//Oh Boy
public JobHandle GradientDescentBackpropigate(NativeSlice2D <float> inputData, NativeSlice2D <float> expectedOutput, int testCaseCount, out float errorSum)
{
Profiler.BeginSample("NetworkEvaluator::GradientDescentBackpropigate");
var resultArray = new NativeArray2D <float>(testCaseCount, OutputLayer.Size);
var feedforwardHandle = Evaluate(inputData, resultArray.Slice(0, testCaseCount), testCaseCount);
//Now we need to use all the jobs created before to evaluate this.
//Compute output error
/* var computeOutputErrorJob = new ErrorEvaluators.QuadraticSigmoidOutputErrorEvaluator()
* {
* Expected = expectedOutput,
* Actuall = resultArray,
* WeightedActivation = OutputLayer.WeightedInput,
* ErrorOut = OutputLayer.Error
* };*/
var computeOutputErrorJob = new ErrorEvaluators.CrossEntropySigmoidOutputErrorEvaluator()
{
Expected = expectedOutput,
Actuall = resultArray.Slice(0, testCaseCount),
WeightedActivation = OutputLayer.SliceWeightedInputs(0, testCaseCount),
ErrorOut = OutputLayer.SliceError(0, testCaseCount)
};
var outputErrorHandle = computeOutputErrorJob.Schedule(OutputLayer.Error.Dimensions.y, 4, feedforwardHandle);
//Convert that output error to the output node gradient
//outputErrorHandle.Complete();
//Perform backpropigation
JobHandle backpropigationHandle = outputErrorHandle;
for (int layerIndex = _layers.Count - 2; layerIndex >= 0; layerIndex--)
{
var targetLayer = _layers[layerIndex];
var nextLayer = _layers[layerIndex + 1];
if (layerIndex != 0)
{
backpropigationHandle = targetLayer.ModelLayer.BackpropigateLayer(targetLayer, nextLayer, testCaseCount, backpropigationHandle);
}
var accumulateGradientOverWeightJob = new ErrorEvaluators.AccumulateGradientOverWeight()
{
PreviousActivation = targetLayer.SliceActivations(0, testCaseCount),
NextError = nextLayer.SliceError(0, testCaseCount),
WeightGradients = nextLayer.WeightGradients
};
backpropigationHandle =
accumulateGradientOverWeightJob.Schedule(nextLayer.WeightGradients.Length, 4,
backpropigationHandle);
}
JobHandle.ScheduleBatchedJobs();
//Update weights (all layers but first)
JobHandle updateNetworkJobHandle = backpropigationHandle;
for (int layerIndex = 1; layerIndex < _layers.Count; layerIndex++)
{
var layer = _layers[layerIndex];
var applyGradientsToWeightsJob = new ErrorEvaluators.ApplyGradientToLayerWeights()
{
TestCount = 1,
LearningRate = LearningRate,
WeightGradients = layer.WeightGradients,
LayerWeights = layer.ModelLayer.Weights
};
updateNetworkJobHandle =
applyGradientsToWeightsJob.Schedule(layer.ModelLayer.Weights.Length, 4, updateNetworkJobHandle);
//updateNetworkJobHandle.Complete();
var applyGradientsToBiasesJob = new ErrorEvaluators.ApplyGradientToLayerBiases()
{
TestCount = 1,
LearningRate = LearningRate,
LayerBiases = layer.ModelLayer.Biases,
LayerErrors = layer.Error.Slice(0, testCaseCount)
};
updateNetworkJobHandle =
applyGradientsToBiasesJob.Schedule(layer.ModelLayer.Biases.Length, 4, updateNetworkJobHandle);
//updateNetworkJobHandle.Complete();
int x = 1;
}
//return error i guess ?
updateNetworkJobHandle.Complete();
//Compute average error sum for test cases
ComputeErrorSum(expectedOutput, testCaseCount, out errorSum);
resultArray.Dispose();
Profiler.EndSample();
return(updateNetworkJobHandle);
}
public virtual NativeSlice2D <float> SliceWeightedInputs(int start, int count)
{
return(WeightedInput.Slice(start, count));
}
public virtual NativeSlice2D <float> SliceError(int start, int count)
{
return(Error.Slice(start, count));
}
public NativeArray <float> WeightGradients; //This does not need to scale by slice size
public virtual NativeSlice2D <float> SliceActivations(int start, int count)
{
return(OutputActivation.Slice(start, count));
}