Gitiles
Code Review Sign In
review.mlplatform.org / ml / armnn / 3537c2ca7ebf31c1673b9ec2bb0c17b0406bbae0 / . / src / armnnTfParser / TfParser.cpp
blob: 834c0dd41b8e0057db3ac47617d2dfdd2d68871e [file] [log] [blame]
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// See LICENSE file in the project root for full license information.
//
#include "TfParser.hpp"
#include <armnn/INetwork.hpp>
#include <armnn/Utils.hpp>
#include <armnn/TypesUtils.hpp>
#include <armnn/Exceptions.hpp>
#include <armnn/Descriptors.hpp>
#include <GraphTopologicalSort.hpp>
#include <Permute.hpp>
#include <google/protobuf/io/zero_copy_stream_impl.h>
#include <google/protobuf/text_format.h>
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/framework/node_def.pb.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/framework/tensor.pb.h"
#include "tensorflow/core/framework/tensor_shape.pb.h"
#include <boost/assert.hpp>
#include <boost/format.hpp>
#include <boost/core/ignore_unused.hpp>
#include <boost/log/trivial.hpp>
#include <boost/numeric/conversion/cast.hpp>
#include <boost/polymorphic_cast.hpp>
#include <memory>
#include <sstream>
#include <numeric>
#include <functional>
using namespace armnn;
namespace armnnTfParser
{
namespace
{
const PermutationVector NHWCToArmNN = { 0, 2, 3, 1 };
const PermutationVector ArmNNToNHWC = { 0, 3, 1, 2 };
IConnectableLayer* AddSwizzleLayer(INetwork& network, IOutputSlot& input, const PermutationVector& mapping,
const std::string& name)
{
// Add swizzle layer
IConnectableLayer* const layer = network.AddPermuteLayer(mapping, name.c_str());
// Connect intput to swizzle layer
input.Connect(layer->GetInputSlot(0));
// Setup swizzled output
const TensorInfo outInfo = armnnUtils::Permuted(input.GetTensorInfo(), mapping);
layer->GetOutputSlot(0).SetTensorInfo(outInfo);
return layer;
}
IConnectableLayer* SwizzleInDeswizzleOut(INetwork& network, IOutputSlot& input, IConnectableLayer& layer,
const std::string& name)
{
// Add swizzle layer
IConnectableLayer* const swizzleLayer = AddSwizzleLayer(network, input, NHWCToArmNN, "swizzle_for-" + name);
// Connect swizzledInput to layer
swizzleLayer->GetOutputSlot(0).Connect(layer.GetInputSlot(0));
// Add deswizzle layer
IConnectableLayer* const deswizzleLayer = AddSwizzleLayer(network, layer.GetOutputSlot(0), ArmNNToNHWC,
"deswizzle_for-" + name);
return deswizzleLayer;
}
template <typename Callable>
void ReadMandatoryNodeAttributeImpl(const tensorflow::NodeDef& nodeDef,
const std::string& attribName,
tensorflow::AttrValue::ValueCase expectedValueCase,
Callable callable)
{
auto iter = nodeDef.attr().find(attribName);
if (iter != nodeDef.attr().end())
{
const auto& attrValue = iter->second;
if (attrValue.value_case() == expectedValueCase)
{
callable(attrValue);
}
else
{
throw ParseException(boost::str(boost::format(
"Attribute %1% of node %2% expected to have %3% as tensorflow::AttrValue::ValueCase, "
"but found %4% instead")
% attribName
% nodeDef.name()
% static_cast<int>(expectedValueCase)
% static_cast<int>(attrValue.value_case())));
}
}
else
{
throw ParseException(boost::str(boost::format("Could not find required attribute %1% in node %2%")
% attribName % nodeDef.name()));
}
}
template <typename Callable>
void ReadOptionalNodeAttributeImpl(const tensorflow::NodeDef& nodeDef,
const std::string& attribName,
tensorflow::AttrValue::ValueCase expectedValueCase,
Callable callable)
{
auto iter = nodeDef.attr().find(attribName);
if (iter != nodeDef.attr().end())
{
const auto& attrValue = iter->second;
if (attrValue.value_case() == expectedValueCase)
{
callable(attrValue);
}
else
{
throw ParseException(boost::str(boost::format(
"Attribute %1% of node %2% expected to have %3% as tensorflow::AttrValue::ValueCase, "
"but found %4% instead")
% attribName
% nodeDef.name()
% static_cast<int>(expectedValueCase)
% static_cast<int>(attrValue.value_case())));
}
}
}
float ReadMandatoryNodeFloatAttribute(const tensorflow::NodeDef& nodeDef, const std::string& name)
{
float attribValue = 0.0f;
ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kF,
[&attribValue](const tensorflow::AttrValue& attrValue)
{
attribValue = attrValue.f();
});
return attribValue;
}
uint32_t ReadMandatoryNodeUint32Attribute(const tensorflow::NodeDef& nodeDef, const std::string& name)
{
uint32_t attribValue = 0u;
ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kI,
[&attribValue](const tensorflow::AttrValue& attrValue)
{
attribValue = static_cast<uint32_t>(attrValue.i());
});
return attribValue;
}
std::string ReadMandatoryNodeStringAttribute(const tensorflow::NodeDef& nodeDef, const std::string& name)
{
std::string attribValue = "";
ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kS,
[&attribValue](const tensorflow::AttrValue& attrValue)
{
attribValue = attrValue.s();
});
return attribValue;
}
std::vector<uint32_t> ReadMandatoryNodeUint32ListAttribute(const tensorflow::NodeDef& nodeDef,
const std::string& name)
{
std::vector<uint32_t> attriList;
ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kList,
[&attriList](const tensorflow::AttrValue& attrValue)
{
for (int attriNum = 0; attriNum < attrValue.list().i_size(); ++attriNum)
{
attriList.push_back(static_cast<uint32_t>(attrValue.list().i(attriNum)));
}
});
return attriList;
}
std::vector<uint32_t> ReadOptionalNodeUint32ListAttribute(const tensorflow::NodeDef& nodeDef,
const std::string& name)
{
std::vector<uint32_t> attriList;
ReadOptionalNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kList,
[&attriList](const tensorflow::AttrValue& attrValue)
{
for (int attriNum = 0; attriNum < attrValue.list().i_size(); ++attriNum)
{
attriList.push_back(static_cast<uint32_t>(attrValue.list().i(attriNum)));
}
});
return attriList;
}
bool ReadOptionalNodeBoolAttribute(const tensorflow::NodeDef& nodeDef,
const std::string& name,
bool defaultValue = false)
{
bool attribValue = defaultValue;
ReadOptionalNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kB,
[&attribValue](const tensorflow::AttrValue& attrValue)
{
attribValue = attrValue.b();
});
return attribValue;
}
tensorflow::DataType ReadMandatoryNodeTypeAttribute(const tensorflow::NodeDef& nodeDef, const std::string& name)
{
tensorflow::DataType attribValue = tensorflow::DT_INVALID;
ReadMandatoryNodeAttributeImpl(nodeDef, name, tensorflow::AttrValue::kType,
[&attribValue](const tensorflow::AttrValue& attrValue)
{
attribValue = attrValue.type();
});
return attribValue;
}
TensorInfo PrepareReshape(const TensorInfo& input, const std::vector<int32_t>& targetDims)
{
std::vector<unsigned int> outDims(targetDims.begin(), targetDims.end());
const auto stretchDim = std::find(targetDims.begin(), targetDims.end(), -1);
if (stretchDim != targetDims.end())
{
if (std::find(std::next(stretchDim), targetDims.end(), -1) != targetDims.end())
{
throw ParseException("At most one component of shape can be -1");
}
auto targetNumElements = boost::numeric_cast<unsigned int>(std::accumulate(targetDims.begin(), targetDims.end(),
-1, std::multiplies<int32_t>()));
auto stretchIndex = static_cast<size_t>(std::distance(targetDims.begin(), stretchDim));
outDims[stretchIndex] = input.GetNumElements() / targetNumElements;
}
TensorInfo reshapeInfo = input;
reshapeInfo.SetShape(TensorShape{ static_cast<unsigned int>(outDims.size()), outDims.data() });
return reshapeInfo;
}
// We need the input0Slot to guide the reshape for input1Slot
IOutputSlot* BroadcastForAddandMul(IOutputSlot* input0Slot, IOutputSlot* input1Slot, bool isNHWC, INetwork& m_Network,
const tensorflow::NodeDef& nodeDef)
{
const TensorInfo& input1Info = input1Slot->GetTensorInfo();
const TensorInfo inputTensorInfo = input0Slot->GetTensorInfo();
const unsigned int matchDim = inputTensorInfo.GetNumDimensions() - (isNHWC ? 1 : 3);
std::array<unsigned int, MaxNumOfTensorDimensions> reshapedDimensions;
std::fill_n(reshapedDimensions.begin(), inputTensorInfo.GetNumDimensions(), 1);
reshapedDimensions[matchDim] = input1Info.GetShape()[0];
armnn::TensorInfo reshapedInfo = input1Info;
reshapedInfo.SetShape(TensorShape{ inputTensorInfo.GetNumDimensions(), reshapedDimensions.data() });
const std::string reshapeLayerName = "reshape_for-" + nodeDef.name();
ReshapeDescriptor reshapeDesc;
reshapeDesc.m_TargetShape = reshapedInfo.GetShape();
IConnectableLayer* const reshapeLayer = m_Network.AddReshapeLayer(reshapeDesc, reshapeLayerName.c_str());
input1Slot->Connect(reshapeLayer->GetInputSlot(0));
reshapeLayer->GetOutputSlot(0).SetTensorInfo(reshapedInfo);
input1Slot = &reshapeLayer->GetOutputSlot(0);
return input1Slot;
}
OutputId ParseOutputId(const std::string & name)
{
unsigned int outputNum = 0;
size_t colonPos = name.find_last_of(":");
if (colonPos != std::string::npos)
{
int n = std::stoi(name.substr(colonPos+1));
if (n<0 || n>100)
{
throw ParseException("Output tensor id is out of range for "+name);
}
outputNum = static_cast<unsigned int>(n);
}
return OutputId(name.substr(0,colonPos),outputNum);
}
} // namespace
const std::map<std::string, TfParser::OperationParsingFunction> TfParser::ms_OperationNameToParsingFunctions = {
{ "Const", &TfParser::ParseConst },
{ "Add", &TfParser::ParseAdd },
{ "BiasAdd", &TfParser::ParseBiasAdd },
{ "Identity", &TfParser::ParseIdentity },
{ "Conv2D", &TfParser::ParseConv2D },
{ "DepthwiseConv2dNative", &TfParser::ParseDepthwiseConv2D },
{ "FusedBatchNorm", &TfParser::ParseFusedBatchNorm },
{ "ConcatV2", &TfParser::ParseConcat },
{ "LRN", &TfParser::ParseLrn },
{ "MatMul", &TfParser::ParseMatMul },
{ "Mul", &TfParser::ParseMul },
{ "Placeholder", &TfParser::ParsePlaceholder },
{ "Relu", &TfParser::ParseRelu },
{ "Relu6", &TfParser::ParseRelu6 },
{ "Reshape", &TfParser::ParseReshape },
{ "ResizeBilinear", &TfParser::ParseResizeBilinear },
{ "Shape", &TfParser::ParseShape },
{ "Squeeze", &TfParser::ParseSqueeze },
{ "Sigmoid", &TfParser::ParseSigmoid },
{ "Softmax", &TfParser::ParseSoftmax },
{ "Softplus", &TfParser::ParseSoftplus },
{ "Tanh", &TfParser::ParseTanh },
{ "MaxPool", &TfParser::ParseMaxPool },
{ "AvgPool", &TfParser::ParseAvgPool },
};
ITfParser* ITfParser::CreateRaw()
{
return new TfParser();
}
ITfParserPtr ITfParser::Create()
{
return ITfParserPtr(CreateRaw(), &ITfParser::Destroy);
}
void ITfParser::Destroy(ITfParser* parser)
{
delete parser;
}
inline void CalculateSamePadding(uint32_t inputSize, uint32_t stride,
uint32_t filterSize, bool samePadding,
uint32_t* paddingFront, uint32_t* paddingBack) {
*paddingFront = 0;
*paddingBack = 0;
if (samePadding) {
uint32_t outputSize = (inputSize + stride - 1) / stride;
uint32_t temp = (outputSize - 1) * stride + filterSize;
if (temp > inputSize) {
*paddingFront = (temp - inputSize) / 2;
*paddingBack = (temp - inputSize) - *paddingFront;
}
}
}
void CalcPadding(uint32_t input, uint32_t kernel, uint32_t stride, uint32_t& outPadHead, uint32_t& outPadTail,
bool samePadding)
{
CalculateSamePadding(input, stride, kernel, samePadding, &outPadHead, &outPadTail);
}
/// An Abstract base class which represents a single tensorflow operation (node)
/// that has been (potentially partially) converted to Armnn.
/// It may not yet have been fully converted into actual Armnn layers.
class ParsedTfOperation
{
public:
ParsedTfOperation(TfParser* parser, const tensorflow::NodeDef& node)
: m_Parser(parser)
, m_Node(node)
{
}
virtual ~ParsedTfOperation() {};
const tensorflow::NodeDef& GetNode() const { return m_Node; }
/// Gets the ArmNN IOutputSlot corresponding to the given output index of the Tensorflow operation.
/// This may result in the creation of Armnn layers if this was deferred (e.g. see ParsedConstTfOperation).
virtual IOutputSlot& ResolveArmnnOutputSlot(unsigned int tfOutputIndex) = 0;
/// If this operation is an Identity then this will follow return the 'parent' operation (recursively).
virtual ParsedTfOperation* ResolveIdentityOperations()
{
return this;
}
protected:
TfParser* m_Parser;
const tensorflow::NodeDef& m_Node;
};
/// An ParsedTfOperation where the Armnn equivalent is a single layer,
/// with output slots that correspond directly to the Tf node outputs.
class SingleLayerParsedTfOperation : public ParsedTfOperation
{
public:
SingleLayerParsedTfOperation(TfParser* parser, const tensorflow::NodeDef& node, IConnectableLayer* layer)
: ParsedTfOperation(parser, node)
, m_Layer(layer)
{
}
IOutputSlot& ResolveArmnnOutputSlot(unsigned int tfOutputIndex) override
{
BOOST_ASSERT(m_Layer);
// Assume one-to-one mapping between Tf and armnn output slots.
unsigned int armnnOutputSlotIdx = tfOutputIndex;
if (armnnOutputSlotIdx >= m_Layer->GetNumOutputSlots())
{
throw ParseException(
boost::str(boost::format("The requested output slot #%1% "
"for %2% does not exist") % armnnOutputSlotIdx % m_Layer->GetName()));
}
return m_Layer->GetOutputSlot(armnnOutputSlotIdx);
}
protected:
IConnectableLayer* m_Layer;
};
/// A SingleLayerParsedTfOperation for deferred layer creation
class DeferredSingleLayerParsedTfOperation : public SingleLayerParsedTfOperation
{
public:
DeferredSingleLayerParsedTfOperation(TfParser* parser, const tensorflow::NodeDef& node)
: SingleLayerParsedTfOperation(parser, node, nullptr)
{
}
IOutputSlot& ResolveArmnnOutputSlot(unsigned int tfOutputIndex) override
{
if (!m_Layer)
{
CreateLayerDeferred();
}
return SingleLayerParsedTfOperation::ResolveArmnnOutputSlot(tfOutputIndex);
}
private:
virtual void CreateLayerDeferred() = 0;
};
TfParser::TfParser()
: m_Network(nullptr, nullptr)
{
}
const tensorflow::NodeDef* TfParser::ResolveIdentityNode(const tensorflow::NodeDef* nodeDef)
{
if (nodeDef->op() != "Identity")
{
return nodeDef;
}
if (nodeDef->input_size() != 1)
{
throw ParseException("Identity node does not have correct amount of inputs!");
}
auto it = m_NodesByName.find(nodeDef->input(0));
if (it != m_NodesByName.end())
{
const tensorflow::NodeDef* inputNode = it->second;
return ResolveIdentityNode(inputNode);
}
else
{
throw ParseException("Cannot find what the Identity node is linked to!");
}
}
std::vector<OutputOfConstNodeDef>
TfParser::GetTfInputNodes(const tensorflow::NodeDef& nodeDef) const
{
std::vector<OutputOfConstNodeDef> ret;
if (nodeDef.op() == "Const")
{
// For some reason const node can have "Control Inputs". We ignore them for now.
return ret;
}
ret.reserve(boost::numeric_cast<size_t>(nodeDef.input_size()));
for (int j = 0; j < nodeDef.input_size(); ++j)
{
OutputId outputId = ParseOutputId(nodeDef.input(j));
if (nodeDef.input(j)[0] == '^') // I couldn't find a better test for control inputs.
{
throw ParseException(
"Node '" + nodeDef.name() + "' has Control Input '" + nodeDef.input(j) + "' which is unsupported.");
}
auto inputIt = m_NodesByName.find(outputId.m_IndexedValue);
if (inputIt == m_NodesByName.end())
{
throw ParseException(
"Can't find node '" + nodeDef.input(j) +
"', which is listed as an input of '" + nodeDef.name() + "'");
}
ret.push_back(OutputOfConstNodeDef(inputIt->second,outputId.m_Index));
}
return ret;
}
std::vector<OutputOfParsedTfOperation>
TfParser::GetInputParsedTfOperationsChecked(const tensorflow::NodeDef& nodeDef,
std::size_t expectedNumInputs)
{
// Fetch the tensorflow nodes connected as inputs and validate the size.
std::vector<OutputOfConstNodeDef> nodes = GetTfInputNodes(nodeDef);
const std::size_t numInputs = nodes.size();
if (numInputs != expectedNumInputs)
{
throw ParseException(boost::str(boost::format("Unexpected number of inputs for node %1%. "
"Expected %2%, found %3%") % nodeDef.name() % expectedNumInputs % numInputs));
}
// Fetch the corresponding ParsedTfOperation operations
std::vector<OutputOfParsedTfOperation> result;
for (auto&& node : nodes)
{
auto it = m_ParsedTfOperations.find(node.m_IndexedValue->name());
if (it == m_ParsedTfOperations.end())
{
throw ParseException("Node with name '" + node.m_IndexedValue->name() + "' has not been parsed");
}
ParsedTfOperation* parsedOp = it->second.get();
// Transparently 'skip' any Identity operations. This simplifies the logic inside the ParseXXX() functions.
parsedOp = parsedOp->ResolveIdentityOperations();
result.push_back(OutputOfParsedTfOperation(parsedOp,node.m_Index));
}
return result;
}
ParsedTfOperationPtr TfParser::ParseAdd(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2);
// If one of the inputs is a MatMul and the other is a const, then we handle both nodes together as FullyConnected
if (inputs[0].m_IndexedValue->GetNode().op() == "MatMul" &&
HasParsedConstTensor<float>(inputs[1].m_IndexedValue->GetNode().name()))
{
IConnectableLayer* layer =
AddFullyConnectedLayer(inputs[0].m_IndexedValue->GetNode(),
&nodeDef,nodeDef.name().c_str());
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
else if (HasParsedConstTensor<float>(inputs[0].m_IndexedValue->GetNode().name()) &&
inputs[1].m_IndexedValue->GetNode().op() == "MatMul")
{
IConnectableLayer* layer =
AddFullyConnectedLayer(inputs[1].m_IndexedValue->GetNode(),
&nodeDef,nodeDef.name().c_str());
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
else
{
// Otherwise it's just a regular addition
return AddAdditionLayer(nodeDef);
}
}
ParsedTfOperationPtr TfParser::ParseBiasAdd(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
return AddAdditionLayer(nodeDef, true);
}
/// An ParsedTfOperation which forwards to another (used for Identity nodes).
class ParsedIdentityTfOperation : public ParsedTfOperation
{
public:
ParsedIdentityTfOperation(TfParser* parser, const tensorflow::NodeDef& node, ParsedTfOperation* representative)
: ParsedTfOperation(parser, node)
, m_Representative(representative)
{
}
virtual IOutputSlot& ResolveArmnnOutputSlot(unsigned int tfOutputIndex) override
{
BOOST_ASSERT(m_Representative);
return m_Representative->ResolveArmnnOutputSlot(tfOutputIndex);
}
virtual ParsedTfOperation* ResolveIdentityOperations() override
{
return m_Representative->ResolveIdentityOperations();
}
private:
ParsedTfOperation* m_Representative;
};
ParsedTfOperationPtr TfParser::ParseIdentity(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1);
// Any requests for the output slots of this node should be forwarded to the node connected as input.
return std::make_unique<ParsedIdentityTfOperation>(this, nodeDef, inputs[0].m_IndexedValue);
}
/// An ParsedTfOperation for a Const node.
/// Creation of the armnn ConstLayer is deferred until it is actually needed, because Const nodes are mostly used
/// for weight inputs to MatMul/Conv2D nodes and in these cases armnn doesn't need a ConstLayer.
template <typename T>
class ParsedConstTfOperation : public DeferredSingleLayerParsedTfOperation
{
public:
ParsedConstTfOperation(TfParser* parser, const tensorflow::NodeDef& node,
const T* tensorData, const TensorInfo& tensorInfo)
: DeferredSingleLayerParsedTfOperation(parser, node),
m_Storage(tensorData, tensorData + tensorInfo.GetNumElements()),
m_TensorInfo(tensorInfo)
{
BOOST_ASSERT(tensorInfo.GetDataType() == GetDataType<T>());
}
void CreateLayerDeferred() override
{
BOOST_ASSERT(m_Layer == nullptr);
m_Layer = m_Parser->m_Network->AddConstantLayer(ConstTensor(m_TensorInfo, m_Storage), m_Node.name().c_str());
m_Layer->GetOutputSlot(0).SetTensorInfo(m_TensorInfo);
}
ConstTensor GetConstTensor(bool swizzleForConvolutionWeights, std::vector<T>& outputTensorData) const
{
// Mappings from TensorFlow filter tensors to the ArmNN filter tensors.
// Tensorflow weights are [H, W, In, Out]
// ArmNN weights are [Out, In, H, W]
static const PermutationVector HWIOToOIHW = {2, 3, 1, 0};
const TensorInfo outInfo = swizzleForConvolutionWeights
? armnnUtils::Permuted(m_TensorInfo, HWIOToOIHW)
: m_TensorInfo;
outputTensorData.resize(m_TensorInfo.GetNumElements());
// Copy or swizzle from the permanent storage into the storage the caller provided.
if (swizzleForConvolutionWeights)
{
armnnUtils::Permute(outInfo.GetShape(), HWIOToOIHW, m_Storage.data(), outputTensorData.data());
}
else
{
memcpy(outputTensorData.data(), m_Storage.data(), m_TensorInfo.GetNumBytes());
}
// Update the result to point to the user provided storage
ConstTensor constTensor(outInfo, outputTensorData);
return constTensor;
}
private:
///< Manages the lifetime of the tensor data.
std::vector<T> m_Storage;
///< Describes the layout of the tensor and points to the data in m_Storage.
TensorInfo m_TensorInfo;
};
DataType ConvertTfTensorDataType(const tensorflow::DataType tfDataType)
{
switch (tfDataType)
{
case tensorflow::DT_FLOAT:
return DataType::Float32;
break;
case tensorflow::DT_INT32:
return DataType::Signed32;
break;
default:
throw ParseException(boost::str(
boost::format("Unknown DataType %1% for node")
% tensorflow::DataType_Name(tfDataType)));
}
}
struct ParseTfTensorValueList
{
template<typename DataType>
static void Parse(
const tensorflow::TensorProto& tfTensor,
unsigned int dstElements,
std::vector<int8_t>& outputData);
template <typename DataType>
static void ReadData(const void* srcData, unsigned int numSrcElements,
std::vector<int8_t>& dstData, unsigned int numDstElements)
{
// If there are no entries in the list, perform no action
if (numSrcElements == 0)
{
return;
}
// If no size was provided, use the length of the value list
if (numDstElements == 0)
{
numDstElements = numSrcElements;
}
// Allocate memory
dstData.resize(std::max(numSrcElements, numDstElements) * sizeof(DataType));
const DataType* srcTensor = reinterpret_cast<const DataType*>(srcData);
DataType* dstTensor = reinterpret_cast<DataType*>(dstData.data());
// Copy the value list entries into the destination
std::copy(srcTensor, srcTensor + numSrcElements, dstTensor);
if (numDstElements > numSrcElements)
{
// Use the last element in the list to fill the remaining entries
std::fill(dstTensor + numSrcElements, dstTensor + numDstElements, srcTensor[numSrcElements - 1]);
}
}
};
template <>
void ParseTfTensorValueList::Parse<float>(const tensorflow::TensorProto& tfTensor,
unsigned int dstElements, std::vector<int8_t>& outputData)
{
ReadData<float>(tfTensor.float_val().data(), static_cast<unsigned int>(tfTensor.float_val_size()),
outputData, dstElements);
}
template <>
void ParseTfTensorValueList::Parse<int32_t>(const tensorflow::TensorProto& tfTensor,
unsigned int dstElements, std::vector<int8_t>& outputData)
{
ReadData<int32_t>(tfTensor.int_val().data(), static_cast<unsigned int>(tfTensor.int_val_size()),
outputData, dstElements);
}
template <template<typename> class OperatorType, typename T = int8_t>
struct MakeTfOperation
{
template<typename DataType, class... Args>
inline static std::unique_ptr<OperatorType<DataType>> Parse(TfParser* parser, const tensorflow::NodeDef& node,
Args&&... args)
{
return std::make_unique<OperatorType<DataType>>(parser, node, std::forward<Args>(args)...);
}
};
template <>
struct MakeTfOperation<ParsedConstTfOperation>
{
template<typename DataType, class... Args>
inline static std::unique_ptr<ParsedConstTfOperation<DataType>> Parse(TfParser* parser,
const tensorflow::NodeDef& node, const std::vector<int8_t>& tensorData, const TensorInfo& tensorInfo)
{
return std::make_unique<ParsedConstTfOperation<DataType>>(parser, node,
reinterpret_cast<const DataType*>(tensorData.data()), tensorInfo);
}
};
template <class FuncType>
struct InvokeParseFunction
{
template<class ResType, class... Args>
inline static ResType Result(DataType dataType, Args&&... args)
{
if (dataType == DataType::Float32)
{
return FuncType::template Parse<float>(std::forward<Args>(args)...);
}
else if (dataType == DataType::Signed32)
{
return FuncType::template Parse<int32_t>(std::forward<Args>(args)...);
}
return ResType();
}
template<class... Args>
inline static void Result(DataType dataType, Args&&... args)
{
if (dataType == DataType::Float32)
{
FuncType::template Parse<float>(std::forward<Args>(args)...);
}
else if (dataType == DataType::Signed32)
{
FuncType::template Parse<int32_t>(std::forward<Args>(args)...);
}
}
};
ParsedTfOperationPtr TfParser::ParseConst(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
BOOST_ASSERT(nodeDef.op() == "Const");
if (nodeDef.attr().count("value") == 0)
{
throw ParseException(boost::str(
boost::format("Value not found for Const node - %1%")
% nodeDef.name()));
}
const tensorflow::TensorProto& tfTensor = nodeDef.attr().at("value").tensor();
const tensorflow::TensorShapeProto& tfTensorShape = tfTensor.tensor_shape();
const tensorflow::DataType tfDataType = ReadMandatoryNodeTypeAttribute(nodeDef, "dtype");
const auto GetDimensionSize = [](auto& d) { return d.size(); };
std::vector<unsigned int> dimensionSizes;
std::transform(tfTensorShape.dim().begin(), tfTensorShape.dim().end(),
std::back_inserter(dimensionSizes), GetDimensionSize);
// Calculate number of elements
const DataType dataType = ConvertTfTensorDataType(tfDataType);
unsigned int numElements = 0U;
if (!dimensionSizes.empty())
{
numElements = std::accumulate(dimensionSizes.begin(), dimensionSizes.end(),
1U, std::multiplies<unsigned int>());
}
std::vector<int8_t> tensorData;
// Get tensor data from the list of values attribute
if (tfTensor.tensor_content().empty())
{
InvokeParseFunction<ParseTfTensorValueList>::Result<void>(dataType, tfTensor, numElements, tensorData);
// If the tensor shape is not defined, but there is a value list, then interpret the data as a 1D
// tensor of the provided number of elements
if (numElements == 0)
{
const unsigned int tfNumElements = static_cast<unsigned int>(tensorData.size()) / GetDataTypeSize(dataType);
dimensionSizes.push_back(tfNumElements);
}
}
// Get tensor data from tensor content attribute
else
{
tensorData.assign(tfTensor.tensor_content().begin(), tfTensor.tensor_content().end());
// Check if a tensor shape is defined for the tensor content
if (numElements == 0)
{
throw ParseException(boost::str(
boost::format("No tensor shape found for Const node - %1%")
% nodeDef.name()));
}
}
// Const node requires at least a list of values or a content attribute
if (tensorData.empty())
{
throw ParseException(boost::str(
boost::format("No tensor data found for Const node - %1%")
% nodeDef.name()));
}
const TensorInfo tensorInfo(static_cast<unsigned int>(dimensionSizes.size()), dimensionSizes.data(), dataType);
// If we have a list of values, then the length of the list must be
// less than or equal to the number of elements implied by the shape argument
if (tensorData.size() > tensorInfo.GetNumBytes())
{
throw ParseException(boost::str(
boost::format("Number of elements (%1%) should be less than or equal \
to the number of elements implied by the shape argument (%2%) for Const node - %3%")
% (tensorData.size() / GetDataTypeSize(dataType))
% tensorInfo.GetNumElements()
% nodeDef.name()));
}
return InvokeParseFunction<MakeTfOperation<ParsedConstTfOperation>>::Result<ParsedTfOperationPtr>(
dataType, this, nodeDef, tensorData, tensorInfo);
}
template<typename Type>
bool TfParser::HasParsedConstTensor(const std::string & nodeName) const
{
auto it = m_ParsedTfOperations.find(nodeName);
if (it == m_ParsedTfOperations.end() ||
dynamic_cast<ParsedConstTfOperation<Type>*>(it->second.get()) == nullptr)
{
return false;
}
else
{
return true;
}
}
ParsedTfOperationPtr TfParser::ParseConv2D(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2);
IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
TensorInfo inputTensorInfo = inputSlot.GetTensorInfo();
if (!HasParsedConstTensor<float>(inputs[1].m_IndexedValue->GetNode().name()))
{
throw ParseException("ArmNN only supports Convolution layers with constant weights");
}
ParsedConstTfOperation<float>* weightNode =
boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[1].m_IndexedValue);
std::string paddingString = ReadMandatoryNodeStringAttribute(nodeDef, "padding");
std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format");
std::vector<uint32_t> strides = ReadMandatoryNodeUint32ListAttribute(nodeDef, "strides");
// read the dilations, if present - only [1,1,1,1] (the default) is supported
std::vector<uint32_t> dilations = ReadOptionalNodeUint32ListAttribute(nodeDef, "dilations");
if (!dilations.empty())
{
for (auto dilation : dilations)
{
if (dilation != 1u)
{
throw ParseException("ArmNN only supports Convolution layers with dilations [1,1,1,1]");
}
}
}
Convolution2dDescriptor desc;
desc.m_BiasEnabled = false;
if (dataFormat == "NHWC")
{
desc.m_StrideX = strides[2];
desc.m_StrideY = strides[1];
// Swizzle input to supported memory layout
inputTensorInfo = armnnUtils::Permuted(inputSlot.GetTensorInfo(), NHWCToArmNN);
}
else if (dataFormat == "NCHW")
{
desc.m_StrideX = strides[3];
desc.m_StrideY = strides[2];
}
else
{
throw ParseException("Unsupported data format passed for Conv2D. Only NHWC and NCHW supported");
}
uint32_t inputHeight = inputTensorInfo.GetShape()[2];
uint32_t inputWidth = inputTensorInfo.GetShape()[3];
std::vector<float> outputTensorData;
ConstTensor weightTensor = weightNode->GetConstTensor(true, outputTensorData);
uint32_t weightHeight = weightTensor.GetShape()[2];
uint32_t weightWidth = weightTensor.GetShape()[3];
bool padding = false;
TensorInfo outputInfo;
if (paddingString == "SAME")
{
padding = true;
outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0],
weightTensor.GetShape()[0],
static_cast<uint32_t>(ceil(
static_cast<float>(inputHeight) /
static_cast<float>(desc.m_StrideY))),
static_cast<uint32_t>(ceil(
static_cast<float>(inputWidth) /
static_cast<float>(desc.m_StrideX)))
}, DataType::Float32);
}
else if (paddingString == "VALID")
{
padding = false;
outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0],
weightTensor.GetShape()[0],
static_cast<uint32_t>(ceil(
static_cast<float>(inputHeight - weightHeight + 1) /
static_cast<float>(desc.m_StrideY))),
static_cast<uint32_t>(ceil(
static_cast<float>(inputWidth - weightWidth + 1) /
static_cast<float>(desc.m_StrideX)))
}, DataType::Float32);
}
else
{
throw ParseException("Only 'SAME' and 'VALID' padding supported");
}
CalcPadding(inputHeight, weightHeight, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, padding);
CalcPadding(inputWidth, weightWidth, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, padding);
IConnectableLayer* layer = m_Network->AddConvolution2dLayer(desc, weightTensor, nodeDef.name().c_str());
layer->GetOutputSlot(0).SetTensorInfo(outputInfo);
if (dataFormat == "NHWC")
{
layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name());
}
else
{
inputSlot.Connect(layer->GetInputSlot(0));
}
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParseDepthwiseConv2D(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2);
IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
TensorInfo inputTensorInfo = inputSlot.GetTensorInfo();
if (!HasParsedConstTensor<float>(inputs[1].m_IndexedValue->GetNode().name()))
{
throw ParseException("ArmNN only supports Depthwise Convolution layers with constant weights");
}
ParsedConstTfOperation<float>* weightNode =
boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[1].m_IndexedValue);
std::string paddingString = ReadMandatoryNodeStringAttribute(nodeDef, "padding");
std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format");
std::vector<uint32_t> strides = ReadMandatoryNodeUint32ListAttribute(nodeDef, "strides");
DepthwiseConvolution2dDescriptor desc;
desc.m_BiasEnabled = false;
if (dataFormat == "NHWC")
{
desc.m_StrideX = strides[2];
desc.m_StrideY = strides[1];
// Swizzle input to supported memory layout
inputTensorInfo = armnnUtils::Permuted(inputSlot.GetTensorInfo(), NHWCToArmNN);
}
else if (dataFormat == "NCHW")
{
desc.m_StrideX = strides[3];
desc.m_StrideY = strides[2];
}
else
{
throw ParseException("Unsupported data format passed for DepthwiseConv2dNative. Only NHWC and NCHW supported");
}
uint32_t inputHeight = inputTensorInfo.GetShape()[2];
uint32_t inputWidth = inputTensorInfo.GetShape()[3];
std::vector<float> outputTensorData;
ConstTensor weightTensor = weightNode->GetConstTensor(true, outputTensorData);
uint32_t weightHeight = weightTensor.GetShape()[2];
uint32_t weightWidth = weightTensor.GetShape()[3];
bool padding = false;
TensorInfo outputInfo;
if (paddingString == "SAME")
{
padding = true;
outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0],
weightTensor.GetShape()[0] * weightTensor.GetShape()[1],
static_cast<uint32_t>(ceil(
static_cast<float>(inputHeight) /
static_cast<float>(desc.m_StrideY))),
static_cast<uint32_t>(ceil(
static_cast<float>(inputWidth) /
static_cast<float>(desc.m_StrideX)))
}, DataType::Float32);
}
else if (paddingString == "VALID")
{
padding = false;
outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0],
weightTensor.GetShape()[0] * weightTensor.GetShape()[1],
static_cast<uint32_t>(ceil(
static_cast<float>(inputHeight - weightHeight + 1) /
static_cast<float>(desc.m_StrideY))),
static_cast<uint32_t>(ceil(
static_cast<float>(inputWidth - weightWidth + 1) /
static_cast<float>(desc.m_StrideX)))
}, DataType::Float32);
}
else
{
throw ParseException("Only 'SAME' and 'VALID' padding supported");
}
CalcPadding(inputHeight, weightHeight, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, padding);
CalcPadding(inputWidth, weightWidth, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, padding);
IConnectableLayer* layer = m_Network->AddDepthwiseConvolution2dLayer(desc, weightTensor, nodeDef.name().c_str());
layer->GetOutputSlot(0).SetTensorInfo(outputInfo);
if (dataFormat == "NHWC")
{
layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name());
}
else
{
inputSlot.Connect(layer->GetInputSlot(0));
}
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParseFusedBatchNorm(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 5);
if (!HasParsedConstTensor<float>(inputs[1].m_IndexedValue->GetNode().name()))
{
throw ParseException("ArmNN only supports FusedBatchNormalization layers with constant scale");
}
ParsedConstTfOperation<float>* scaleNode =
boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[1].m_IndexedValue);
if (!HasParsedConstTensor<float>(inputs[2].m_IndexedValue->GetNode().name()))
{
throw ParseException("ArmNN only supports FusedBatchNormalization layers with constant offset");
}
ParsedConstTfOperation<float>* offsetNode =
boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[2].m_IndexedValue);
if (!HasParsedConstTensor<float>(inputs[3].m_IndexedValue->GetNode().name()))
{
throw ParseException("ArmNN only supports FusedBatchNormalization layers with constant mean");
}
ParsedConstTfOperation<float>* meanNode =
boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[3].m_IndexedValue);
if (!HasParsedConstTensor<float>(inputs[4].m_IndexedValue->GetNode().name()))
{
throw ParseException("ArmNN only supports FusedBatchNormalization layers with constant variance");
}
ParsedConstTfOperation<float>* varianceNode =
boost::polymorphic_downcast<ParsedConstTfOperation<float> *>(inputs[4].m_IndexedValue);
// The descriptor only has the epsilon attribute
BatchNormalizationDescriptor desc;
desc.m_Eps = ReadMandatoryNodeFloatAttribute(nodeDef, "epsilon");
// data for the parsed tensor args (scale, offset, mean, variance) must be stored locally until the layer is added
std::vector<float> scaleTensorData;
ConstTensor scaleTensor = scaleNode->GetConstTensor(false, scaleTensorData);
std::vector<float> offsetTensorData;
ConstTensor offsetTensor = offsetNode->GetConstTensor(false, offsetTensorData);
std::vector<float> meanTensorData;
ConstTensor meanTensor = meanNode->GetConstTensor(false, meanTensorData);
std::vector<float> varianceTensorData;
ConstTensor varianceTensor = varianceNode->GetConstTensor(false, varianceTensorData);
IConnectableLayer* layer = m_Network->AddBatchNormalizationLayer(desc,
meanTensor,
varianceTensor,
offsetTensor,
scaleTensor,
nodeDef.name().c_str());
IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
const std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format");
if (dataFormat == "NHWC")
{
const TensorInfo outputTensorInfo = armnnUtils::Permuted(inputSlot.GetTensorInfo(), NHWCToArmNN);
layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name());
}
else
{
layer->GetOutputSlot(0).SetTensorInfo(inputSlot.GetTensorInfo());
inputSlot.Connect(layer->GetInputSlot(0));
}
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParseConcat(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfConstNodeDef> nodes = GetTfInputNodes(nodeDef);
// In tensorflow, we have the last input of the Concat layer as the axis for concatenation
unsigned int numInputs = static_cast<unsigned int>(nodes.size());
unsigned int numConcatView = numInputs - 1;
OriginsDescriptor concatDescriptor(static_cast<uint32_t>(numConcatView), MaxNumOfTensorDimensions);
std::vector<unsigned int>mergeDimSizes(MaxNumOfTensorDimensions, 0u);
unsigned int mergeDim = 0;
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, numInputs);
// The last input is the axis for concatenation
if (!HasParsedConstTensor<int32_t>(inputs[numInputs - 1].m_IndexedValue->GetNode().name()))
{
throw ParseException("ArmNN only supports Concat with constant axis");
}
ParsedConstTfOperation<int32_t>* shapeNode =
boost::polymorphic_downcast<ParsedConstTfOperation<int32_t>*>(inputs[numInputs - 1].m_IndexedValue);
std::vector<int32_t> axisTensorData;
ConstTensor axisTensor = shapeNode->GetConstTensor(false, axisTensorData);
// This concatDim indicates the data format: 3 is the NHWC, 1 is the NCHW
const unsigned int concatDimInput = static_cast<unsigned int>(axisTensorData[0]);
// Armnn supports concatenation along the channel dimension for data format NHWC and NCHW
if (concatDimInput == 0 || concatDimInput == 2)
{
throw ParseException("The dimension for concatenation is not supported by Armnn");
}
// This is the only concatDim we support in Armnn
const unsigned int concatDim = 1;
for (unsigned int viewIndex = 0; viewIndex < numConcatView; ++viewIndex)
{
// need to double check whether it should be
IOutputSlot& inputSlot =
inputs[viewIndex].m_IndexedValue->ResolveArmnnOutputSlot(inputs[viewIndex].m_Index);
TensorInfo inputTensorInfo = inputSlot.GetTensorInfo();
if (inputTensorInfo.GetNumDimensions() != MaxNumOfTensorDimensions)
{
throw ParseException("The number of dimensions for input tensors of the concatenation op should be 4");
}
if (concatDimInput == 3)
{
inputTensorInfo = armnnUtils::Permuted(inputTensorInfo, NHWCToArmNN);
}
for (unsigned int dim = 0; dim < MaxNumOfTensorDimensions; ++dim)
{
mergeDimSizes[dim] = inputTensorInfo.GetShape()[dim];
}
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 < MaxNumOfTensorDimensions; ++j)
{
concatDescriptor.SetViewOriginCoord(viewIndex, j, 0);
}
}
mergeDimSizes[concatDim] = mergeDim;
armnn::IConnectableLayer *layer = m_Network->AddMergerLayer(concatDescriptor, nodeDef.name().c_str());
layer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo(MaxNumOfTensorDimensions, mergeDimSizes.data(),
DataType::Float32));
for (unsigned int v = 0; v < numConcatView; ++v)
{
IOutputSlot& inputSlot = inputs[v].m_IndexedValue->ResolveArmnnOutputSlot(inputs[v].m_Index);
if (concatDimInput == 3)
{
IConnectableLayer* const swizzleLayer = AddSwizzleLayer(*m_Network, inputSlot, NHWCToArmNN,
"swizzle_for-" + nodeDef.name());
swizzleLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(v));
}
else
{
inputSlot.Connect(layer->GetInputSlot(v));
}
}
if (concatDimInput == 3)
{
IConnectableLayer* const deswizzleLayer = AddSwizzleLayer(*m_Network, layer->GetOutputSlot(0), ArmNNToNHWC,
"deswizzle_for-" + nodeDef.name());
layer = deswizzleLayer;
}
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParseShape(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
// Note: The Shape layer is handled in a special way, because:
// 1. ARMNN doesn't support int32 tensors which it outputs
// 2. ARMNN works with statically shaped tensors which are known at parse time
// 3. because of 1. and 2. we treat the output of Shape as a temporary const int32
// tensor which may be used as an input to other ops, most likely a Reshape
const tensorflow::DataType tfDataType = ReadMandatoryNodeTypeAttribute(nodeDef, "out_type");
if (tfDataType != tensorflow::DT_INT32)
{
throw ParseException("Armnn only supports DT_INT32 as out_type");
}
const std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1);
IOutputSlot& prevLayerOutputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
const TensorInfo& prevLayerTensorInfo = prevLayerOutputSlot.GetTensorInfo();
unsigned int prevLayerDimensions = prevLayerTensorInfo.GetNumDimensions();
std::vector<int32_t> shapeTensorData;
shapeTensorData.reserve(prevLayerDimensions);
for (unsigned int i=0; i<prevLayerDimensions; ++i)
{
shapeTensorData.push_back(static_cast<int32_t>(prevLayerTensorInfo.GetShape()[i]));
}
TensorInfo shapeTensorInfo(1, &prevLayerDimensions, DataType::Signed32);
return std::make_unique<ParsedConstTfOperation<int32_t>>(this,
nodeDef,
&shapeTensorData[0],
shapeTensorInfo);
}
ParsedTfOperationPtr TfParser::ParseReshape(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2);
ParsedTfOperation* inputNode = inputs[0].m_IndexedValue;
if (!HasParsedConstTensor<int32_t>(inputs[1].m_IndexedValue->GetNode().name()))
{
throw ParseException("ArmNN only supports Reshape layers with constant shapes");
}
ParsedConstTfOperation<int32_t>* shapeNode =
boost::polymorphic_downcast<ParsedConstTfOperation<int32_t>*>(inputs[1].m_IndexedValue);
armnn::IOutputSlot& prevLayerOutputSlot = inputNode->ResolveArmnnOutputSlot(inputs[0].m_Index);
TensorInfo inputTensorInfo = prevLayerOutputSlot.GetTensorInfo();
std::vector<int32_t> shapeTensorData;
ConstTensor shapeTensor = shapeNode->GetConstTensor(false, shapeTensorData);
const TensorInfo outputTensorInfo = PrepareReshape(inputTensorInfo, shapeTensorData);
TensorShape targetShape = outputTensorInfo.GetShape();
ReshapeDescriptor reshapeDesc;
reshapeDesc.m_TargetShape = targetShape;
IConnectableLayer* layer = m_Network->AddReshapeLayer(reshapeDesc, nodeDef.name().c_str());
prevLayerOutputSlot.Connect(layer->GetInputSlot(0));
layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParseResizeBilinear(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2);
if (!HasParsedConstTensor<int32_t>(inputs[1].m_IndexedValue->GetNode().name()))
{
throw ParseException("ArmNN only supports ResizeBilinear layers with constant sizes");
}
ParsedConstTfOperation<int32_t>* sizeNode =
boost::polymorphic_downcast<ParsedConstTfOperation<int32_t>*>(inputs[1].m_IndexedValue);
// Check the align_corners attribute is not set
if (ReadOptionalNodeBoolAttribute(nodeDef, "align_corners", false))
{
throw ParseException("ArmNN only supports ResizeBilinear layers with align_corners set to false");
}
// data for the parsed tensor args (size) must be stored locally
std::vector<int32_t> sizeTensorData;
ConstTensor sizeTensor = sizeNode->GetConstTensor(false, sizeTensorData);
// The descriptor only has target height and width attributes, which we get from the size tensor
ResizeBilinearDescriptor desc;
desc.m_TargetHeight = static_cast<uint32_t> (sizeTensorData[0]);
desc.m_TargetWidth = static_cast<uint32_t> (sizeTensorData[1]);
IConnectableLayer* layer = m_Network->AddResizeBilinearLayer(desc, nodeDef.name().c_str());
IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
TensorInfo inputTensorInfo = inputSlot.GetTensorInfo();
// the input shape is always in BHWC format, this will be swizzled below; for now,
// get the batch and channels to make up the ArmNN output shape with the target size
unsigned int outBatch = inputTensorInfo.GetShape()[0];
unsigned int outChannels = inputTensorInfo.GetShape()[3];
unsigned int outHeight = desc.m_TargetHeight;
unsigned int outWidth = desc.m_TargetWidth;
TensorShape outShape({outBatch, outChannels, outHeight, outWidth});
// The output DataType is always Float32, regardless of the input DataType
const TensorInfo outputTensorInfo(outShape, armnn::DataType::Float32);
layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
// TensorFlow ResizeBilinear input is always in BHWC format, so add swizzle and deswizzle layers
layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name());
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
TensorInfo OutputShapeOfSqueeze(const tensorflow::NodeDef& nodeDef, TensorInfo inputTensorInfo)
{
BOOST_ASSERT(nodeDef.op() == "Squeeze");
tensorflow::DataType tfDataType = ReadMandatoryNodeTypeAttribute(nodeDef, "T");
DataType type;
if (tfDataType == tensorflow::DT_FLOAT)
{
type = DataType::Float32;
}
else if (tfDataType == tensorflow::DT_INT32)
{
type = DataType::Signed32;
}
else
{
throw ParseException(boost::str(
boost::format("Unsupported DataType %1% for Squeeze operation")
% tensorflow::DataType_Name(tfDataType)));
}
std::vector<uint32_t> squeezeDims = ReadOptionalNodeUint32ListAttribute(nodeDef, "squeeze_dims");
if (squeezeDims.empty())
{
for(unsigned int i = 0; i < inputTensorInfo.GetNumDimensions(); i++)
{
if (inputTensorInfo.GetShape()[i] == 1)
{
squeezeDims.push_back(i);
}
}
}
std::vector<uint32_t> outputDims;
for(unsigned int i = 0; i < inputTensorInfo.GetNumDimensions(); i++)
{
bool includeDimension = (std::find(squeezeDims.begin(), squeezeDims.end(), i) == squeezeDims.end());
if (includeDimension)
{
outputDims.push_back(inputTensorInfo.GetShape()[i]);
}
}
if (outputDims.size() > 4)
{
throw ParseException("Unsupported shape for Squeeze");
}
TensorInfo outTensorInfo = TensorInfo(boost::numeric_cast<unsigned int>(outputDims.size()),
outputDims.data(),
type);
return outTensorInfo;
}
ParsedTfOperationPtr TfParser::ParseSqueeze(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1);
IOutputSlot& prevLayerOutputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
TensorInfo inputTensorInfo = prevLayerOutputSlot.GetTensorInfo();
TensorInfo outputInfo;
outputInfo = OutputShapeOfSqueeze(nodeDef, inputTensorInfo);
ReshapeDescriptor reshapeDesc;
reshapeDesc.m_TargetShape = outputInfo.GetShape();
IConnectableLayer* layer = m_Network->AddReshapeLayer(reshapeDesc, nodeDef.name().c_str());
prevLayerOutputSlot.Connect(layer->GetInputSlot(0));
layer->GetOutputSlot(0).SetTensorInfo(outputInfo);
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParseLrn(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1);
NormalizationDescriptor normalizationDescriptor;
normalizationDescriptor.m_NormMethodType = NormalizationAlgorithmMethod::LocalBrightness;
normalizationDescriptor.m_NormChannelType = NormalizationAlgorithmChannel::Across;
normalizationDescriptor.m_Alpha = ReadMandatoryNodeFloatAttribute(nodeDef, "alpha");
normalizationDescriptor.m_Beta = ReadMandatoryNodeFloatAttribute(nodeDef, "beta");
normalizationDescriptor.m_K = ReadMandatoryNodeFloatAttribute(nodeDef, "bias");
normalizationDescriptor.m_NormSize = ReadMandatoryNodeUint32Attribute(nodeDef, "depth_radius");
// The window size must be an odd value. For a window size of (2 * n + 1), TensorFlow defines depth_radius = n.
normalizationDescriptor.m_NormSize = normalizationDescriptor.m_NormSize * 2 + 1;
IOutputSlot& prevLayerOutputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
IConnectableLayer* layer = m_Network->AddNormalizationLayer(normalizationDescriptor,
nodeDef.name().c_str());
const TensorInfo permutedInfo = armnnUtils::Permuted(prevLayerOutputSlot.GetTensorInfo(), NHWCToArmNN);
layer->GetOutputSlot(0).SetTensorInfo(permutedInfo);
layer = SwizzleInDeswizzleOut(*m_Network, prevLayerOutputSlot, *layer, nodeDef.name());
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
/// An ParsedTfOperation for a MatMul node.
/// Creation of the armnn FullyConnected layer is deferred until it is actually needed, because MatMul nodes are
/// often used for the first part of a biased FullyConnected (MatMul followed by Add) and in these cases armnn doesn't
/// need a separate layer for the MatMul.
class ParsedMatMulTfOperation : public DeferredSingleLayerParsedTfOperation
{
public:
ParsedMatMulTfOperation(TfParser* parser, const tensorflow::NodeDef& node)
: DeferredSingleLayerParsedTfOperation(parser, node)
{
}
void CreateLayerDeferred() override
{
BOOST_ASSERT(m_Layer == nullptr);
m_Layer = m_Parser->AddFullyConnectedLayer(m_Node, nullptr, m_Node.name().c_str());
}
};
ParsedTfOperationPtr TfParser::ParseMatMul(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
// Defer the creation of the layer (see ParsedMatMulTfOperation).
return std::make_unique<ParsedMatMulTfOperation>(this, nodeDef);
}
ParsedTfOperationPtr TfParser::ParseMul(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
boost::ignore_unused(graphDef);
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2);
IConnectableLayer* const layer = m_Network->AddMultiplicationLayer(nodeDef.name().c_str());
IOutputSlot* input0Slot = &inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
IOutputSlot* input1Slot = &inputs[1].m_IndexedValue->ResolveArmnnOutputSlot(inputs[1].m_Index);
auto const input0NumDims = input0Slot->GetTensorInfo().GetNumDimensions();
auto const input1NumDims = input1Slot->GetTensorInfo().GetNumDimensions();
if (input0NumDims < input1NumDims)
{
const bool isNHWC = true;
input0Slot = BroadcastForAddandMul(input1Slot, input0Slot, isNHWC, *m_Network, nodeDef);
}
if (input1NumDims < input0NumDims)
{
const bool isNHWC = true;
input1Slot = BroadcastForAddandMul(input0Slot, input1Slot, isNHWC, *m_Network, nodeDef);
}
input0Slot->Connect(layer->GetInputSlot(0));
input1Slot->Connect(layer->GetInputSlot(1));
if (input0NumDims < input1NumDims)
{
layer->GetOutputSlot(0).SetTensorInfo(input1Slot->GetTensorInfo());
}
else
{
layer->GetOutputSlot(0).SetTensorInfo(input0Slot->GetTensorInfo());
}
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParsePlaceholder(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
boost::ignore_unused(graphDef);
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 0);
const LayerBindingId layerId = boost::numeric_cast<LayerBindingId>(m_NetworkInputsBindingInfo.size());
auto it = m_InputShapes.find(nodeDef.name());
if (it == m_InputShapes.end())
{
throw ParseException("Missing input shape for Placeholder '" + nodeDef.name() + "'");
}
TensorInfo tensorInfo(it->second, DataType::Float32);
IConnectableLayer* const layer = m_Network->AddInputLayer(layerId, nodeDef.name().c_str());
layer->GetOutputSlot(0).SetTensorInfo(tensorInfo);
TrackInputBinding(layer, layerId, tensorInfo);
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParseRelu(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
boost::ignore_unused(graphDef);
ActivationDescriptor activationDesc;
activationDesc.m_Function = ActivationFunction::ReLu;
return AddActivationLayer(nodeDef, activationDesc);
}
ParsedTfOperationPtr TfParser::ParseRelu6(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
boost::ignore_unused(graphDef);
ActivationDescriptor activationDesc;
activationDesc.m_Function = ActivationFunction::BoundedReLu;
activationDesc.m_A = 6.0f;
activationDesc.m_B = 0.0f;
return AddActivationLayer(nodeDef, activationDesc);
}
ParsedTfOperationPtr TfParser::ParseSigmoid(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
boost::ignore_unused(graphDef);
ActivationDescriptor activationDesc;
activationDesc.m_Function = ActivationFunction::Sigmoid;
return AddActivationLayer(nodeDef, activationDesc);
}
ParsedTfOperationPtr TfParser::ParseSoftmax(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
boost::ignore_unused(graphDef);
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1);
SoftmaxDescriptor softmaxDescriptor;
IConnectableLayer* const layer = m_Network->AddSoftmaxLayer(softmaxDescriptor, nodeDef.name().c_str());
IOutputSlot& prevLayerSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
prevLayerSlot.Connect(layer->GetInputSlot(0));
layer->GetOutputSlot(0).SetTensorInfo(prevLayerSlot.GetTensorInfo());
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParseSoftplus(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
boost::ignore_unused(graphDef);
ActivationDescriptor activationDesc;
activationDesc.m_Function = ActivationFunction::SoftReLu;
return AddActivationLayer(nodeDef, activationDesc);
}
ParsedTfOperationPtr TfParser::ParseTanh(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
boost::ignore_unused(graphDef);
ActivationDescriptor activationDesc;
activationDesc.m_Function = ActivationFunction::TanH;
activationDesc.m_A = 1.0f;
activationDesc.m_B = 1.0f;
return AddActivationLayer(nodeDef, activationDesc);
}
ParsedTfOperationPtr TfParser::AddActivationLayer(const tensorflow::NodeDef& nodeDef,
ActivationDescriptor& activationDesc)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1);
IConnectableLayer* const layer = m_Network->AddActivationLayer(activationDesc, nodeDef.name().c_str());
IOutputSlot& prevLayerOutputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
prevLayerOutputSlot.Connect(layer->GetInputSlot(0));
layer->GetOutputSlot(0).SetTensorInfo(prevLayerOutputSlot.GetTensorInfo());
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::ParseMaxPool(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
return ParsePooling2d(nodeDef, graphDef, PoolingAlgorithm::Max);
}
ParsedTfOperationPtr TfParser::ParseAvgPool(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef)
{
return ParsePooling2d(nodeDef, graphDef, PoolingAlgorithm::Average);
}
ParsedTfOperationPtr TfParser::ParsePooling2d(const tensorflow::NodeDef& nodeDef,
const tensorflow::GraphDef& graphDef, PoolingAlgorithm pooltype)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 1);
IOutputSlot& inputSlot = inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
TensorInfo inputTensorInfo = inputSlot.GetTensorInfo();
if (inputs.size() != 1)
{
throw ParseException("2D Pooling expects one input!");
}
std::string paddingString = ReadMandatoryNodeStringAttribute(nodeDef, "padding");
std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format");
std::vector<uint32_t> strides = ReadMandatoryNodeUint32ListAttribute(nodeDef, "strides");
std::vector<uint32_t> ksize = ReadMandatoryNodeUint32ListAttribute(nodeDef, "ksize"); // size of pool windows
Pooling2dDescriptor pooling2dDescriptor;
pooling2dDescriptor.m_PoolType = pooltype;
pooling2dDescriptor.m_PaddingMethod = PaddingMethod::Exclude;
pooling2dDescriptor.m_OutputShapeRounding = OutputShapeRounding::Floor;
if (dataFormat == "NHWC")
{
pooling2dDescriptor.m_StrideX = strides[2];
pooling2dDescriptor.m_StrideY = strides[1];
pooling2dDescriptor.m_PoolWidth = ksize[2];
pooling2dDescriptor.m_PoolHeight = ksize[1];
// Swizzle input to supported memory layout
inputTensorInfo = armnnUtils::Permuted(inputSlot.GetTensorInfo(), NHWCToArmNN);
}
else if (dataFormat == "NCHW")
{
pooling2dDescriptor.m_StrideX = strides[3];
pooling2dDescriptor.m_StrideY = strides[2];
pooling2dDescriptor.m_PoolWidth = ksize[3];
pooling2dDescriptor.m_PoolHeight = ksize[2];
}
else
{
throw ParseException("Only NHWC or NCHW supported for Pooling2d");
}
uint32_t inputHeight = inputTensorInfo.GetShape()[2];
uint32_t inputWidth = inputTensorInfo.GetShape()[3];
bool padding = false;
TensorInfo outputInfo;
if (paddingString == "SAME")
{
padding = true;
outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0],
inputTensorInfo.GetShape()[1],
static_cast<uint32_t>(ceil(
static_cast<float>(inputHeight) /
static_cast<float>(pooling2dDescriptor.m_StrideY))),
static_cast<uint32_t>(ceil(
static_cast<float>(inputWidth) /
static_cast<float>(pooling2dDescriptor.m_StrideX)))
}, DataType::Float32);
}
else if (paddingString == "VALID")
{
padding = false;
outputInfo = TensorInfo({ inputTensorInfo.GetShape()[0],
inputTensorInfo.GetShape()[1],
static_cast<uint32_t>(ceil(
static_cast<float>(inputHeight - pooling2dDescriptor.m_PoolHeight + 1) /
static_cast<float>(pooling2dDescriptor.m_StrideY))),
static_cast<uint32_t>(ceil(
static_cast<float>(inputWidth - pooling2dDescriptor.m_PoolWidth + 1) /
static_cast<float>(pooling2dDescriptor.m_StrideX)))
}, DataType::Float32);
}
else
{
throw ParseException("Only 'SAME' and 'VALID' padding supported");
}
CalcPadding(inputWidth, pooling2dDescriptor.m_PoolWidth, pooling2dDescriptor.m_StrideX,
pooling2dDescriptor.m_PadLeft, pooling2dDescriptor.m_PadRight, padding);
CalcPadding(inputHeight, pooling2dDescriptor.m_PoolHeight, pooling2dDescriptor.m_StrideY,
pooling2dDescriptor.m_PadTop, pooling2dDescriptor.m_PadBottom, padding);
IConnectableLayer* layer = m_Network->AddPooling2dLayer(pooling2dDescriptor, nodeDef.name().c_str());
if (layer == nullptr)
{
throw ParseException("Failed to add pooling2d layer");
}
layer->GetOutputSlot(0).SetTensorInfo(outputInfo);
if (dataFormat == "NHWC")
{
layer = SwizzleInDeswizzleOut(*m_Network, inputSlot, *layer, nodeDef.name());
}
else
{
inputSlot.Connect(layer->GetInputSlot(0));
}
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
ParsedTfOperationPtr TfParser::AddAdditionLayer(const tensorflow::NodeDef& nodeDef, bool isBiasAdd)
{
std::vector<OutputOfParsedTfOperation> inputs = GetInputParsedTfOperationsChecked(nodeDef, 2);
IOutputSlot* input0Slot = &inputs[0].m_IndexedValue->ResolveArmnnOutputSlot(inputs[0].m_Index);
IOutputSlot* input1Slot = &inputs[1].m_IndexedValue->ResolveArmnnOutputSlot(inputs[1].m_Index);
const TensorInfo& input0Info = input0Slot->GetTensorInfo();
const TensorInfo& input1Info = input1Slot->GetTensorInfo();
if (isBiasAdd)
{
// BiasAdd takes bias as a 1D tensor. We need to add a reshape layer to create a 4D tensor
// with the same data in the correct dimension for broadcast in addition.
if(input1Info.GetNumDimensions() != 1)
{
throw ParseException("Unsupported bias for BiasAdd. It should be a 1D vector.");
}
const std::string dataFormat = ReadMandatoryNodeStringAttribute(nodeDef, "data_format");
const bool isNHWC = (dataFormat == "NHWC");
const bool isNCHW = (dataFormat == "NCHW");
if (!isNHWC && ! isNCHW)
{
throw ParseException("Only NHWC or NCHW supported for BiasAdd");
}
input1Slot = BroadcastForAddandMul(input0Slot, input1Slot, isNHWC, *m_Network, nodeDef);
}
else
{
if (input0Info.GetNumDimensions() == 1)
{
const bool isNHWC = true;
input0Slot = BroadcastForAddandMul(input1Slot, input0Slot, isNHWC, *m_Network, nodeDef);
}
if (input1Info.GetNumDimensions() == 1)
{
const bool isNHWC = true;
input1Slot = BroadcastForAddandMul(input0Slot, input1Slot, isNHWC, *m_Network, nodeDef);
}
}
IConnectableLayer* const layer = m_Network->AddAdditionLayer(nodeDef.name().c_str());
input0Slot->Connect(layer->GetInputSlot(0));
input1Slot->Connect(layer->GetInputSlot(1));
if (input0Info.GetNumDimensions() == 1 && isBiasAdd == false)
{
layer->GetOutputSlot(0).SetTensorInfo(input1Slot->GetTensorInfo());
}
else
{
layer->GetOutputSlot(0).SetTensorInfo(input0Slot->GetTensorInfo());
}
return std::make_unique<SingleLayerParsedTfOperation>(this, nodeDef, layer);
}
IConnectableLayer* TfParser::AddFullyConnectedLayer(const tensorflow::NodeDef& matMulNodeDef,
const tensorflow::NodeDef* addNodeDef, const char* armnnLayerName)
{
// find bias const (if applicable)
ParsedConstTfOperation<float>* biasNode = nullptr;
if (addNodeDef != nullptr)
{
std::vector<OutputOfParsedTfOperation> addInputs = GetInputParsedTfOperationsChecked(*addNodeDef, 2);
// find our inputs
if (HasParsedConstTensor<float>(addInputs[0].m_IndexedValue->GetNode().name()))
{
biasNode = boost::polymorphic_downcast<ParsedConstTfOperation<float>*>(addInputs[0].m_IndexedValue);
}
else if (HasParsedConstTensor<float>(addInputs[1].m_IndexedValue->GetNode().name()))
{
biasNode = boost::polymorphic_downcast<ParsedConstTfOperation<float>*>(addInputs[1].m_IndexedValue);
}
else
{
throw ParseException("ArmNN only supports fully connected layers with constant bias");
}
}
// find matmul inputs
ParsedConstTfOperation<float>* weightNode = nullptr;
ParsedTfOperation* inputNode = nullptr;
unsigned int inputIdx = 0;
std::vector<OutputOfParsedTfOperation> mulInputs = GetInputParsedTfOperationsChecked(matMulNodeDef, 2);
if (HasParsedConstTensor<float>(mulInputs[0].m_IndexedValue->GetNode().name()))
{
weightNode = boost::polymorphic_downcast<ParsedConstTfOperation<float>*>(mulInputs[0].m_IndexedValue);
inputNode = mulInputs[1].m_IndexedValue;
inputIdx = mulInputs[1].m_Index;
}
else if (HasParsedConstTensor<float>(mulInputs[1].m_IndexedValue->GetNode().name()))
{
weightNode = boost::polymorphic_downcast<ParsedConstTfOperation<float>*>(mulInputs[1].m_IndexedValue);
inputNode = mulInputs[0].m_IndexedValue;
inputIdx = mulInputs[0].m_Index;
}
else
{
throw ParseException("ArmNN only supports fully connected layers with constant weights");
}
std::vector<float> weightTensorData;
// handle weight
ConstTensor weights = weightNode->GetConstTensor(false, weightTensorData);
FullyConnectedDescriptor desc;
desc.m_BiasEnabled = addNodeDef != nullptr;
IConnectableLayer* layer = nullptr;
// make the layer
if (addNodeDef != nullptr)
{
std::vector<float> biasTensorData;
ConstTensor biases = biasNode->GetConstTensor(false, biasTensorData);
if (weights.GetShape()[1] != biases.GetShape()[0])
{
throw ParseException("shape of matmul and bias do not match");
}
layer = m_Network->AddFullyConnectedLayer(desc, weights, biases, armnnLayerName);
}
else
{
layer = m_Network->AddFullyConnectedLayer(desc, weights, armnnLayerName);
}
BOOST_ASSERT(layer != nullptr);
inputNode->ResolveArmnnOutputSlot(inputIdx).Connect(layer->GetInputSlot(0));
unsigned int batches = inputNode->ResolveArmnnOutputSlot(inputIdx).GetTensorInfo().GetShape()[0];
// handle output
TensorInfo outputInfo({ batches, weights.GetShape()[1] }, DataType::Float32);
layer->GetOutputSlot(0).SetTensorInfo(outputInfo);
return layer;
}
void TfParser::LoadNodeDef(const tensorflow::NodeDef& nodeDef, const tensorflow::GraphDef& graphDef)
{
// get the type of the node (assume float)
tensorflow::DataType type = tensorflow::DT_FLOAT;
if (nodeDef.attr().count("T") != 0)
{
auto attr = nodeDef.attr().at("T");
type = attr.type();
}
else if (nodeDef.attr().count("dtype") != 0)
{
auto attr = nodeDef.attr().at("dtype");
type = attr.type();
}
if (type != tensorflow::DT_FLOAT && nodeDef.op() != "Const")
{
throw ParseException("Currently only FLOAT is supported for tensorflow nodes (apart from Const)");
}
const std::string& operation = nodeDef.op();
auto it = ms_OperationNameToParsingFunctions.find(operation);
if (it != ms_OperationNameToParsingFunctions.end())
{
auto func = it->second;
ParsedTfOperationPtr parsedTfOperation = (this->*func)(nodeDef, graphDef);
ParsedTfOperation* parsedTfOperationRaw = parsedTfOperation.get();
// Store the parsed operation so that dependent layers can connect to it
auto it = m_ParsedTfOperations.find(nodeDef.name());
if (it != m_ParsedTfOperations.end())
{
throw ParseException(boost::str(boost::format("Name %1% used by more than one node") % nodeDef.name()));
}
m_ParsedTfOperations[nodeDef.name()] = std::move(parsedTfOperation);
// If this node was requested as an output from the network then add an ArmNN output layer
if (std::find(m_RequestedOutputs.begin(), m_RequestedOutputs.end(), nodeDef.name()) !=
m_RequestedOutputs.end())
{
auto outId = ParseOutputId(nodeDef.name());
const LayerBindingId layerId = boost::numeric_cast<LayerBindingId>(m_NetworkOutputsBindingInfo.size());
IOutputSlot& prevSlot = parsedTfOperationRaw->ResolveArmnnOutputSlot(outId.m_Index);
TensorInfo tensorInfo = prevSlot.GetTensorInfo();
IConnectableLayer* outputLayer = m_Network->AddOutputLayer(layerId, nodeDef.name().c_str());
prevSlot.Connect(outputLayer->GetInputSlot(0));
TrackOutputBinding(outputLayer, layerId, tensorInfo);
}
}
else
{
throw ParseException(boost::str(
boost::format("Unsupported operation %1% in tensorflow::GraphDef") % operation));
}
}
void TfParser::LoadGraphDef(const tensorflow::GraphDef& graphDef)
{
// add all nodes to our map
m_NodesByName.clear();
m_NetworkInputsBindingInfo.clear();
m_NetworkOutputsBindingInfo.clear();
for (int i = 0; i < graphDef.node_size(); ++i)
{
const tensorflow::NodeDef& node = graphDef.node(i);
m_NodesByName[node.name()] = &node;
}
// Find the output nodes the user requested
std::vector<const tensorflow::NodeDef*> targetNodes;
for (const std::string& requestedOutputName : m_RequestedOutputs)
{
auto nodeIt = m_NodesByName.find(requestedOutputName);
if (nodeIt == m_NodesByName.end())
{
throw ParseException("Couldn't find requested output node '" + requestedOutputName + "' in graph");
}
targetNodes.push_back(nodeIt->second);
}
// Sort them into a linear ordering such that all inputs of a node are before the node itself
std::vector<const tensorflow::NodeDef*> sortedNodes;
if (!armnnUtils::GraphTopologicalSort<const tensorflow::NodeDef*>(
targetNodes,
[this](const tensorflow::NodeDef* node)
{
auto outputs = GetTfInputNodes(*node);
std::vector<const tensorflow::NodeDef*> nodesOnly;
for (const auto & o : outputs) {
nodesOnly.push_back(o.m_IndexedValue);
}
return nodesOnly;
},
sortedNodes))
{
throw ParseException("Cycle detected in graph");
}
// Parse each node in order, knowing that all inputs of a node will be processed before the node itself
for (const auto& it : sortedNodes)
{
const tensorflow::NodeDef& currentNode = *it;
LoadNodeDef(currentNode, graphDef);
}
}
INetworkPtr TfParser::CreateNetworkFromTextFile(const char* graphFile,
const std::map<std::string, TensorShape>& inputShapes,
const std::vector<std::string>& requestedOutputs)
{
FILE* fd = fopen(graphFile, "r");
if (fd == nullptr)
{
std::stringstream error;
error << "Graph file " << graphFile << " failed to open";
throw FileNotFoundException(error.str());
}
// Parse the file into a message
tensorflow::GraphDef graphDef;
auto input = new google::protobuf::io::FileInputStream(fileno(fd));
bool success = google::protobuf::TextFormat::Parse(input, &graphDef);
delete input;
fclose(fd);
if (!success)
{
std::stringstream error;
error << "Failed to parse graph file";
throw ParseException(error.str());
}
return CreateNetworkFromGraphDef(graphDef, inputShapes, requestedOutputs);
}
INetworkPtr TfParser::CreateNetworkFromString(const char* protoText,
const std::map<std::string, TensorShape>& inputShapes,
const std::vector<std::string>& requestedOutputs)
{
// Parse the string into a message
tensorflow::GraphDef graphDef;
bool success = google::protobuf::TextFormat::ParseFromString(protoText, &graphDef);
if (!success)
{
std::stringstream error;
error << "Failed to parse graph file";
throw ParseException(error.str());
}
return CreateNetworkFromGraphDef(graphDef, inputShapes, requestedOutputs);
}
INetworkPtr TfParser::CreateNetworkFromBinaryFile(const char* graphFile,
const std::map<std::string, TensorShape>& inputShapes,
const std::vector<std::string>& requestedOutputs)
{
FILE* fd = fopen(graphFile, "rb");
if (fd == nullptr)
{
std::stringstream error;
error << "Graph file " << graphFile << " failed to open";
throw FileNotFoundException(error.str());
}
// Parse the file into a message
tensorflow::GraphDef graphDef;
google::protobuf::io::FileInputStream inStream(fileno(fd));
google::protobuf::io::CodedInputStream codedStream(&inStream);
codedStream.SetTotalBytesLimit(INT_MAX, INT_MAX);
bool success = graphDef.ParseFromCodedStream(&codedStream);
fclose(fd);
if (!success)
{
std::stringstream error;
error << "Failed to parse protobuf file" << graphFile;
throw ParseException(error.str());
}
return CreateNetworkFromGraphDef(graphDef, inputShapes, requestedOutputs);
}
INetworkPtr TfParser::CreateNetworkFromGraphDef(const tensorflow::GraphDef& graphDef,
const std::map<std::string, TensorShape>& inputShapes,
const std::vector<std::string>& requestedOutputs)
{
m_Network = INetwork::Create();
m_InputShapes = inputShapes;
if (requestedOutputs.size() == 0)
{
throw ParseException("requestedOutputs must have at least one entry");
}
m_RequestedOutputs = requestedOutputs;
try
{
LoadGraphDef(graphDef);
}
catch (const ParseException& e)
{
Cleanup();
throw e;
}
Cleanup();
return std::move(m_Network);
}
void TfParser::Cleanup()
{
// cleanup, in case we reuse this parser
m_InputShapes.clear();
m_RequestedOutputs.clear();
m_NodesByName.clear();
m_ParsedTfOperations.clear();
}
BindingPointInfo TfParser::GetNetworkInputBindingInfo(const std::string& name) const
{
return GetBindingInfo(name, "input", m_NetworkInputsBindingInfo);
}
BindingPointInfo TfParser::GetNetworkOutputBindingInfo(const std::string& name) const
{
return GetBindingInfo(name, "output", m_NetworkOutputsBindingInfo);
}
std::pair<LayerBindingId, TensorInfo> TfParser::GetBindingInfo(const std::string& layerName,
const char* bindingPointDesc,
const std::unordered_map<std::string, BindingPointInfo>& nameToBindingInfo)
{
auto it = nameToBindingInfo.find(layerName);
if (it == nameToBindingInfo.end())
{
throw InvalidArgumentException(boost::str(boost::format("Unknown %1% '%2%'") % bindingPointDesc % layerName));
}
return it->second;
}
void TfParser::TrackInputBinding(IConnectableLayer* layer, LayerBindingId id, const TensorInfo& tensorInfo)
{
return TrackBindingPoint(layer, id, tensorInfo, "input", m_NetworkInputsBindingInfo);
}
void TfParser::TrackOutputBinding(IConnectableLayer* layer, LayerBindingId id, const TensorInfo& tensorInfo)
{
return TrackBindingPoint(layer, id, tensorInfo, "output", m_NetworkOutputsBindingInfo);
}
void TfParser::TrackBindingPoint(IConnectableLayer* layer,
LayerBindingId id,
const TensorInfo& tensorInfo,
const char* bindingPointDesc,
std::unordered_map<std::string, BindingPointInfo>& nameToBindingInfo)
{
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") % id % bindingPointDesc));
}
}
} // namespace armnnTfParser