BatchSize * BottleneckWidth);
//prediction = ResultPredicted[threadID];
bottleneck = ResultBottleneck[threadID];
foreach (var tensor in Output)
{
tensor.Dispose();
}
}
public float Train(IntPtr d_data, IntPtr d_labels, IntPtr d_weights, float learningRate, int threadID, out long[] argmax, out float[] probability)
{
if (!ForTraining)
{
throw new Exception("Network was loaded in prediction mode, but asked to train.");
}
lock (TFHelper.DeviceSync[DeviceID])
{
GPU.CopyDeviceToHostPinned(d_data, TensorMicTile[threadID].Data, BatchSize * (int)BoxDimensionsTrain.Elements());
GPU.CopyDeviceToHostPinned(d_labels, TensorTrainingLabels[threadID].Data, BatchSize * (int)BoxDimensionsTrain.Elements() * NClasses);
GPU.CopyDeviceToHostPinned(d_weights, TensorTrainingWeights[threadID].Data, BatchSize * (int)BoxDimensionsTrain.Elements());
Marshal.Copy(new[] { learningRate }, 0, TensorLearningRate[threadID].Data, 1);
var Output = RunnerTraining[threadID].Run();
Marshal.Copy(Output[1].Data, ResultArgMax[threadID], 0, BatchSize * (int)BoxDimensionsTrain.Elements());
Marshal.Copy(Output[2].Data, ResultSoftMax[threadID], 0, BatchSize * (int)BoxDimensionsTrain.Elements() * NClasses);
Marshal.Copy(Output[3].Data, ResultLoss[threadID], 0, BatchSize);
argmax = ResultArgMax[threadID];
probability = ResultSoftMax[threadID];
foreach (var tensor in Output)
{
tensor.Dispose();
}
return(MathHelper.Mean(ResultLoss[threadID]));
}
}
public void Train(IntPtr d_source,
IntPtr d_target,
IntPtr d_weightSource,
IntPtr d_weightTarget,
float dropoutRate,
float learningRate,
float orthogonalityRate,
int threadID,
out float[] prediction,
out float[] bottleneck,
out float[] loss,
out float[] lossKL)
{
if (!ForTraining)
{
throw new Exception("Network was loaded in prediction mode, but asked to train.");
}
GPU.CopyDeviceToHostPinned(d_source, TensorSource[threadID].Data, BatchSize * (int)BoxDimensions.ElementsFFT() * 2);
GPU.CopyDeviceToHostPinned(d_target, TensorTarget[threadID].Data, BatchSize * (int)BoxDimensions.ElementsFFT() * 2);
GPU.CopyDeviceToHostPinned(d_weightSource, TensorWeightSource[threadID].Data, BatchSize * (int)BoxDimensions.ElementsFFT());
GPU.CopyDeviceToHostPinned(d_weightTarget, TensorWeightTarget[threadID].Data, BatchSize * (int)BoxDimensions.ElementsFFT());
Marshal.Copy(new[] { dropoutRate }, 0, TensorDropoutRate[threadID].Data, 1);
Marshal.Copy(new[] { learningRate }, 0, TensorLearningRate[threadID].Data, 1);
Marshal.Copy(new[] { orthogonalityRate }, 0, TensorOrthogonalityRate[threadID].Data, 1);
var Output = RunnerTraining[threadID].Run();
Marshal.Copy(Output[0].Data, ResultPredicted[threadID], 0, BatchSize * (int)BoxDimensions.ElementsFFT() * 2);
Marshal.Copy(Output[1].Data, ResultLoss[threadID], 0, 1);
Marshal.Copy(Output[2].Data, ResultLossKL[threadID], 0, 1);
Marshal.Copy(Output[3].Data, ResultBottleneck[threadID], 0, BatchSize * BottleneckWidth);
prediction = ResultPredicted[threadID];
bottleneck = ResultBottleneck[threadID];
loss = ResultLoss[threadID];
lossKL = ResultLossKL[threadID];
foreach (var tensor in Output)
{
tensor.Dispose();
}
}
public void Predict(IntPtr d_data, int threadID, out long[] argmax, out float[] probability)
{
//if (ForTraining)
// throw new Exception("Network was loaded in training mode, but asked to predict.");
GPU.CopyDeviceToHostPinned(d_data, TensorMicTilePredict[threadID].Data, BatchSize * (int)BoxDimensionsPredict.Elements());
var Output = RunnerPrediction[threadID].Run();
Marshal.Copy(Output[0].Data, ResultArgMax[threadID], 0, BatchSize * (int)BoxDimensionsPredict.Elements());
Marshal.Copy(Output[1].Data, ResultSoftMax[threadID], 0, BatchSize * (int)BoxDimensionsPredict.Elements() * 3);
argmax = ResultArgMax[threadID];
probability = ResultSoftMax[threadID];
foreach (var tensor in Output)
{
tensor.Dispose();
}
}
public void Predict(IntPtr d_data, int threadID, out float[] prediction)
{
lock (TFHelper.DeviceSync[DeviceID])
{
//if (ForTraining)
// throw new Exception("Network was loaded in training mode, but asked to predict.");
GPU.CopyDeviceToHostPinned(d_data, TensorSource[threadID].Data, BatchSize * (int)BoxDimensions.Elements());
var Output = RunnerPrediction[threadID].Run();
Marshal.Copy(Output[0].Data, ResultPredicted[threadID], 0, BatchSize * (int)BoxDimensions.Elements());
prediction = ResultPredicted[threadID];
foreach (var tensor in Output)
{
tensor.Dispose();
}
}
}
public void Train(IntPtr d_data, float[] labels, float learningRate, int threadID, out long[] argmax, out float[] probability)
{
if (!ForTraining)
{
throw new Exception("Network was loaded in prediction mode, but asked to train.");
}
GPU.CopyDeviceToHostPinned(d_data, TensorMicTile[threadID].Data, BatchSize * (int)BoxDimensions.Elements());
Marshal.Copy(labels, 0, TensorTrainingLabels[threadID].Data, BatchSize * 2);
Marshal.Copy(new[] { learningRate }, 0, TensorLearningRate[threadID].Data, 1);
var Output = RunnerTraining[threadID].Run();
Marshal.Copy(Output[0].Data, ResultArgMax[threadID], 0, BatchSize);
Marshal.Copy(Output[1].Data, ResultSoftMax[threadID], 0, BatchSize * 2);
argmax = ResultArgMax[threadID];
probability = ResultSoftMax[threadID];
foreach (var tensor in Output)
{
tensor.Dispose();
}
}