+ +

#include <CaffeParser.hpp>

+

+Inheritance diagram for CaffeParserBase:

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + +

+Public Member Functions
virtual armnn::INetworkPtr	CreateNetworkFromTextFile (const char *graphFile, const std::map< std::string, armnn::TensorShape > &inputShapes, const std::vector< std::string > &requestedOutputs) override
	Create the network from a protobuf text file on disk. More...

virtual armnn::INetworkPtr	CreateNetworkFromString (const char *protoText, const std::map< std::string, armnn::TensorShape > &inputShapes, const std::vector< std::string > &requestedOutputs) override
	Creates the network directly from protobuf text in a string. Useful for debugging/testing. More...

virtual BindingPointInfo	GetNetworkInputBindingInfo (const std::string &name) const override
	Retrieves binding info (layer id and tensor info) for the network input identified by the given layer name. More...

virtual BindingPointInfo	GetNetworkOutputBindingInfo (const std::string &name) const override
	Retrieves binding info (layer id and tensor info) for the network output identified by the given layer name. More...

	CaffeParserBase ()

Public Member Functions inherited from ICaffeParser
virtual armnn::INetworkPtr	CreateNetworkFromBinaryFile (const char *graphFile, const std::map< std::string, armnn::TensorShape > &inputShapes, const std::vector< std::string > &requestedOutputs)=0
	Create the network from a protobuf binary file on the disk. More...

+ + + +

+Protected Types
using	OperationParsingFunction = void(CaffeParserBase::*)(const caffe::LayerParameter &layerParam)

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +

+Protected Member Functions
armnn::TensorInfo	BlobShapeToTensorInfo (const caffe::BlobShape &blobShape) const
	Converts Caffe's protobuf tensor shape format to ArmNN's. More...

void	TrackInputBinding (armnn::IConnectableLayer *layer, armnn::LayerBindingId id, const armnn::TensorInfo &tensorInfo)

void	TrackOutputBinding (armnn::IConnectableLayer *layer, armnn::LayerBindingId id, const armnn::TensorInfo &tensorInfo)

void	SetArmnnOutputSlotForCaffeTop (const std::string &caffeTopName, armnn::IOutputSlot &armnnOutputSlot)

armnn::IOutputSlot &	GetArmnnOutputSlotForCaffeTop (const std::string &caffeTopName) const
	Retrieves the Armnn IOutputSlot representing the given Caffe top. More...

void	Cleanup ()

armnn::INetworkPtr	CreateNetworkFromNetParameter (caffe::NetParameter &netParam, const std::map< std::string, armnn::TensorShape > &inputShapes, const std::vector< std::string > &requestedOutputs)
	Parses a NetParameter loaded into memory from one of the other CreateNetwork*. More...

void	LoadNetParam (caffe::NetParameter &netParameter)
	does the actual conversion from caffe::NetParameter to armnn::INetwork More...

std::vector< const caffe::LayerParameter * >	GetInputs (const caffe::LayerParameter &layerParam)
	Find the Caffe layers listed as inputs (bottoms) for a given layer. More...

void	ResolveInPlaceLayers (caffe::NetParameter &netParameter)
	Modifies the Caffe network to replace "in-place" layers (whose top() and bottom() are both the same) with regular layers. More...


void	ParseInputLayer (const caffe::LayerParameter &layerParam)
	Adds an armnn layer to m_Network given a Caffe LayerParameter of the correct type and is responsible for recording any newly created IOutputSlots using SetArmnnOutputSlotForCaffeTop(). More...

void	ParseConvLayer (const caffe::LayerParameter &layerParam)

void	ParsePoolingLayer (const caffe::LayerParameter &layerParam)

void	ParseReluLayer (const caffe::LayerParameter &layerParam)

void	ParseLRNLayer (const caffe::LayerParameter &layerParam)

void	ParseInnerProductLayer (const caffe::LayerParameter &layerParam)

void	ParseSoftmaxLayer (const caffe::LayerParameter &layerParam)

void	ParseEltwiseLayer (const caffe::LayerParameter &layerParam)

void	ParseConcatLayer (const caffe::LayerParameter &layerParam)

void	ParseBatchNormLayer (const caffe::LayerParameter &layerParam)

void	ParseScaleLayer (const caffe::LayerParameter &layerParam)

void	ParseSplitLayer (const caffe::LayerParameter &layerParam)

void	ParseDropoutLayer (const caffe::LayerParameter &layerParam)


void	AddConvLayerWithSplits (const caffe::LayerParameter &layerParam, const armnn::Convolution2dDescriptor &desc, unsigned int kernelW, unsigned int kernelH)
	ParseConv may use these helpers depending on the group parameter. More...

void	AddConvLayerWithDepthwiseConv (const caffe::LayerParameter &layerParam, const armnn::Convolution2dDescriptor &desc, unsigned int kernelW, unsigned int kernelH)

Protected Member Functions inherited from ICaffeParser
virtual	~ICaffeParser ()

+ + + + + +

+Static Protected Member Functions
static void	TrackBindingPoint (armnn::IConnectableLayer layer, armnn::LayerBindingId id, const armnn::TensorInfo &tensorInfo, const char bindingPointDesc, std::unordered_map< std::string, BindingPointInfo > &nameToBindingInfo)

static std::pair< armnn::LayerBindingId, armnn::TensorInfo >	GetBindingInfo (const std::string &layerName, const char *bindingPointDesc, const std::unordered_map< std::string, BindingPointInfo > &bindingInfos)

+ + + + + + + + + + + + + + + + + + +

+Protected Attributes
std::unordered_map< std::string, BindingPointInfo >	m_NetworkInputsBindingInfo
	maps input layer names to their corresponding ids and tensor infos More...

std::unordered_map< std::string, BindingPointInfo >	m_NetworkOutputsBindingInfo
	maps output layer names to their corresponding ids and tensor infos More...

armnn::INetworkPtr	m_Network

std::map< std::string, armnn::TensorShape >	m_InputShapes

std::unordered_map< std::string, armnn::IOutputSlot * >	m_ArmnnOutputSlotForCaffeTop
	As we add armnn layers we store the armnn IOutputSlot which corresponds to the Caffe tops. More...

std::vector< std::string >	m_RequestedOutputs

std::map< std::string, const caffe::LayerParameter * >	m_CaffeLayersByTopName

+ + + + +

+Static Protected Attributes
static const std::map< std::string, OperationParsingFunction >	ms_CaffeLayerNameToParsingFunctions
	Maps Caffe layer names to parsing member functions. More...

+ + + + + + + + +

+Additional Inherited Members
Static Public Member Functions inherited from ICaffeParser
static ICaffeParser *	CreateRaw ()

static ICaffeParserPtr	Create ()

static void	Destroy (ICaffeParser *parser)

+

Detailed Description

+

Definition at line 26 of file CaffeParser.hpp.

+

Member Typedef Documentation

+ +

◆ OperationParsingFunction

+ +

+

+ + + + + +

+ + + + +

using OperationParsingFunction = void(CaffeParserBase::*)(const caffe::LayerParameter& layerParam)

+

+protected

+

+ +

Definition at line 115 of file CaffeParser.hpp.

+ +

+

Constructor & Destructor Documentation

+ +

◆ CaffeParserBase()

+ +

+

+ + + + + + + +

CaffeParserBase ( )

+

+ +

Definition at line 274 of file CaffeParser.cpp.

+

     : m_Network(nullptr, nullptr)
 {
 
 }

+

+

Member Function Documentation

+ +

◆ AddConvLayerWithDepthwiseConv()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +

void AddConvLayerWithDepthwiseConv	(	const caffe::LayerParameter &	layerParam,
		const armnn::Convolution2dDescriptor &	desc,
		unsigned int	kernelW,
		unsigned int	kernelH
	)

+

+protected

+

+ +

Definition at line 611 of file CaffeParser.cpp.

+ +

References CaffeParserBase::BlobShapeToTensorInfo(), CHECK_LOCATION, IOutputSlot::Connect(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), armnnUtils::GetTensorInfo(), Convolution2dDescriptor::m_BiasEnabled, CaffeParserBase::m_Network, Convolution2dDescriptor::m_PadBottom, Convolution2dDescriptor::m_PadLeft, DepthwiseConvolution2dDescriptor::m_PadLeft, Convolution2dDescriptor::m_PadRight, Convolution2dDescriptor::m_PadTop, Convolution2dDescriptor::m_StrideX, Convolution2dDescriptor::m_StrideY, armnn::numeric_cast(), CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), and armnnCaffeParser::TensorDescToBlobShape().

+ +

Referenced by CaffeParserBase::ParseConvLayer().

+

 {
     BOOST_ASSERT(layerParam.type() == "Convolution");
     ValidateNumInputsOutputs(layerParam, 1, 1);
 
     ConvolutionParameter convParam  = layerParam.convolution_param();
     BlobShape inputShape = TensorDescToBlobShape(GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo());
 
     DepthwiseConvolution2dDescriptor desc;
     desc.m_PadLeft      = convDesc.m_PadLeft;
     desc.m_PadRight     = convDesc.m_PadRight;
     desc.m_PadTop       = convDesc.m_PadTop;
     desc.m_PadBottom    = convDesc.m_PadBottom;
     desc.m_StrideX      = convDesc.m_StrideX;
     desc.m_StrideY      = convDesc.m_StrideY;
     desc.m_BiasEnabled  = convDesc.m_BiasEnabled;
 
     unsigned int numFilters = convParam.num_output();
 
     BlobShape outputShape;
     outputShape.add_dim(0);
     outputShape.set_dim(0, inputShape.dim(0));
     outputShape.add_dim(1);
     outputShape.set_dim(1, numFilters);
     outputShape.add_dim(2);
     outputShape.set_dim(
         2, (static_cast<int>(
                 static_cast<float>(inputShape.dim(2) + 2 * desc.m_PadBottom - kernelH) /
                 static_cast<float>(desc.m_StrideY)) + 1));
     outputShape.add_dim(3);
     outputShape.set_dim(
         3, (static_cast<int>(
                 static_cast<float>(inputShape.dim(3) + 2 * desc.m_PadRight - kernelW) /
                 static_cast<float>(desc.m_StrideX)) + 1));
 
     // Load the weight data
     size_t allWeightsSize = boost::numeric_cast<size_t>(inputShape.dim(1) * kernelH * kernelW);
     vector<float> weightData(allWeightsSize);
 
     GetDataFromBlob(layerParam, weightData, 0);
 
     // depth multiplier will be 1 for the depthwise convolution
     const unsigned int weightDimSizes[4] = {
         static_cast<unsigned int>(1),                 // depth multiplier
         static_cast<unsigned int>(inputShape.dim(1)), // #channels
         kernelH,
         kernelW};
 
     armnn::IConnectableLayer* returnLayer = nullptr;
     ConstTensor weights(TensorInfo(4, weightDimSizes, DataType::Float32), weightData.data());
     Optional<ConstTensor> optionalBiases;
     vector<float> biasData;
     if (desc.m_BiasEnabled)
     {
         TensorInfo biasInfo;
 
         biasData.resize(boost::numeric_cast<size_t>(outputShape.dim(1)), 1.f);
         GetDataFromBlob(layerParam, biasData, 1);
 
         const unsigned int biasDimSizes[1] = {static_cast<unsigned int>(outputShape.dim(1))};
         biasInfo = TensorInfo(1, biasDimSizes, DataType::Float32);
 
         ConstTensor biases(biasInfo, biasData.data());
         optionalBiases = Optional<ConstTensor>(biases);
     }
     returnLayer = m_Network->AddDepthwiseConvolution2dLayer(desc,
                                                             weights,
                                                             optionalBiases,
                                                             layerParam.name().c_str());
 
     if (!returnLayer)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Failed to create depthwise convolution layer. "
                     "Layer=%1% #filters=%2% %3%") %
                     layerParam.name() %
                     numFilters %
                     CHECK_LOCATION().AsString()));
     }
     armnn::IOutputSlot& inputConnection = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0));
     inputConnection.Connect(returnLayer->GetInputSlot(0));
     returnLayer->GetOutputSlot(0).SetTensorInfo(BlobShapeToTensorInfo(outputShape));
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), returnLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ AddConvLayerWithSplits()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +

void AddConvLayerWithSplits	(	const caffe::LayerParameter &	layerParam,
		const armnn::Convolution2dDescriptor &	desc,
		unsigned int	kernelW,
		unsigned int	kernelH
	)

+

+protected

+

+ +

ParseConv may use these helpers depending on the group parameter.

+ +

Definition at line 419 of file CaffeParser.cpp.

+ +

References CaffeParserBase::BlobShapeToTensorInfo(), CHECK_LOCATION, IOutputSlot::Connect(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), IConnectableLayer::GetInputSlot(), IConnectableLayer::GetNumOutputSlots(), IConnectableLayer::GetOutputSlot(), armnnUtils::GetTensorInfo(), Convolution2dDescriptor::m_BiasEnabled, CaffeParserBase::m_Network, Convolution2dDescriptor::m_PadBottom, Convolution2dDescriptor::m_PadRight, Convolution2dDescriptor::m_StrideX, Convolution2dDescriptor::m_StrideY, armnn::numeric_cast(), CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), IOutputSlot::SetTensorInfo(), OriginsDescriptor::SetViewOriginCoord(), ViewsDescriptor::SetViewOriginCoord(), ViewsDescriptor::SetViewSize(), and armnnCaffeParser::TensorDescToBlobShape().

+ +

Referenced by CaffeParserBase::ParseConvLayer().

+

 {
     BOOST_ASSERT(layerParam.type() == "Convolution");
     ValidateNumInputsOutputs(layerParam, 1, 1);
 
     ConvolutionParameter convParam = layerParam.convolution_param();
     BlobShape inputShape = TensorDescToBlobShape(GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo());
     const unsigned int numGroups = convParam.has_group() ? convParam.group() : 1;
 
     // asusme these were already verified by the caller ParseConvLayer() function
     BOOST_ASSERT(numGroups < inputShape.dim(1));
     BOOST_ASSERT(numGroups > 1);
 
     // Handle grouping
     armnn::IOutputSlot& inputConnection = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0));
 
     vector<string> convLayerNames(numGroups);
     vector<armnn::IConnectableLayer*> convLayers(numGroups);
     convLayerNames[0] = layerParam.name();
 
     // This convolution is to be applied to chunks of the input data so add a splitter layer
 
     // Redirect the convolution input to the splitter
     unsigned int splitterDimSizes[4] = {static_cast<unsigned int>(inputShape.dim(0)),
                                         static_cast<unsigned int>(inputShape.dim(1)),
                                         static_cast<unsigned int>(inputShape.dim(2)),
                                         static_cast<unsigned int>(inputShape.dim(3))};
 
     // Split dimension 1 of the splitter output shape and conv input shapes
     // according to the number of groups
 
     splitterDimSizes[1] /= numGroups;
     inputShape.set_dim(1, splitterDimSizes[1]);
 
     // This is used to describe how the input is to be split
     ViewsDescriptor splitterDesc(numGroups);
 
     // Create an output node for each group, giving each a unique name
     for (unsigned int g = 0; g < numGroups; ++g)
     {
         // Work out the names of the splitter layers child convolutions
         stringstream ss;
         ss << layerParam.name() << "_" << g;
         convLayerNames[g] = ss.str();
 
         splitterDesc.SetViewOriginCoord(g, 1, splitterDimSizes[1] * g);
 
         // Set the size of the views.
         for (unsigned int dimIdx=0; dimIdx < 4; dimIdx++)
         {
             splitterDesc.SetViewSize(g, dimIdx, splitterDimSizes[dimIdx]);
         }
     }
 
     const std::string splitterLayerName = std::string("splitter_") + layerParam.bottom(0);
     armnn::IConnectableLayer* splitterLayer = m_Network->AddSplitterLayer(splitterDesc, splitterLayerName.c_str());
 
     inputConnection.Connect(splitterLayer->GetInputSlot(0));
     for (unsigned int i = 0; i < splitterLayer->GetNumOutputSlots(); i++)
     {
         splitterLayer->GetOutputSlot(i).SetTensorInfo(BlobShapeToTensorInfo(inputShape));
     }
 
     unsigned int numFilters = convParam.num_output();
 
     // Populates convolution output tensor descriptor dimensions.
     BlobShape outputShape;
     outputShape.add_dim(0);
     outputShape.set_dim(0, inputShape.dim(0));
     outputShape.add_dim(1);
     // Ensures that dimension 1 of the convolution output is split according to the number of groups.
     outputShape.set_dim(1, numFilters / numGroups);
     outputShape.add_dim(2);
     outputShape.set_dim(
         2, (static_cast<int>(
                 static_cast<float>(inputShape.dim(2) + 2 * desc.m_PadBottom - kernelH) /
                 static_cast<float>(desc.m_StrideY)) + 1));
     outputShape.add_dim(3);
     outputShape.set_dim(
         3, (static_cast<int>(
                 static_cast<float>(inputShape.dim(3) + 2 * desc.m_PadRight - kernelW) /
                 static_cast<float>(desc.m_StrideX)) + 1));
 
     // Load the weight data for ALL groups
     vector<float> weightData(boost::numeric_cast<size_t>(numGroups *
                                                          inputShape.dim(1) *  // number of input channels
                                                          outputShape.dim(1) * // number of output channels
                                                          kernelH *
                                                          kernelW));
     GetDataFromBlob(layerParam, weightData, 0);
 
     const unsigned int weightDimSizes[4] = {
         static_cast<unsigned int>(outputShape.dim(1)),
         static_cast<unsigned int>(inputShape.dim(1)),
         kernelH,
         kernelW};
 
     TensorInfo biasInfo;
     vector<float> biasData;
 
     if (desc.m_BiasEnabled)
     {
         biasData.resize(boost::numeric_cast<size_t>(numGroups * outputShape.dim(1)), 1.f);
         GetDataFromBlob(layerParam, biasData, 1);
 
         const unsigned int biasDimSizes[1] = {static_cast<unsigned int>(outputShape.dim(1))};
         biasInfo = TensorInfo(1, biasDimSizes, DataType::Float32);
     }
 
     const unsigned int numWeightsPerGroup = boost::numeric_cast<unsigned int>(weightData.size()) / numGroups;
     const unsigned int numBiasesPerGroup  = boost::numeric_cast<unsigned int>(biasData.size()) / numGroups;
 
     for (unsigned int g = 0; g < numGroups; ++g)
     {
         // Sets the slot index, group 0 should be connected to the 0th output of the splitter
         // group 1 should be connected to the 1st output of the splitter.
 
         // Pulls out the weights for this group from that loaded from the model file earlier.
         ConstTensor weights(TensorInfo(4, weightDimSizes, DataType::Float32),
                             weightData.data() + numWeightsPerGroup * g);
 
         IConnectableLayer* convLayer = nullptr;
         Optional<ConstTensor> optionalBiases;
         if (desc.m_BiasEnabled)
         {
             // Pulls out the biases for this group from that loaded from the model file earlier.
             ConstTensor biases(biasInfo, biasData.data() + numBiasesPerGroup * g);
             optionalBiases = Optional<ConstTensor>(biases);
         }
         convLayer = m_Network->AddConvolution2dLayer(desc,
                                                      weights,
                                                      optionalBiases,
                                                      convLayerNames[g].c_str());
         convLayers[g] = convLayer;
 
         // If we have more than one group then the input to the nth convolution the splitter layer's nth output,
         // otherwise it's the regular input to this layer.
         armnn::IOutputSlot& splitterInputConnection =
             splitterLayer ? splitterLayer->GetOutputSlot(g) : inputConnection;
         splitterInputConnection.Connect(convLayer->GetInputSlot(0));
         convLayer->GetOutputSlot(0).SetTensorInfo(BlobShapeToTensorInfo(outputShape));
     }
 
     // If the convolution was performed in chunks, add a layer to concatenate the results
 
     // The merge input shape matches that of the convolution output
     unsigned int concatDimSizes[4] = {static_cast<unsigned int>(outputShape.dim(0)),
                                       static_cast<unsigned int>(outputShape.dim(1)),
                                       static_cast<unsigned int>(outputShape.dim(2)),
                                       static_cast<unsigned int>(outputShape.dim(3))};
 
     // This is used to describe how the input is to be concatenated
     OriginsDescriptor concatDesc(numGroups);
 
     // Now create an input node for each group, using the name from
     // the output of the corresponding convolution
     for (unsigned int g = 0; g < numGroups; ++g)
     {
         concatDesc.SetViewOriginCoord(g, 1, concatDimSizes[1] * g);
     }
 
     // Make sure the output from the concat is the correct size to hold the data for all groups
     concatDimSizes[1] *= numGroups;
     outputShape.set_dim(1, concatDimSizes[1]);
 
     // Finally add the concat layer
     IConnectableLayer* concatLayer = m_Network->AddConcatLayer(concatDesc, layerParam.name().c_str());
 
     if (!concatLayer)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Failed to create final concat layer for Split+Convolution+Concat. "
                     "Layer=%1% #groups=%2% #filters=%3% %4%") %
                     layerParam.name() %
                     numGroups %
                     numFilters %
                     CHECK_LOCATION().AsString()));
     }
 
     for (unsigned int g = 0; g < numGroups; ++g)
     {
         convLayers[g]->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(g));
     }
     concatLayer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo(4, concatDimSizes, DataType::Float32));
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), concatLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ BlobShapeToTensorInfo()

+ +

+

+ + + + + +

+ + + + + + + + +

TensorInfo BlobShapeToTensorInfo ( const caffe::BlobShape & blobShape ) const

+

+protected

+

+ +

Converts Caffe's protobuf tensor shape format to ArmNN's.

+ +

Definition at line 314 of file CaffeParser.cpp.

+ +

Referenced by CaffeParserBase::AddConvLayerWithDepthwiseConv(), CaffeParserBase::AddConvLayerWithSplits(), CaffeParserBase::ParseConvLayer(), and CaffeParserBase::ParseInputLayer().

+

 {
     std::vector<unsigned int> shape;
     for (int j = 0; j < blobShape.dim_size(); ++j)
     {
         shape.push_back(static_cast<unsigned int>(blobShape.dim(j)));
     }
 
     return TensorInfo(boost::numeric_cast<unsigned int>(shape.size()), shape.data(), DataType::Float32);
 }

+

+

+ +

◆ Cleanup()

+ +

+

+ + + + + +

+ + + + + + + +

void Cleanup ( )

+

+protected

+

+ +

Definition at line 1860 of file CaffeParser.cpp.

+ +

References CaffeParserBase::m_ArmnnOutputSlotForCaffeTop, CaffeParserBase::m_CaffeLayersByTopName, CaffeParserBase::m_InputShapes, and CaffeParserBase::m_RequestedOutputs.

+ +

Referenced by RecordByRecordCaffeParser::CreateNetworkFromBinaryFile(), and CaffeParserBase::CreateNetworkFromNetParameter().

+

                               {
     // cleanup, in case we reuse this parser
     m_InputShapes.clear();
     m_RequestedOutputs.clear();
     m_ArmnnOutputSlotForCaffeTop.clear();
     // NOTE: when we get the text/string format
     //       optimised for memory then this data structure can
     //       also move to the CaffeParser class
     m_CaffeLayersByTopName.clear();
 }

+
+
+
+

+

+ +

◆ CreateNetworkFromNetParameter()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + + + + + +

INetworkPtr CreateNetworkFromNetParameter	(	caffe::NetParameter &	netParam,
		const std::map< std::string, armnn::TensorShape > &	inputShapes,
		const std::vector< std::string > &	requestedOutputs
	)

+

+protected

+

+ +

Parses a NetParameter loaded into memory from one of the other CreateNetwork*.

+ +

Definition at line 1829 of file CaffeParser.cpp.

+ +

References CaffeParserBase::Cleanup(), INetwork::Create(), CaffeParserBase::LoadNetParam(), CaffeParserBase::m_InputShapes, CaffeParserBase::m_Network, CaffeParserBase::m_NetworkInputsBindingInfo, CaffeParserBase::m_NetworkOutputsBindingInfo, and CaffeParserBase::m_RequestedOutputs.

+ +

Referenced by CaffeParser::CreateNetworkFromBinaryFile(), CaffeParserBase::CreateNetworkFromString(), and CaffeParserBase::CreateNetworkFromTextFile().

+

 {
     m_NetworkInputsBindingInfo.clear();
     m_NetworkOutputsBindingInfo.clear();
 
     m_Network = INetwork::Create();
 
     m_InputShapes = inputShapes;
     if (requestedOutputs.size() == 0)
     {
         throw ParseException("requestedOutputs must have at least one entry");
     }
     m_RequestedOutputs = requestedOutputs;
 
     try
     {
         LoadNetParam(netParam);
     }
     catch (const ParseException& e)
     {
         Cleanup();
         throw e;
     }
 
     Cleanup();
 
     return move(m_Network);
 }

+
+
+
+
+
+
+
+
+

+

+ +

◆ CreateNetworkFromString()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + + + + + +

INetworkPtr CreateNetworkFromString	(	const char *	protoText,
		const std::map< std::string, armnn::TensorShape > &	inputShapes,
		const std::vector< std::string > &	requestedOutputs
	)

+

+overridevirtual

+

+ +

Creates the network directly from protobuf text in a string. Useful for debugging/testing.

+ +

Implements ICaffeParser.

+ +

Definition at line 1769 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, and CaffeParserBase::CreateNetworkFromNetParameter().

+

 {
     // Parses the string into a message.
     NetParameter netParam;
     bool         success = google::protobuf::TextFormat::ParseFromString(protoText, &netParam);
 
     if (!success)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Failed to parse graph string %1%") %
                     CHECK_LOCATION().AsString()));
     }
 
     return CreateNetworkFromNetParameter(netParam, inputShapes, requestedOutputs);
 }

+
+
+

+

+ +

◆ CreateNetworkFromTextFile()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + + + + + +

INetworkPtr CreateNetworkFromTextFile	(	const char *	graphFile,
		const std::map< std::string, armnn::TensorShape > &	inputShapes,
		const std::vector< std::string > &	requestedOutputs
	)

+

+overridevirtual

+

+ +

Create the network from a protobuf text file on disk.

+ +

Implements ICaffeParser.

+ +

Definition at line 1733 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, and CaffeParserBase::CreateNetworkFromNetParameter().

+

 {
     FILE* fd = fopen(graphFile, "r");
 
     if (fd == nullptr)
     {
         throw FileNotFoundException(
             boost::str(
                 boost::format(
                     "Failed to open graph file: %1% %2%") %
                     graphFile %
                     CHECK_LOCATION().AsString()));
     }
 
     // Parses the file into a message.
     NetParameter netParam;
     auto         input   = new google::protobuf::io::FileInputStream(fileno(fd));
     bool         success = google::protobuf::TextFormat::Parse(input, &netParam);
     delete input;
     fclose(fd);
 
     if (!success)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Failed to parse graph file: %1% %2%") %
                     graphFile %
                     CHECK_LOCATION().AsString()));
     }
 
     return CreateNetworkFromNetParameter(netParam, inputShapes, requestedOutputs);
 }

+
+
+
+

+

+ +

◆ GetArmnnOutputSlotForCaffeTop()

+ +

+

+ + + + + +

+ + + + + + + + +

armnn::IOutputSlot & GetArmnnOutputSlotForCaffeTop ( const std::string & caffeTopName ) const

+

+protected

+

+ +

Retrieves the Armnn IOutputSlot representing the given Caffe top.

+

Throws if it cannot be found (e.g. not parsed yet).

+ +

Definition at line 1533 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, and CaffeParserBase::m_ArmnnOutputSlotForCaffeTop.

+ +

+

 {
     auto it = m_ArmnnOutputSlotForCaffeTop.find(caffeTopName);
     if (it != m_ArmnnOutputSlotForCaffeTop.end())
     {
         return *it->second;
     }
     else
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Could not find armnn output slot for Caffe top '%1%' %2%") %
                     caffeTopName %
                     CHECK_LOCATION().AsString()));
     }
 }

+
+
+

+

+ +

◆ GetBindingInfo()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + + + + + +

std::pair< armnn::LayerBindingId, armnn::TensorInfo > GetBindingInfo	(	const std::string &	layerName,
		const char *	bindingPointDesc,
		const std::unordered_map< std::string, BindingPointInfo > &	bindingInfos
	)

+

+staticprotected

+

+ +

Definition at line 296 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION.

+ +

Referenced by CaffeParserBase::GetNetworkInputBindingInfo(), and CaffeParserBase::GetNetworkOutputBindingInfo().

+

 {
     auto it = nameToBindingInfo.find(layerName);
     if (it == nameToBindingInfo.end())
     {
         throw InvalidArgumentException(
             boost::str(
                 boost::format(
                     "Unknown binding %1% for layer '%2%'. %3%") %
                     bindingPointDesc %
                     layerName %
                     CHECK_LOCATION().AsString()));
     }
     return it->second;
 }

+
+

+

+ +

◆ GetInputs()

+ +

+

+ + + + + +

+ + + + + + + + +

vector< const LayerParameter * > GetInputs ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Find the Caffe layers listed as inputs (bottoms) for a given layer.

+ +

Definition at line 339 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, and CaffeParserBase::m_CaffeLayersByTopName.

+ +

Referenced by CaffeParserBase::LoadNetParam().

+

 {
     std::vector<const caffe::LayerParameter*> ret;
     ret.reserve(boost::numeric_cast<size_t>(layerParam.bottom_size()));
     for (int j = 0; j < layerParam.bottom_size(); ++j)
     {
         std::string inputName = layerParam.bottom(j);
         auto inputIt = m_CaffeLayersByTopName.find(inputName);
         if (inputIt == m_CaffeLayersByTopName.end())
         {
             throw ParseException(
                 boost::str(
                     boost::format(
                         "Can't find Caffe layer with top called '%1%', "
                         "which is listed as an input of '%2%'. %3%") %
                         inputName %
                         layerParam.name() %
                         CHECK_LOCATION().AsString()));
         }
         ret.push_back(inputIt->second);
     }
 
     return ret;
 }

+
+
+

+

+ +

◆ GetNetworkInputBindingInfo()

+ +

+

+ + + + + +

+ + + + + + + + +

BindingPointInfo GetNetworkInputBindingInfo ( const std::string & name ) const

+

+overridevirtual

+

+ +

Retrieves binding info (layer id and tensor info) for the network input identified by the given layer name.

+ +

Implements ICaffeParser.

+ +

Definition at line 286 of file CaffeParser.cpp.

+ +

References CaffeParserBase::GetBindingInfo(), and CaffeParserBase::m_NetworkInputsBindingInfo.

+

 {
     return GetBindingInfo(name, "input", m_NetworkInputsBindingInfo);
 }

+
+

+

+ +

◆ GetNetworkOutputBindingInfo()

+ +

+

+ + + + + +

+ + + + + + + + +

BindingPointInfo GetNetworkOutputBindingInfo ( const std::string & name ) const

+

+overridevirtual

+

+ +

Retrieves binding info (layer id and tensor info) for the network output identified by the given layer name.

+ +

Implements ICaffeParser.

+ +

Definition at line 291 of file CaffeParser.cpp.

+ +

References CaffeParserBase::GetBindingInfo(), and CaffeParserBase::m_NetworkOutputsBindingInfo.

+

 {
     return GetBindingInfo(name, "output", m_NetworkOutputsBindingInfo);
 }

+
+

+

+ +

◆ LoadNetParam()

+ +

+

+ + + + + +

+ + + + + + + + +

void LoadNetParam ( caffe::NetParameter & netParameter )

+

+protected

+

+ +

does the actual conversion from caffe::NetParameter to armnn::INetwork

+ +

Definition at line 1632 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, IOutputSlot::Connect(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), CaffeParserBase::GetInputs(), CaffeParserBase::m_CaffeLayersByTopName, CaffeParserBase::m_Network, CaffeParserBase::m_NetworkOutputsBindingInfo, CaffeParserBase::m_RequestedOutputs, CaffeParserBase::ms_CaffeLayerNameToParsingFunctions, armnn::numeric_cast(), CaffeParserBase::ResolveInPlaceLayers(), and CaffeParserBase::TrackOutputBinding().

+ +

Referenced by CaffeParserBase::CreateNetworkFromNetParameter().

+

 {
     // Caffe models sometimes have an implicit input layer.
     // In that case, add an explicit one.
     if (netParameter.input_size() > 0)
     {
         LayerParameter* newLayer = netParameter.add_layer();
 
         newLayer->set_type("Input");
         newLayer->set_name(netParameter.input(0));
         newLayer->add_top(netParameter.input(0));
 
         InputParameter* inputParam = newLayer->mutable_input_param();
         BlobShape* shape = inputParam->add_shape();
 
         int dim_size = netParameter.input_dim_size();
         for (int i = 0; i < dim_size; ++i)
         {
             shape->add_dim(netParameter.input_dim(i));
         }
     }
 
     // Replaces in-place layers with regular ones to make the rest of the parsing easier.
     ResolveInPlaceLayers(netParameter);
 
     // Creates a lookup of Caffe layers by name.
     for (int i = 0; i < netParameter.layer_size(); ++i)
     {
         const caffe::LayerParameter& layer = netParameter.layer(i);
         for (int i = 0; i < layer.top_size(); ++i)
         {
             m_CaffeLayersByTopName[layer.top(i)] = &layer;
         }
     }
 
     // Finds the output layers the user requested.
     std::vector<const caffe::LayerParameter*> targetLayers;
     for (const std::string& requestedOutputName : m_RequestedOutputs)
     {
         auto nodeIt = m_CaffeLayersByTopName.find(requestedOutputName);
         if (nodeIt == m_CaffeLayersByTopName.end())
         {
             throw ParseException(
                 boost::str(
                     boost::format(
                         "Couldn't find requested output layer '%1%' in graph %2%") %
                         requestedOutputName %
                         CHECK_LOCATION().AsString()));
         }
         targetLayers.push_back(nodeIt->second);
     }
 
     // Sorts them into a linear ordering such that all inputs of a node are before the node itself.
     std::vector<const caffe::LayerParameter*> sortedNodes;
     if (!armnnUtils::GraphTopologicalSort<const caffe::LayerParameter*>(
         targetLayers,
         [this](const caffe::LayerParameter* node)
         {
             return GetInputs(*node);
         },
         sortedNodes))
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Cycle detected in graph. #nodes: %1% %2%") %
                     sortedNodes.size() %
                     CHECK_LOCATION().AsString()));
     }
 
     // Parses each node in order, knowing that all inputs of a node will be processed before the node itself.
     for (const caffe::LayerParameter* current : sortedNodes)
     {
         auto it = ms_CaffeLayerNameToParsingFunctions.find(current->type());
         if (it == ms_CaffeLayerNameToParsingFunctions.end())
         {
             throw ParseException(
                 boost::str(
                     boost::format("Unsupported layer type: '%1%' for layer %2% %3%") %
                     current->type() %
                     current->name() %
                     CHECK_LOCATION().AsString()));
         }
         auto func = it->second;
         (this->*func)(*current);
     }
 
     // Adds ArmNN output layers connected to each requested output.
     for (const std::string& requestedOutput : m_RequestedOutputs)
     {
         armnn::IOutputSlot& outputSlot = GetArmnnOutputSlotForCaffeTop(requestedOutput);
 
         const armnn::LayerBindingId outputId = boost::numeric_cast<armnn::LayerBindingId>(
             m_NetworkOutputsBindingInfo.size());
         armnn::IConnectableLayer* const outputLayer = m_Network->AddOutputLayer(outputId, requestedOutput.c_str());
         outputSlot.Connect(outputLayer->GetInputSlot(0));
 
         TrackOutputBinding(outputLayer, outputId, outputLayer->GetInputSlot(0).GetConnection()->GetTensorInfo());
     }
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseBatchNormLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseBatchNormLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1338 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, Connect(), armnn::Float32, CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), IConnectableLayer::GetInputSlot(), IConnectableLayer::GetOutputSlot(), TensorInfo::GetShape(), armnnUtils::GetTensorInfo(), BatchNormalizationDescriptor::m_Eps, CaffeParserBase::m_Network, CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), and IOutputSlot::SetTensorInfo().

+

 {
     ValidateNumInputsOutputs(layerParam, 1, 1);
 
     const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo();
 
     string name = layerParam.name();
 
     BatchNormParameter param = layerParam.batch_norm_param();
     // If use_global_stats is not explicitly set in the model, assume it to be true (its default value
     // when the network is in the testing phase).
     if (param.has_use_global_stats())
     {
         if (!param.use_global_stats())
         {
             throw ParseException(
                 boost::str(
                     boost::format(
                         "Error parsing Batch Norm layer '%1%': "
                         "Parameter 'use_global_stats' is set to false, which is "
                         "unsupported (value used for training). %2%") %
                         name %
                         CHECK_LOCATION().AsString()));
         }
     }
 
     BatchNormalizationDescriptor desc;
     desc.m_Eps = param.eps();
 
     unsigned int channels = inputInfo.GetShape()[1];
     unsigned int shape[]  = {channels};
 
     vector<float> meanData(channels);
     GetDataFromBlob(layerParam, meanData, 0);
 
     vector<float> varianceData(channels);
     GetDataFromBlob(layerParam, varianceData, 1);
 
     // Reads moving average factor and applies scaling (if required).
     const BlobProto& blob = layerParam.blobs(boost::numeric_cast<int>(2));
     const float movingAverageFactor = blob.data(boost::numeric_cast<int>(0));
     if(movingAverageFactor != 0.0f)
     {
         const float scaleFactor = 1.0f / movingAverageFactor;
         auto scaleFunction = [scaleFactor](float f) -> float { return f * scaleFactor; };
 
         std::transform(varianceData.begin(), varianceData.end(), varianceData.begin(), scaleFunction);
         std::transform(meanData.begin(), meanData.end(), meanData.begin(), scaleFunction);
     }
 
     // Identifies scale operation.
     vector<float> betaData(channels, 0.0f);
     vector<float> gammaData(channels, 1.0f);
 
     ConstTensor mean(TensorInfo(1, shape, armnn::DataType::Float32), meanData);
     ConstTensor variance(TensorInfo(1, shape, armnn::DataType::Float32), varianceData);
     ConstTensor beta(TensorInfo(1, shape, armnn::DataType::Float32), betaData);
     ConstTensor gamma(TensorInfo(1, shape, armnn::DataType::Float32), gammaData);
 
     armnn::IConnectableLayer* const batchNormLayer = m_Network->AddBatchNormalizationLayer(desc,
         mean, variance, beta, gamma, name.c_str());
     GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(batchNormLayer->GetInputSlot(0));
     batchNormLayer->GetOutputSlot(0).SetTensorInfo(inputInfo);
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), batchNormLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseConcatLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseConcatLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1277 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, IOutputSlot::Connect(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), IConnectableLayer::GetInputSlot(), TensorInfo::GetNumDimensions(), IConnectableLayer::GetOutputSlot(), TensorInfo::GetShape(), armnnUtils::GetTensorInfo(), CaffeParserBase::m_Network, CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), IOutputSlot::SetTensorInfo(), and OriginsDescriptor::SetViewOriginCoord().

+

 {
     unsigned int numInputs = static_cast<unsigned int>(layerParam.bottom_size());
     // We assume concat happens along the channel dimension, which is 1 in (0, 1, 2, 3).
     unsigned int concatDim = 1;
     unsigned int numOfDims = 4;
 
     // we only consider 4-D tensor here
     OriginsDescriptor concatDescriptor(static_cast<uint32_t>(numInputs), numOfDims);
     std::vector<unsigned int>mergeDimSizes(numOfDims, 0u);
 
     unsigned int mergeDim = 0;
     for (unsigned int viewIndex = 0; viewIndex < numInputs; ++viewIndex)
     {
         const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(
             layerParam.bottom(boost::numeric_cast<int>(viewIndex))).GetTensorInfo();
         // Checks whether the dimensions of the input tensors are actually 4.
         if (inputInfo.GetNumDimensions()!=4)
         {
             throw ParseException(
                 boost::str(
                     boost::format(
                         "The number of dimensions for input tensors of "
                         "the concatenation op should be 4. Inputs of %1% has "
                         "%2% dimensions. %3%") %
                         layerParam.name() %
                         inputInfo.GetNumDimensions() %
                         CHECK_LOCATION().AsString()));
         }
 
         mergeDimSizes[0] = inputInfo.GetShape()[0];
         mergeDimSizes[1] = inputInfo.GetShape()[1];
         mergeDimSizes[2] = inputInfo.GetShape()[2];
         mergeDimSizes[3] = inputInfo.GetShape()[3];
 
         for (unsigned int j = 0; j < concatDim; ++j)
         {
             concatDescriptor.SetViewOriginCoord(viewIndex, j, 0);
         }
 
         concatDescriptor.SetViewOriginCoord(viewIndex, concatDim, mergeDim);
         mergeDim += mergeDimSizes[concatDim];
 
         for (unsigned int j = concatDim+1; j < numOfDims; ++j)
         {
             concatDescriptor.SetViewOriginCoord(viewIndex, j, 0);
         }
     }
     mergeDimSizes[concatDim] = mergeDim;
 
     armnn::IConnectableLayer* concatlayer = m_Network->AddConcatLayer(concatDescriptor, layerParam.name().c_str());
     for (unsigned int i = 0; i < numInputs; ++i)
     {
         armnn::IOutputSlot& outputSlot = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(boost::numeric_cast<int>(i)));
         outputSlot.Connect(concatlayer->GetInputSlot(i));
     }
 
     concatlayer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo(numOfDims, mergeDimSizes.data(), DataType::Float32));
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), concatlayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseConvLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseConvLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 701 of file CaffeParser.cpp.

+ +

References CaffeParserBase::AddConvLayerWithDepthwiseConv(), CaffeParserBase::AddConvLayerWithSplits(), CaffeParserBase::BlobShapeToTensorInfo(), CHECK_LOCATION, IOutputSlot::Connect(), GET_OPTIONAL_WITH_VECTOR_FALLBACK, CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), armnnUtils::GetTensorInfo(), Convolution2dDescriptor::m_BiasEnabled, CaffeParserBase::m_Network, Convolution2dDescriptor::m_PadBottom, Convolution2dDescriptor::m_PadLeft, Convolution2dDescriptor::m_PadRight, Convolution2dDescriptor::m_PadTop, Convolution2dDescriptor::m_StrideX, Convolution2dDescriptor::m_StrideY, CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), and armnnCaffeParser::TensorDescToBlobShape().

+

 {
     // Ignored Caffe Parameters
     // * Dilation Size
     // * Weight Filler
     // * Bias Filler
     // * Engine
     // * Force nd_im2col
     // * Axis
 
     // Not Available ArmNN Interface Parameters
     // * Rounding policy;
 
     BOOST_ASSERT(layerParam.type() == "Convolution");
     ValidateNumInputsOutputs(layerParam, 1, 1);
 
     ConvolutionParameter convParam = layerParam.convolution_param();
     BlobShape inputShape = TensorDescToBlobShape(GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo());
     const unsigned int numGroups = convParam.has_group() ? convParam.group() : 1;
     unsigned int numFilters = convParam.num_output();
 
     const auto notFound = std::numeric_limits<unsigned int>::max();
 
     unsigned int kernelH = GET_OPTIONAL_WITH_VECTOR_FALLBACK(convParam, ConvolutionParameter,
                                                              kernel_h, kernel_size, unsigned int, notFound);
     unsigned int kernelW = GET_OPTIONAL_WITH_VECTOR_FALLBACK(convParam, ConvolutionParameter,
                                                              kernel_w, kernel_size, unsigned int, notFound);
 
     unsigned int strideH = GET_OPTIONAL_WITH_VECTOR_FALLBACK(convParam, ConvolutionParameter,
                                                              stride_h, stride, unsigned int, 1u);
     unsigned int strideW = GET_OPTIONAL_WITH_VECTOR_FALLBACK(convParam, ConvolutionParameter,
                                                              stride_w, stride, unsigned int, 1u);
 
     unsigned int padH = GET_OPTIONAL_WITH_VECTOR_FALLBACK(convParam, ConvolutionParameter,
                                                           pad_h, pad, unsigned int, 0u);
     unsigned int padW = GET_OPTIONAL_WITH_VECTOR_FALLBACK(convParam, ConvolutionParameter,
                                                           pad_w, pad, unsigned int, 0u);
 
     Convolution2dDescriptor convolution2dDescriptor;
     convolution2dDescriptor.m_PadLeft     = padW;
     convolution2dDescriptor.m_PadRight    = padW;
     convolution2dDescriptor.m_PadTop      = padH;
     convolution2dDescriptor.m_PadBottom   = padH;
     convolution2dDescriptor.m_StrideX     = strideW;
     convolution2dDescriptor.m_StrideY     = strideH;
     convolution2dDescriptor.m_BiasEnabled = convParam.has_bias_term() ? convParam.bias_term() : true;
 
     if (numGroups > numFilters)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Error parsing Convolution: %1%. "
                     "The 'group'=%2% parameter cannot be larger than the "
                     "number of filters supplied ='%3%'. %4%") %
                     layerParam.name() %
                     numGroups %
                     numFilters %
                     CHECK_LOCATION().AsString()));
     }
 
     if (inputShape.dim_size() != 4)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Convolution input shape is expected to have 4 dimensions. "
                     "%1%'s input has only %2%. %3%") %
                     layerParam.name() %
                     inputShape.dim_size() %
                     CHECK_LOCATION().AsString()));
     }
 
     if (numGroups > 1)
     {
         if (numGroups > inputShape.dim(1))
         {
             throw ParseException(
                 boost::str(
                     boost::format(
                         "Error parsing Convolution: %1%. "
                         "The 'group'=%2% parameter cannot be larger than the "
                         "channel of the input shape=%3% (in NCHW format). %4%") %
                         layerParam.name() %
                         numGroups %
                         inputShape.dim(1) %
                         CHECK_LOCATION().AsString()));
         }
         else if (numGroups == inputShape.dim(1))
         {
             // we use a depthwise convolution here, because the number of groups equals to the
             // input channels
             AddConvLayerWithDepthwiseConv(layerParam, convolution2dDescriptor, kernelW, kernelH);
             return;
         }
         else
         {
             // we split the input by channels into channels/groups separate convolutions
             // and concatenate the results afterwards
             AddConvLayerWithSplits(layerParam, convolution2dDescriptor, kernelW, kernelH);
             return;
         }
     }
 
     // NOTE: at this point we only need to handle #group=1 case, all other cases should be
     //       handled by the AddConvLayer* helpers
 
     // Populate convolution output tensor descriptor dimensions
     BlobShape outputShape;
     outputShape.add_dim(0);
     outputShape.set_dim(0, inputShape.dim(0));
     outputShape.add_dim(1);
     outputShape.set_dim(1, numFilters);
     outputShape.add_dim(2);
     outputShape.set_dim(
         2, (static_cast<int>(
                 static_cast<float>(inputShape.dim(2) + 2 * padH - kernelH) /
                 static_cast<float>(strideH)) + 1));
     outputShape.add_dim(3);
     outputShape.set_dim(
         3, (static_cast<int>(
                 static_cast<float>(inputShape.dim(3) + 2 * padW - kernelW) /
                 static_cast<float>(strideW)) + 1));
 
     // Load the weight data for ALL groups
     vector<float> weightData(boost::numeric_cast<size_t>(inputShape.dim(1) *
                                                          outputShape.dim(1) *
                                                          kernelH *
                                                          kernelW));
     GetDataFromBlob(layerParam, weightData, 0);
 
     const unsigned int weightDimSizes[4] = {
         static_cast<unsigned int>(outputShape.dim(1)), // output channels
         static_cast<unsigned int>(inputShape.dim(1)),  // input channels
         kernelH,
         kernelW};
 
     armnn::IConnectableLayer* returnLayer = nullptr;
 
     // Pull out the weights for this group from that loaded from the model file earlier
     ConstTensor weights(TensorInfo(4, weightDimSizes, DataType::Float32), weightData.data());
     Optional<ConstTensor> optionalBiases;
     vector<float> biasData;
     if (convolution2dDescriptor.m_BiasEnabled)
     {
         TensorInfo biasInfo;
 
         biasData.resize(boost::numeric_cast<size_t>(outputShape.dim(1)), 1.f);
         GetDataFromBlob(layerParam, biasData, 1);
 
         const unsigned int biasDimSizes[1] = {static_cast<unsigned int>(outputShape.dim(1))};
         biasInfo = TensorInfo(1, biasDimSizes, DataType::Float32);
 
         // Pull out the biases for this group from that loaded from the model file earlier
         ConstTensor biases(biasInfo, biasData.data());
         optionalBiases = Optional<ConstTensor>(biases);
     }
     returnLayer = m_Network->AddConvolution2dLayer(convolution2dDescriptor,
                                                    weights,
                                                    optionalBiases,
                                                    layerParam.name().c_str());
 
     armnn::IOutputSlot& inputConnection = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0));
     inputConnection.Connect(returnLayer->GetInputSlot(0));
     returnLayer->GetOutputSlot(0).SetTensorInfo(BlobShapeToTensorInfo(outputShape));
 
     if (!returnLayer)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Failed to create Convolution layer. "
                     "Layer=%1% #groups=%2% #filters=%3% %4%") %
                     layerParam.name() %
                     numGroups %
                     numFilters %
                     CHECK_LOCATION().AsString()));
     }
 
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), returnLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseDropoutLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseDropoutLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1477 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), and CaffeParserBase::SetArmnnOutputSlotForCaffeTop().

+

 {
     // Ignored for inference, so patch the single input to its single output.
     if (layerParam.bottom_size() != 1 || layerParam.top_size() != 1)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Dropout layer '%1%' should have exactly 1 bottom and 1 top. "
                     "#bottoms=%2% #tops=%3% %4%") %
                     layerParam.name() %
                     layerParam.bottom_size() %
                     layerParam.top_size() %
                     CHECK_LOCATION().AsString()));
     }
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)));
 }

+
+
+
+

+

+ +

◆ ParseEltwiseLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseEltwiseLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1230 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, Connect(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), IConnectableLayer::GetInputSlot(), IConnectableLayer::GetOutputSlot(), armnnUtils::GetTensorInfo(), CaffeParserBase::m_Network, CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), and IOutputSlot::SetTensorInfo().

+

 {
     ValidateNumInputsOutputs(layerParam, 2, 1);
 
     const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo();
 
     // Ignored Caffe Parameters:
     //      coeff
 
     EltwiseParameter_EltwiseOp operation = EltwiseParameter_EltwiseOp_SUM; // Defaults to sum as per caffe.
 
     if (layerParam.has_eltwise_param() && layerParam.eltwise_param().has_operation())
     {
         operation = layerParam.eltwise_param().operation();
     }
 
     armnn::IConnectableLayer* newLayer = nullptr;
     switch (operation)
     {
         case EltwiseParameter_EltwiseOp_SUM:
         {
             newLayer = m_Network->AddAdditionLayer(layerParam.name().c_str());
             break;
         }
         case EltwiseParameter_EltwiseOp_PROD:
         {
             newLayer = m_Network->AddMultiplicationLayer(layerParam.name().c_str());
             break;
         }
         default:
         {
             throw ParseException(
                 boost::str(
                     boost::format(
                         "Unsupported operation %1% in Eltwise layer %2% %3%") %
                         operation %
                         layerParam.name() %
                         CHECK_LOCATION().AsString()));
         }
     }
 
     GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(newLayer->GetInputSlot(0));
     GetArmnnOutputSlotForCaffeTop(layerParam.bottom(1)).Connect(newLayer->GetInputSlot(1));
     newLayer->GetOutputSlot(0).SetTensorInfo(inputInfo);
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), newLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseInnerProductLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseInnerProductLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1134 of file CaffeParser.cpp.

+ +

References Connect(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), IConnectableLayer::GetInputSlot(), TensorInfo::GetNumDimensions(), IConnectableLayer::GetOutputSlot(), TensorInfo::GetShape(), armnnUtils::GetTensorInfo(), FullyConnectedDescriptor::m_BiasEnabled, CaffeParserBase::m_Network, FullyConnectedDescriptor::m_TransposeWeightMatrix, CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), and IOutputSlot::SetTensorInfo().

+

 {
     InnerProductParameter param = layerParam.inner_product_param();
 
     ValidateNumInputsOutputs(layerParam, 1, 1);
 
     unsigned int outputSize = param.num_output();
 
     // Ignored Caffe Parameters:
     // Weight Filler
     // Bias Filler
     // Engine
     // Axis
 
     FullyConnectedDescriptor tensorFullyConnectedDescriptor;
 
     if (param.has_transpose())
     {
         // If true, assumes transposed weights.
         tensorFullyConnectedDescriptor.m_TransposeWeightMatrix = param.transpose();
     }
     else
     {
         // Caffe defaults to transposed.
         tensorFullyConnectedDescriptor.m_TransposeWeightMatrix = true;
     }
 
     const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo();
 
     TensorInfo weightInfo;
     TensorInfo biasInfo;
 
     // Allows implicit flattening of extra dimensions.
     unsigned int inputSize = inputInfo.GetShape()[1];
     for (unsigned int i = 2; i < inputInfo.GetNumDimensions(); ++i)
     {
         inputSize *= inputInfo.GetShape()[i];
     }
 
     const float* weightDataPtr = GetArrayPtrFromBlob(layerParam, 0);
     const unsigned int swTD[2] = { outputSize, inputSize };
     ConstTensor weights(TensorInfo(2, swTD, DataType::Float32), weightDataPtr);
 
     tensorFullyConnectedDescriptor.m_BiasEnabled = true;
     // Todo: check whether bias enabled.
     armnn::IConnectableLayer* fullyConnectedLayer = nullptr;
     if (tensorFullyConnectedDescriptor.m_BiasEnabled)
     {
         // BIAS VALUE
         const float* biasDataPtr = GetArrayPtrFromBlob(layerParam, 1);
 
         const unsigned int sbTD[1] = { outputSize };
 
         ConstTensor biases(TensorInfo(1, sbTD, DataType::Float32), biasDataPtr);
 
         fullyConnectedLayer = m_Network->AddFullyConnectedLayer(tensorFullyConnectedDescriptor,
                                                                 weights,
                                                                 Optional<ConstTensor>(biases),
                                                                 layerParam.name().c_str());
     }
     else
     {
         fullyConnectedLayer = m_Network->AddFullyConnectedLayer(tensorFullyConnectedDescriptor,
                                                                 weights,
                                                                 EmptyOptional(),
                                                                 layerParam.name().c_str());
     }
 
     TensorInfo outputInfo({ inputInfo.GetShape()[0], outputSize }, DataType::Float32);
     GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(fullyConnectedLayer->GetInputSlot(0));
     fullyConnectedLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), fullyConnectedLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseInputLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseInputLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Adds an armnn layer to m_Network given a Caffe LayerParameter of the correct type and is responsible for recording any newly created IOutputSlots using SetArmnnOutputSlotForCaffeTop().

+ +

Definition at line 364 of file CaffeParser.cpp.

+ +

References CaffeParserBase::BlobShapeToTensorInfo(), CHECK_LOCATION, CaffeParserBase::m_InputShapes, CaffeParserBase::m_Network, CaffeParserBase::m_NetworkInputsBindingInfo, armnn::numeric_cast(), CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), TensorInfo::SetShape(), and CaffeParserBase::TrackInputBinding().

+

 {
     BOOST_ASSERT(layerParam.type() == "Input");
     ValidateNumInputsOutputs(layerParam, 0, 1);
 
     const InputParameter& param = layerParam.input_param();
 
     const armnn::LayerBindingId inputId = boost::numeric_cast<armnn::LayerBindingId>(
         m_NetworkInputsBindingInfo.size());
     armnn::IConnectableLayer* const inputLayer = m_Network->AddInputLayer(inputId, layerParam.name().c_str());
 
     // Decides the tensor info for this input. This can be specified in the Caffe network but can also
     // be overriden by user input (m_inputShapes).
     armnn::TensorInfo inputTensorInfo;
 
     const BlobShape* originalShape = param.shape_size() > 0 && param.shape(0).dim_size() > 0 ?
         &param.shape(0) : nullptr;
     if (originalShape)
     {
         inputTensorInfo = BlobShapeToTensorInfo(*originalShape);
     }
 
     auto overrideIt = m_InputShapes.find(layerParam.name());
     if (overrideIt != m_InputShapes.end())
     {
         const TensorShape& overrideShape = overrideIt->second;
         if (originalShape &&
             (    originalShape->dim(1) != overrideShape[1]
               || originalShape->dim(2) != overrideShape[2]
               || originalShape->dim(3) != overrideShape[3]))
         {
             throw ParseException(
                 boost::str(
                     boost::format(
                         "Parsed input shape for '%1%' is incompatible with the override provided. %2%") %
                         layerParam.name() %
                         CHECK_LOCATION().AsString()));
         }
         inputTensorInfo.SetShape(overrideShape);
     }
     else if (!originalShape)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "No input descriptor given for '%1%' and no input shape found in caffe model. %2%") %
                     layerParam.name() %
                     CHECK_LOCATION().AsString()));
     }
 
     TrackInputBinding(inputLayer, inputId, inputTensorInfo);
     inputLayer->GetOutputSlot(0).SetTensorInfo(inputTensorInfo);
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), inputLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseLRNLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseLRNLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1027 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, Connect(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), IConnectableLayer::GetInputSlot(), IConnectableLayer::GetOutputSlot(), armnnUtils::GetTensorInfo(), NormalizationDescriptor::m_Alpha, NormalizationDescriptor::m_Beta, NormalizationDescriptor::m_K, CaffeParserBase::m_Network, NormalizationDescriptor::m_NormChannelType, NormalizationDescriptor::m_NormMethodType, NormalizationDescriptor::m_NormSize, armnn::numeric_cast(), CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), and IOutputSlot::SetTensorInfo().

+

 {
     ValidateNumInputsOutputs(layerParam, 1, 1);
 
     LRNParameter param = layerParam.lrn_param();
 
     const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo();
 
     // Ignored BATCH NORMALIZATION Caffe Parameters.
     // Ignored MVN Caffe Parameters.
     // Ignored LRN Caffe Parameters.
     //      Engine
 
     NormalizationDescriptor normalizationDescriptor;
     if (param.has_norm_region())
     {
         LRNParameter_NormRegion n = param.norm_region();
         switch (n)
         {
             case LRNParameter_NormRegion_ACROSS_CHANNELS:
             {
                 normalizationDescriptor.m_NormChannelType = NormalizationAlgorithmChannel::Across;
                 break;
             }
             case LRNParameter_NormRegion_WITHIN_CHANNEL:
             {
                 normalizationDescriptor.m_NormChannelType = NormalizationAlgorithmChannel::Within;
                 break;
             }
             default:
             {
                 throw ParseException(
                     boost::str(
                         boost::format(
                             "Unknown region %1% for LRN layer %2% %3%") %
                             n %
                             layerParam.name() %
                             CHECK_LOCATION().AsString()));
             }
         }
     }
     else
     {
         // Caffe defaults to normalization across channels.
         normalizationDescriptor.m_NormChannelType = NormalizationAlgorithmChannel::Across;
     }
 
     normalizationDescriptor.m_NormMethodType = NormalizationAlgorithmMethod::LocalBrightness;
     if (param.has_local_size())
     {
         normalizationDescriptor.m_NormSize = param.local_size();
     }
     else
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "local_size not defined for LRN layer %1% %2%") %
                     layerParam.name() %
                     CHECK_LOCATION().AsString()));
     }
 
     if (param.has_alpha())
     {
         normalizationDescriptor.m_Alpha = param.alpha();
         normalizationDescriptor.m_Alpha /= boost::numeric_cast<float>(param.local_size());
     }
     else
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Alpha not defined for LRN layer %1% %2%") %
                     layerParam.name() %
                     CHECK_LOCATION().AsString()));
     }
     if (param.has_beta())
     {
         normalizationDescriptor.m_Beta = param.beta();
     }
     else
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Beta not defined for LRN layer %1% %2%") %
                     layerParam.name() %
                     CHECK_LOCATION().AsString()));
     }
 
     if (param.has_k())
     {
         normalizationDescriptor.m_K = param.k();
     }
     else
     {
         normalizationDescriptor.m_K = 1;
     }
 
     IConnectableLayer* const normLayer = m_Network->AddNormalizationLayer(normalizationDescriptor,
         layerParam.name().c_str());
     GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(normLayer->GetInputSlot(0));
     normLayer->GetOutputSlot(0).SetTensorInfo(inputInfo);
 
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), normLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParsePoolingLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParsePoolingLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 883 of file CaffeParser.cpp.

+ +

+

 {
     // Ignored Caffe Parameters
     //      Stochastic Pooling
     //      Engine
 
     ValidateNumInputsOutputs(layerParam, 1, 1);
     PoolingParameter param = layerParam.pooling_param();
     const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo();
 
     const auto notFound = std::numeric_limits<unsigned int>::max();
 
     unsigned int kernel_h = GET_OPTIONAL_WITH_FALLBACK(param, PoolingParameter,
                                                        kernel_h, kernel_size, unsigned int, notFound);
     unsigned int kernel_w = GET_OPTIONAL_WITH_FALLBACK(param, PoolingParameter,
                                                        kernel_w, kernel_size, unsigned int, notFound);
 
     if ((kernel_h == notFound || kernel_w == notFound) && param.has_global_pooling())
     {
         kernel_h = inputInfo.GetShape()[2];
         kernel_w = inputInfo.GetShape()[3];
     }
 
     unsigned int stride_h = GET_OPTIONAL_WITH_FALLBACK(param, PoolingParameter,
                                                        stride_h, stride, unsigned int, notFound);
     unsigned int stride_w = GET_OPTIONAL_WITH_FALLBACK(param, PoolingParameter,
                                                        stride_h, stride, unsigned int, notFound);
 
     if ((stride_h == notFound || stride_w == notFound) && param.has_global_pooling())
     {
         stride_h = 1;
         stride_w = 1;
     }
 
     unsigned int pad_h = GET_OPTIONAL_WITH_FALLBACK(param, PoolingParameter,
                                                     pad_h, pad, unsigned int, 0u);
     unsigned int pad_w = GET_OPTIONAL_WITH_FALLBACK(param, PoolingParameter,
                                                     pad_w, pad, unsigned int, 0u);
 
     // Populate Weight and Bias Filter Descriptor
     Pooling2dDescriptor pooling2dDescriptor;
     if (param.has_pool())
     {
         PoolingParameter_PoolMethod p = param.pool();
         switch (p)
         {
             case PoolingParameter_PoolMethod_MAX:
             {
                 pooling2dDescriptor.m_PoolType = PoolingAlgorithm::Max;
                 break;
             }
             case PoolingParameter_PoolMethod_AVE:
             {
                 pooling2dDescriptor.m_PoolType = PoolingAlgorithm::Average;
                 break;
             }
             case PoolingParameter_PoolMethod_STOCHASTIC:
             {
                 throw ParseException(
                     boost::str(
                         boost::format(
                             "Pooling Layer: Stochastic Pooling Not Supported. Layer=%1% %2%") %
                             layerParam.name() %
                             CHECK_LOCATION().AsString()));
             }
             default:
             {
                 throw ParseException(
                     boost::str(
                         boost::format(
                             "Pooling Layer: unknown pooling method: %1% for layer: %2% %3%") %
                             p %
                             layerParam.name() %
                             CHECK_LOCATION().AsString()));
             }
         }
     }
     else
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "No Pooling Method Defined for %1% %2%") %
                     layerParam.name() %
                     CHECK_LOCATION().AsString()));
     }
 
     pooling2dDescriptor.m_PadLeft     = pad_w;
     pooling2dDescriptor.m_PadRight    = pad_w;
     pooling2dDescriptor.m_PadTop      = pad_h;
     pooling2dDescriptor.m_PadBottom   = pad_h;
     pooling2dDescriptor.m_StrideX     = stride_w;
     pooling2dDescriptor.m_StrideY     = stride_h;
     pooling2dDescriptor.m_PoolWidth   = kernel_w;
     pooling2dDescriptor.m_PoolHeight  = kernel_h;
 
     pooling2dDescriptor.m_OutputShapeRounding = OutputShapeRounding::Ceiling;
     pooling2dDescriptor.m_PaddingMethod  = PaddingMethod::IgnoreValue;
 
     armnn::IConnectableLayer* poolingLayer = m_Network->AddPooling2dLayer(pooling2dDescriptor,
         layerParam.name().c_str());
 
     TensorInfo outputInfo(
         { inputInfo.GetShape()[0],
           inputInfo.GetShape()[1],
           static_cast<unsigned int>(ceil(
               static_cast<float>(inputInfo.GetShape()[2] + 2 * pad_h - kernel_h) /
               boost::numeric_cast<float>(stride_h))) + 1,
           static_cast<unsigned int>(ceil(
               static_cast<float>(inputInfo.GetShape()[3] + 2 * pad_w - kernel_w) /
               boost::numeric_cast<float>(stride_w))) + 1 },
         DataType::Float32);
 
     GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(poolingLayer->GetInputSlot(0));
     poolingLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), poolingLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseReluLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseReluLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1001 of file CaffeParser.cpp.

+ +

References Connect(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), IConnectableLayer::GetInputSlot(), IConnectableLayer::GetOutputSlot(), armnnUtils::GetTensorInfo(), ActivationDescriptor::m_A, ActivationDescriptor::m_Function, CaffeParserBase::m_Network, CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), and IOutputSlot::SetTensorInfo().

+

 {
     ValidateNumInputsOutputs(layerParam, 1, 1);
 
     const string& name = layerParam.name();
     const ReLUParameter& param = layerParam.relu_param();
 
     ActivationDescriptor activationDescriptor;
     const float negativeSlope = param.negative_slope();
     if (negativeSlope == 0.0f)
     {
         activationDescriptor.m_Function = ActivationFunction::ReLu;
     }
     else
     {
         activationDescriptor.m_Function = ActivationFunction::LeakyReLu;
         activationDescriptor.m_A = negativeSlope;
     }
 
     const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo();
     IConnectableLayer* const activationLayer = m_Network->AddActivationLayer(activationDescriptor, name.c_str());
     GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(activationLayer->GetInputSlot(0));
     activationLayer->GetOutputSlot(0).SetTensorInfo(inputInfo);
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), activationLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseScaleLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseScaleLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1404 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, Connect(), armnn::Float32, CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), IConnectableLayer::GetInputSlot(), IConnectableLayer::GetOutputSlot(), TensorInfo::GetShape(), armnnUtils::GetTensorInfo(), BatchNormalizationDescriptor::m_Eps, CaffeParserBase::m_Network, CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), and IOutputSlot::SetTensorInfo().

+

 {
     // Current unoptimal solution: add a batchnormalization layer with 0 mean and 1 variance.
     ValidateNumInputsOutputs(layerParam, 1, 1);
 
     const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo();
 
     string name = layerParam.name();
 
     ScaleParameter param = layerParam.scale_param();
     if (param.axis() != 1)
     {
         // Would have to use something other than BatchNormalizationLayer in this case
         throw ParseException(
             boost::str(
                 boost::format(
                     "Loading Scale Layer: Only axis 1 is supported currently. "
                     "Layer=%1% Axis=%2% %3%") %
                     layerParam.name() %
                     param.axis() %
                     CHECK_LOCATION().AsString()));
     }
 
     unsigned int     channels = inputInfo.GetShape()[1];
     unsigned int     shape[]  = {channels};
 
     BatchNormalizationDescriptor desc;
     desc.m_Eps = 0.0f; // Don't need epsilon if variance is 1.
     vector<float> meanData(channels, 0.0f);
     vector<float> varianceData(channels, 1.0f);
     vector<float> betaData(channels, 0.0f);
     vector<float> gammaData(channels);
 
     GetDataFromBlob(layerParam, gammaData, 0);
 
     if(param.has_bias_term())
     {
         GetDataFromBlob(layerParam, betaData, 1);
     }
 
     ConstTensor mean(TensorInfo(1, shape, armnn::DataType::Float32), meanData);
     ConstTensor variance(TensorInfo(1, shape, armnn::DataType::Float32), varianceData);
     ConstTensor beta(TensorInfo(1, shape, armnn::DataType::Float32), betaData);
     ConstTensor gamma(TensorInfo(1, shape, armnn::DataType::Float32), gammaData);
 
     armnn::IConnectableLayer* const batchNormLayer = m_Network->AddBatchNormalizationLayer(desc,
         mean, variance, beta, gamma, name.c_str());
     GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(batchNormLayer->GetInputSlot(0));
     batchNormLayer->GetOutputSlot(0).SetTensorInfo(inputInfo);
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), batchNormLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseSoftmaxLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseSoftmaxLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1208 of file CaffeParser.cpp.

+ +

References Connect(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), IConnectableLayer::GetInputSlot(), IConnectableLayer::GetOutputSlot(), armnnUtils::GetTensorInfo(), SoftmaxDescriptor::m_Axis, CaffeParserBase::m_Network, CaffeParserBase::SetArmnnOutputSlotForCaffeTop(), and IOutputSlot::SetTensorInfo().

+

 {
     ValidateNumInputsOutputs(layerParam, 1, 1);
 
     SoftmaxParameter param = layerParam.softmax_param();
 
     const TensorInfo& inputInfo = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).GetTensorInfo();
 
     // Ignored Caffe Parameters:
     //      axis
     //      Engine
 
     armnn::SoftmaxDescriptor softmaxDescriptor;
     softmaxDescriptor.m_Axis = 1;
     armnn::IConnectableLayer* const softmaxLayer = m_Network->AddSoftmaxLayer(
         softmaxDescriptor,
         layerParam.name().c_str());
     GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0)).Connect(softmaxLayer->GetInputSlot(0));
     softmaxLayer->GetOutputSlot(0).SetTensorInfo(inputInfo);
     SetArmnnOutputSlotForCaffeTop(layerParam.top(0), softmaxLayer->GetOutputSlot(0));
 }

+
+
+
+
+
+
+
+
+
+
+
+

+

+ +

◆ ParseSplitLayer()

+ +

+

+ + + + + +

+ + + + + + + + +

void ParseSplitLayer ( const caffe::LayerParameter & layerParam )

+

+protected

+

+ +

Definition at line 1456 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), and CaffeParserBase::SetArmnnOutputSlotForCaffeTop().

+

 {
     // Used in caffe to duplicate memory - not necessary in armnn.
     if (layerParam.bottom_size() != 1)
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Split layer '%1%' should have exactly 1 bottom. "
                     "#bottoms=%2% %3%") %
                     layerParam.name() %
                     layerParam.bottom_size() %
                     CHECK_LOCATION().AsString()));
     }
     armnn::IOutputSlot& outputSlot = GetArmnnOutputSlotForCaffeTop(layerParam.bottom(0));
     for (int i = 0; i < layerParam.top_size(); i++)
     {
         SetArmnnOutputSlotForCaffeTop(layerParam.top(i), outputSlot);
     }
 }

+
+
+
+
+

+

+ +

◆ ResolveInPlaceLayers()

+ +

+

+ + + + + +

+ + + + + + + + +

void ResolveInPlaceLayers ( caffe::NetParameter & netParameter )

+

+protected

+

+ +

Modifies the Caffe network to replace "in-place" layers (whose top() and bottom() are both the same) with regular layers.

+

This simplifies further parsing.

+ +

Definition at line 1572 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION.

+ +

Referenced by RecordByRecordCaffeParser::CreateNetworkFromBinaryFile(), and CaffeParserBase::LoadNetParam().

+

 {
     // Finds layers with the same top.
     std::map<std::string, std::vector<caffe::LayerParameter*>> layersByTop;
     for (int layerIdx = 0; layerIdx < netParameter.layer_size(); ++layerIdx)
     {
         caffe::LayerParameter& layer = *netParameter.mutable_layer(layerIdx);
         std::string name = layer.name();
         for (int i = 0; i < layer.top_size(); ++i)
         {
             layersByTop[layer.top(i)].push_back(&layer);
         }
     }
 
     // For each set of layers with the same top, resolves them to a linear chain rather than in-place layers.
     // Note that for 'regular' layers, there will be a single layer in each group and so this will be a no-op.
     for (auto layersWithSameTopIt : layersByTop)
     {
         const std::string& top = layersWithSameTopIt.first;
         const std::vector<caffe::LayerParameter*>& layersWithSameTop = layersWithSameTopIt.second;
 
         // Chains the layers together in the order that they are listed in the prototxt (hopefully this is correct).
         // Note that the last layer will not have its top modified so that other layers will continue to reference it.
         for (unsigned int layerIdx = 0; layerIdx < layersWithSameTop.size() - 1; ++layerIdx)
         {
             caffe::LayerParameter& layer1 = *layersWithSameTop[layerIdx];
             caffe::LayerParameter& layer2 = *layersWithSameTop[layerIdx+1];
             if (layer1.top_size() != 1)
             {
                 throw ParseException(
                     boost::str(
                         boost::format(
                             "Node '%1%' is an in-place layer but doesn't have exactly one "
                             "top. It has %2% instead. %3%") %
                             layer1.name() %
                             layer1.top_size() %
                             CHECK_LOCATION().AsString()));
             }
             std::string newTop = layer1.name() + "_top";
             layer1.set_top(0, newTop);
             if (layer2.bottom_size() != 1 || layer2.bottom(0) != top)
             {
                 throw ParseException(
                     boost::str(
                         boost::format(
                             "Node '%1%' is an in-place layer but "
                             "doesn't have exactly one bottom, or it doesn't match its top. "
                             "#bottoms=%2%, first bottom is %3%, top is %4% %5%") %
                             layer2.name() %
                             layer2.bottom(0) %
                             top %
                             CHECK_LOCATION().AsString()));
             }
             layer2.set_bottom(0, newTop);
         }
     }
 }

+
+

+

+ +

◆ SetArmnnOutputSlotForCaffeTop()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + +

void SetArmnnOutputSlotForCaffeTop	(	const std::string &	caffeTopName,
		armnn::IOutputSlot &	armnnOutputSlot
	)

+

+protected

+

+ +

Definition at line 1551 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, and CaffeParserBase::m_ArmnnOutputSlotForCaffeTop.

+ +

+

 {
     auto it = m_ArmnnOutputSlotForCaffeTop.find(caffeTopName);
     if (it == m_ArmnnOutputSlotForCaffeTop.end())
     {
         m_ArmnnOutputSlotForCaffeTop[caffeTopName] = &armnnOutputSlot;
     }
     else
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Attempting to add duplicate entry for Caffe top '%1%' %2%") %
                     caffeTopName %
                     CHECK_LOCATION().AsString()));
     }
 }

+
+
+

+

+ +

◆ TrackBindingPoint()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +

void TrackBindingPoint	(	armnn::IConnectableLayer *	layer,
		armnn::LayerBindingId	id,
		const armnn::TensorInfo &	tensorInfo,
		const char *	bindingPointDesc,
		std::unordered_map< std::string, BindingPointInfo > &	nameToBindingInfo
	)

+

+staticprotected

+

+ +

Definition at line 1509 of file CaffeParser.cpp.

+ +

References CHECK_LOCATION, and IConnectableLayer::GetName().

+ +

Referenced by CaffeParserBase::TrackInputBinding(), and CaffeParserBase::TrackOutputBinding().

+

 {
     const std::string layerName = layer->GetName();
     auto it = nameToBindingInfo.find(layerName);
     if (it == nameToBindingInfo.end())
     {
         nameToBindingInfo[layerName] = std::make_pair(id, tensorInfo);
     }
     else
     {
         throw ParseException(
             boost::str(
                 boost::format(
                     "Id %1% used by more than one %2% layer %3%") %
                     id %
                     bindingPointDesc %
                     CHECK_LOCATION().AsString()));
     }
 }

+
+
+

+

+ +

◆ TrackInputBinding()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + + + + + +

void TrackInputBinding	(	armnn::IConnectableLayer *	layer,
		armnn::LayerBindingId	id,
		const armnn::TensorInfo &	tensorInfo
	)

+

+protected

+

+ +

Definition at line 1495 of file CaffeParser.cpp.

+ +

References IConnectableLayer::GetName(), CaffeParserBase::m_NetworkInputsBindingInfo, and CaffeParserBase::TrackBindingPoint().

+ +

Referenced by CaffeParserBase::ParseInputLayer().

+

 {
     return TrackBindingPoint(layer, id, tensorInfo, layer->GetName(), m_NetworkInputsBindingInfo);
 }

+
+
+

+

+ +

◆ TrackOutputBinding()

+ +

+

+ + + + + +

+ + + + + + + + + + + + + + + + + + + + + + + + +

void TrackOutputBinding	(	armnn::IConnectableLayer *	layer,
		armnn::LayerBindingId	id,
		const armnn::TensorInfo &	tensorInfo
	)

+

+protected

+

+ +

Definition at line 1502 of file CaffeParser.cpp.

+ +

References IConnectableLayer::GetName(), CaffeParserBase::m_NetworkOutputsBindingInfo, and CaffeParserBase::TrackBindingPoint().

+ +

Referenced by RecordByRecordCaffeParser::CreateNetworkFromBinaryFile(), and CaffeParserBase::LoadNetParam().

+

 {
     return TrackBindingPoint(layer, id, tensorInfo, layer->GetName(), m_NetworkOutputsBindingInfo);
 }

+
+
+

+

Member Data Documentation

+ +

◆ m_ArmnnOutputSlotForCaffeTop

+ +

+

+ + + + + +

+ + + + +

std::unordered_map<std::string, armnn::IOutputSlot*> m_ArmnnOutputSlotForCaffeTop

+

+protected

+

+ +

As we add armnn layers we store the armnn IOutputSlot which corresponds to the Caffe tops.

+ +

Definition at line 131 of file CaffeParser.hpp.

+ +

Referenced by CaffeParserBase::Cleanup(), CaffeParserBase::GetArmnnOutputSlotForCaffeTop(), and CaffeParserBase::SetArmnnOutputSlotForCaffeTop().

+ +

+

+ +

◆ m_CaffeLayersByTopName

+ +

+

+ + + + + +

+ + + + +

std::map<std::string, const caffe::LayerParameter*> m_CaffeLayersByTopName

+

+protected

+

+ +

Definition at line 139 of file CaffeParser.hpp.

+ +

Referenced by CaffeParserBase::Cleanup(), CaffeParserBase::GetInputs(), and CaffeParserBase::LoadNetParam().

+ +

+

+ +

◆ m_InputShapes

+ +

+

+ + + + + +

+ + + + +

std::map<std::string, armnn::TensorShape> m_InputShapes

+

+protected

+

+ +

Definition at line 128 of file CaffeParser.hpp.

+ +

Referenced by CaffeParserBase::Cleanup(), RecordByRecordCaffeParser::CreateNetworkFromBinaryFile(), CaffeParserBase::CreateNetworkFromNetParameter(), and CaffeParserBase::ParseInputLayer().

+ +

+

+ +

◆ m_Network

+ +

+

+ + + + + +

+ + + + +

armnn::INetworkPtr m_Network

+

+protected

+

+ +

Definition at line 126 of file CaffeParser.hpp.

+ +

+ +

+

+ +

◆ m_NetworkInputsBindingInfo

+ +

+

+ + + + + +

+ + + + +

std::unordered_map<std::string, BindingPointInfo> m_NetworkInputsBindingInfo

+

+protected

+

+ +

maps input layer names to their corresponding ids and tensor infos

+ +

Definition at line 121 of file CaffeParser.hpp.

+ +

Referenced by RecordByRecordCaffeParser::CreateNetworkFromBinaryFile(), CaffeParserBase::CreateNetworkFromNetParameter(), CaffeParserBase::GetNetworkInputBindingInfo(), CaffeParserBase::ParseInputLayer(), and CaffeParserBase::TrackInputBinding().

+ +

+

+ +

◆ m_NetworkOutputsBindingInfo

+ +

+

+ + + + + +

+ + + + +

std::unordered_map<std::string, BindingPointInfo> m_NetworkOutputsBindingInfo

+

+protected

+

+ +

maps output layer names to their corresponding ids and tensor infos

+ +

Definition at line 124 of file CaffeParser.hpp.

+ +

Referenced by RecordByRecordCaffeParser::CreateNetworkFromBinaryFile(), CaffeParserBase::CreateNetworkFromNetParameter(), CaffeParserBase::GetNetworkOutputBindingInfo(), CaffeParserBase::LoadNetParam(), and CaffeParserBase::TrackOutputBinding().

+ +

+

+ +

◆ m_RequestedOutputs

+ +

+

+ + + + + +

+ + + + +

std::vector<std::string> m_RequestedOutputs

+

+protected

+

+ +

Definition at line 133 of file CaffeParser.hpp.

+ +

Referenced by CaffeParserBase::Cleanup(), CaffeParserBase::CreateNetworkFromNetParameter(), and CaffeParserBase::LoadNetParam().

+ +

+

+ +

◆ ms_CaffeLayerNameToParsingFunctions

+ +

+

+ + + + + +

+ + + + +

const std::map< std::string, CaffeParserBase::OperationParsingFunction > ms_CaffeLayerNameToParsingFunctions

+

+staticprotected

+

+Initial value:

= {
    { "Input",        &CaffeParserBase::ParseInputLayer },
    { "Convolution",  &CaffeParserBase::ParseConvLayer },
    { "Pooling",      &CaffeParserBase::ParsePoolingLayer },
    { "ReLU",         &CaffeParserBase::ParseReluLayer },
    { "LRN",          &CaffeParserBase::ParseLRNLayer },
    { "InnerProduct", &CaffeParserBase::ParseInnerProductLayer },
    { "Softmax",      &CaffeParserBase::ParseSoftmaxLayer },
    { "Eltwise",      &CaffeParserBase::ParseEltwiseLayer },
    { "Concat",       &CaffeParserBase::ParseConcatLayer },
    { "BatchNorm",    &CaffeParserBase::ParseBatchNormLayer },
    { "Scale",        &CaffeParserBase::ParseScaleLayer },
    { "Split",        &CaffeParserBase::ParseSplitLayer },
    { "Dropout",      &CaffeParserBase::ParseDropoutLayer},
}

+

Maps Caffe layer names to parsing member functions.

+ +

Definition at line 118 of file CaffeParser.hpp.

+ +

Referenced by RecordByRecordCaffeParser::CreateNetworkFromBinaryFile(), and CaffeParserBase::LoadNetParam().

+ +

+

The documentation for this class was generated from the following files:

src/armnnCaffeParser/CaffeParser.hpp
src/armnnCaffeParser/CaffeParser.cpp

+

+Public Member Functions

+Protected Types

+Protected Member Functions

+Static Protected Member Functions

+Protected Attributes

+Static Protected Attributes

+Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ OperationParsingFunction

Constructor & Destructor Documentation

◆ CaffeParserBase()

Member Function Documentation

◆ AddConvLayerWithDepthwiseConv()

◆ AddConvLayerWithSplits()

◆ BlobShapeToTensorInfo()

◆ Cleanup()

◆ CreateNetworkFromNetParameter()

◆ CreateNetworkFromString()

◆ CreateNetworkFromTextFile()

◆ GetArmnnOutputSlotForCaffeTop()

◆ GetBindingInfo()

◆ GetInputs()

◆ GetNetworkInputBindingInfo()

◆ GetNetworkOutputBindingInfo()

◆ LoadNetParam()

◆ ParseBatchNormLayer()

◆ ParseConcatLayer()

◆ ParseConvLayer()

◆ ParseDropoutLayer()

◆ ParseEltwiseLayer()

◆ ParseInnerProductLayer()

◆ ParseInputLayer()

◆ ParseLRNLayer()

◆ ParsePoolingLayer()

◆ ParseReluLayer()

◆ ParseScaleLayer()

◆ ParseSoftmaxLayer()

◆ ParseSplitLayer()

◆ ResolveInPlaceLayers()

◆ SetArmnnOutputSlotForCaffeTop()

◆ TrackBindingPoint()

◆ TrackInputBinding()

◆ TrackOutputBinding()

Member Data Documentation

◆ m_ArmnnOutputSlotForCaffeTop

◆ m_CaffeLayersByTopName

◆ m_InputShapes

◆ m_Network

◆ m_NetworkInputsBindingInfo

◆ m_NetworkOutputsBindingInfo

◆ m_RequestedOutputs

◆ ms_CaffeLayerNameToParsingFunctions