more on cnn layer (wip)

MADEAPPS · Oct 25, 2023 · eafaca2 · eafaca2
1 parent b22e479
commit eafaca2
Show file tree

Hide file tree

Showing 3 changed files with 39 additions and 8 deletions.
diff --git a/newton-4.00/sdk/dBrain/ndBrainLayerConvolutional.cpp b/newton-4.00/sdk/dBrain/ndBrainLayerConvolutional.cpp
@@ -304,11 +304,10 @@ void ndBrainLayerConvolutional::MakePrediction(const ndBrainVector& input, ndBra
 {
 	//m_weights.Mul(input, output);
 	//output.Add(m_bias);
+	ndAssert(input.GetCount() == GetInputSize());
 
-	//output.Set(ndBrainFloat(ndFloat32(0.0f)));
 	//ndInt32 inputStart = 0;
 	//ndInt32 outputStart = 0;
-	ndAssert(0);
 	output.InitGaussianWeights(0.1f);
 //float xxxxx = 0;
 //	for (ndInt32 i = 0; i < m_numberOfKernels; ++i)

diff --git a/newton-4.00/sdk/dBrain/ndBrainLayerConvolutionalMaxPooling.cpp b/newton-4.00/sdk/dBrain/ndBrainLayerConvolutionalMaxPooling.cpp
@@ -24,15 +24,28 @@
 #include "ndBrainLayerConvolutionalMaxPooling.h"
 
 ndBrainLayerConvolutionalMaxPooling::ndBrainLayerConvolutionalMaxPooling(ndInt32 inputWidth, ndInt32 inputHeight, ndInt32 inputDepth)
-	:ndBrainLayerActivation(0)
+	:ndBrainLayerActivation(((inputWidth + 1) / 2) * ((inputHeight + 1) / 2) * inputDepth)
+	,m_index()
+	,m_width(inputWidth)
+	,m_height(inputHeight)
+	,m_channels(inputDepth)
 {
-	ndAssert(0);
+	m_index.SetCount(m_neurons);
 }
 
 ndBrainLayerConvolutionalMaxPooling::ndBrainLayerConvolutionalMaxPooling(const ndBrainLayerConvolutionalMaxPooling& src)
 	:ndBrainLayerActivation(src)
+	,m_index()
+	,m_width(src.m_width)
+	,m_height(src.m_height)
+	,m_channels(src.m_channels)
 {
-	ndAssert(0);
+	m_index.SetCount(src.m_index.GetCount());
+}
+
+ndInt32 ndBrainLayerConvolutionalMaxPooling::GetInputSize() const
+{
+	return m_width * m_height * m_channels;
 }
 
 ndBrainLayer* ndBrainLayerConvolutionalMaxPooling::Clone() const
@@ -47,10 +60,22 @@ const char* ndBrainLayerConvolutionalMaxPooling::GetLabelId() const
 
 void ndBrainLayerConvolutionalMaxPooling::MakePrediction(const ndBrainVector& input, ndBrainVector& output) const
 {
-	ndAssert(0);
+	ndAssert(input.GetCount() == GetInputSize());
+	ndAssert(output.GetCount() == GetOutputSize());
 	//ndAssert(input.GetCount() == output.GetCount());
-	//for (ndInt32 i = input.GetCount() - 1; i >= 0; --i)
-	//{
+	ndInt32 base = 0;
+	for (ndInt32 k = 0; k < m_channels; ++k)
+	{
+		for (ndInt32 i = 0; i < (m_height & -1); i += 2)
+		{
+			for (ndInt32 j = 0; j < (m_width & -1); j += 2)
+			{
+
+			}
+
+			base += m_width * 2;
+		}
+	}
 	//	output[i] = (input[i] > ndBrainFloat(0.0f)) ? input[i] : ndBrainFloat(0.0f);
 	//	ndAssert(ndCheckFloat(output[i]));
 	//}

diff --git a/newton-4.00/sdk/dBrain/ndBrainLayerConvolutionalMaxPooling.h b/newton-4.00/sdk/dBrain/ndBrainLayerConvolutionalMaxPooling.h
@@ -33,9 +33,16 @@ class ndBrainLayerConvolutionalMaxPooling : public ndBrainLayerActivation
 	ndBrainLayer* Clone() const;
 	static ndBrainLayer* Load(const ndBrainLoad* const loadSave);
 
+	virtual ndInt32 GetInputSize() const;
+
 	const char* GetLabelId() const;
 	void MakePrediction(const ndBrainVector& input, ndBrainVector& output) const;
 	void InputDerivative(const ndBrainVector& output, const ndBrainVector& outputDerivative, ndBrainVector& inputDerivative) const;
+
+	mutable ndArray<ndInt32> m_index;
+	ndInt32 m_width;
+	ndInt32 m_height;
+	ndInt32 m_channels;
 };
 
 #endif