#include <HGCalConcentratorAutoEncoderImpl.h>

Public Member Functions
	HGCalConcentratorAutoEncoderImpl (const edm::ParameterSet &conf)

void	select (unsigned nLinks, const std::vector< l1t::HGCalTriggerCell > &trigCellVecInput, std::vector< l1t::HGCalTriggerCell > &trigCellVecOutput, std::vector< l1t::HGCalConcentratorData > &ae_EncodedOutput)

void	setGeometry (const HGCalTriggerGeometryBase *const geom)

Private Attributes
std::array< double, nEncodedLayerNodes_ >	ae_encodedLayer_

bool	bitShiftNormalization_

int	bitsPerInput_

std::vector< int >	cellRemap_

std::vector< int >	cellRemapNoDuplicates_

std::vector< uint >	decoderShape_

std::vector< uint >	encoderShape_

std::unique_ptr< tensorflow::GraphDef >	graphDef_decoder_

std::unique_ptr< tensorflow::GraphDef >	graphDef_encoder_

std::string	inputTensorName_decoder_

std::string	inputTensorName_encoder_

std::vector< unsigned int >	linkToGraphMap_

int	maxBitsPerOutput_

std::vector< edm::ParameterSet >	modelFilePaths_

unsigned int	nInputs_

std::vector< int >	outputBitsPerLink_

std::string	outputTensorName_decoder_

std::string	outputTensorName_encoder_

bool	preserveModuleSum_

bool	saveEncodedValues_

std::vector< std::unique_ptr< tensorflow::Session > >	session_decoder_

std::vector< std::unique_ptr< tensorflow::Session > >	session_encoder_

HGCalTriggerTools	triggerTools_

double	zeroSuppresionThreshold_

Static Private Attributes
static constexpr int	ae_outputCellU_ [nTriggerCells_]

static constexpr int	ae_outputCellV_ [nTriggerCells_]

static constexpr int	cellUVremap_ [cellUVSize_][cellUVSize_]

static constexpr int	cellUVSize_ = 8

static constexpr int	decoderTensorDims_ = 2

static constexpr int	encoderTensorDims_ = 4

static constexpr int	maxAEInputSize_ = 192

static constexpr unsigned int	maxNumberOfLinks_ = 13

static constexpr int	nEncodedLayerNodes_ = 16

static constexpr int	nTriggerCells_ = 48

Detailed Description

Definition at line 14 of file HGCalConcentratorAutoEncoderImpl.h.

Constructor & Destructor Documentation

◆ HGCalConcentratorAutoEncoderImpl()

HGCalConcentratorAutoEncoderImpl::HGCalConcentratorAutoEncoderImpl ( const edm::ParameterSet & conf )

Definition at line 8 of file HGCalConcentratorAutoEncoderImpl.cc.

References cellRemap_, cellRemapNoDuplicates_, tensorflow::createSession(), decoderShape_, decoderTensorDims_, encoderShape_, encoderTensorDims_, Exception, contentValuesFiles::fullPath, graphDef_decoder_, graphDef_encoder_, mps_fire::i, inputTensorName_decoder_, inputTensorName_encoder_, linkToGraphMap_, tensorflow::loadGraphDef(), maxNumberOfLinks_, modelFilePaths_, Skims_PA_cff::name, nInputs_, nTriggerCells_, outputTensorName_decoder_, outputTensorName_encoder_, session_decoder_, session_encoder_, and AlCaHLTBitMon_QueryRunRegistry::string.

     : cellRemap_(conf.getParameter<std::vector<int>>("cellRemap")),
       cellRemapNoDuplicates_(conf.getParameter<std::vector<int>>("cellRemapNoDuplicates")),
       encoderShape_(conf.getParameter<std::vector<uint>>("encoderShape")),
       decoderShape_(conf.getParameter<std::vector<uint>>("decoderShape")),
       bitsPerInput_(conf.getParameter<int>("nBitsPerInput")),
       maxBitsPerOutput_(conf.getParameter<int>("maxBitsPerOutput")),
       outputBitsPerLink_(conf.getParameter<std::vector<int>>("bitsPerLink")),
       modelFilePaths_(conf.getParameter<std::vector<edm::ParameterSet>>("modelFiles")),
       linkToGraphMap_(conf.getParameter<std::vector<unsigned int>>("linkToGraphMap")),
       zeroSuppresionThreshold_(conf.getParameter<double>("zeroSuppresionThreshold")),
       bitShiftNormalization_(conf.getParameter<bool>("bitShiftNormalization")),
       saveEncodedValues_(conf.getParameter<bool>("saveEncodedValues")),
       preserveModuleSum_(conf.getParameter<bool>("preserveModuleSum")) {
   // find total size of the expected input shape
   // used for checking the maximum size used in cell Remap
   nInputs_ = 1;
   for (const auto& i : encoderShape_) {
     nInputs_ *= i;
   }
 
   // check the size of the inputs shapes
   if (encoderShape_.size() != encoderTensorDims_) {
     throw cms::Exception("BadInitialization")
         << "Encoder input shapes are currently expected to be " << encoderTensorDims_ << " values";
   }
 
   if (decoderShape_.size() != decoderTensorDims_) {
     throw cms::Exception("BadInitialization")
         << "Encoder input shapes are currently expected to be " << decoderTensorDims_ << " values long";
   }
 
   if (cellRemap_.size() != nInputs_) {
     throw cms::Exception("BadInitialization")
         << "Size of cellRemap (" << cellRemap_.size()
         << ") does not agree with the total size specified for the encoder inputs based on the encoderShape variable ("
         << nInputs_ << ")";
   }
 
   if (cellRemap_.size() != cellRemapNoDuplicates_.size()) {
     throw cms::Exception("BadInitialization")
         << "Size of cellRemap (" << cellRemap_.size() << ") does not agree with size of cellRemapNoDuplicates ("
         << cellRemapNoDuplicates_.size() << ")";
   }
 
   for (unsigned i = 0; i < cellRemap_.size(); i++) {
     if (cellRemap_[i] > nTriggerCells_ - 1) {
       throw cms::Exception("BadInitialization")
           << "cellRemap value " << cellRemap_[i] << " is larger than the number of trigger cells " << nTriggerCells_;
     }
     if (cellRemapNoDuplicates_[i] > nTriggerCells_ - 1) {
       throw cms::Exception("BadInitialization") << "cellRemapNoDuplicates value " << cellRemapNoDuplicates_[i]
                                                 << " is larger than the number of trigger cells " << nTriggerCells_;
     }
   }
 
   for (const auto& modelFilePset : modelFilePaths_) {
     std::string encoderPath = modelFilePset.getParameter<edm::FileInPath>("encoderModelFile").fullPath();
     std::string decoderPath = modelFilePset.getParameter<edm::FileInPath>("decoderModelFile").fullPath();
 
     graphDef_encoder_ = std::unique_ptr<tensorflow::GraphDef>{tensorflow::loadGraphDef(encoderPath)};
 
     // create a new session and add the graphDef
     session_encoder_.push_back(
         std::unique_ptr<tensorflow::Session>{tensorflow::createSession(graphDef_encoder_.get())});
 
     graphDef_decoder_ = std::unique_ptr<tensorflow::GraphDef>{tensorflow::loadGraphDef(decoderPath)};
 
     // create a new session and add the graphDef
     session_decoder_.push_back(
         std::unique_ptr<tensorflow::Session>{tensorflow::createSession(graphDef_decoder_.get())});
 
     //extract encoder tenser names from first graph, check that rest of the names are consistent
     if (modelFilePset == modelFilePaths_.front()) {
       inputTensorName_encoder_ = graphDef_encoder_.get()->node(0).name();
       outputTensorName_encoder_ = graphDef_encoder_.get()->node(graphDef_encoder_.get()->node_size() - 1).name();
       inputTensorName_decoder_ = graphDef_decoder_.get()->node(0).name();
       outputTensorName_decoder_ = graphDef_decoder_.get()->node(graphDef_decoder_.get()->node_size() - 1).name();
     } else {
       if (inputTensorName_encoder_ != graphDef_encoder_.get()->node(0).name()) {
         throw cms::Exception("BadInitialization") << "provided list of encoder graphs have different input nodes";
       }
       if (outputTensorName_encoder_ != graphDef_encoder_.get()->node(graphDef_encoder_.get()->node_size() - 1).name()) {
         throw cms::Exception("BadInitialization") << "provided list of encoder graphs have different output nodes";
       }
       if (inputTensorName_decoder_ != graphDef_decoder_.get()->node(0).name()) {
         throw cms::Exception("BadInitialization") << "provided list of decoder graphs have different input nodes";
       }
       if (outputTensorName_decoder_ != graphDef_decoder_.get()->node(graphDef_decoder_.get()->node_size() - 1).name()) {
         throw cms::Exception("BadInitialization") << "provided list of decoder graphs have different output nodes";
       }
     }
   }
 
   // check that the appropriate number of links have been specified
   if (linkToGraphMap_.size() <= maxNumberOfLinks_) {
     throw cms::Exception("BadInitialization")
         << "Autoencoder graph number must be specified for all link allocation possibilities. Only "
         << linkToGraphMap_.size() << " values specified while " << maxNumberOfLinks_ << "links are possible";
   }
 
   // check that all graph indices specified exist in the model file lists
   for (const auto& graphNumber : linkToGraphMap_) {
     if (graphNumber >= modelFilePaths_.size()) {
       throw cms::Exception("BadInitialization")
           << "Autoencoder graph number  " << graphNumber << " is larger than the size of the provided list of graphs "
           << modelFilePaths_.size();
     }
   }
 }

Member Function Documentation

◆ select()

void HGCalConcentratorAutoEncoderImpl::select	(	unsigned	nLinks,
		const std::vector< l1t::HGCalTriggerCell > &	trigCellVecInput,
		std::vector< l1t::HGCalTriggerCell > &	trigCellVecOutput,
		std::vector< l1t::HGCalConcentratorData > &	ae_EncodedOutput
	)

Definition at line 119 of file HGCalConcentratorAutoEncoderImpl.cc.

References gpuClustering::adc, ae_encodedLayer_, ae_outputCellU_, ae_outputCellV_, bitShiftNormalization_, bitsPerInput_, cellRemap_, cellRemapNoDuplicates_, cellUVremap_, ztail::d, data, decoderShape_, encoderShape_, l1tnanotables_cff::et, PVValHelper::eta, Exception, HGCalTriggerTools::getTCPosition(), HGCalTriggerTools::getTriggerGeometry(), mps_fire::i, l1ctLayer2EG_cff::id, inputTensorName_decoder_, inputTensorName_encoder_, createfilelist::int, HGCalTriggerTools::isScintillator(), nano_mu_digi_cff::layer, linkToGraphMap_, maxBitsPerOutput_, nEncodedLayerNodes_, nInputs_, PostProcessor_cff::normalization, nTriggerCells_, outputBitsPerLink_, outputTensorName_decoder_, outputTensorName_encoder_, point, funct::pow(), preserveModuleSum_, tensorflow::run(), saveEncodedValues_, session_decoder_, session_encoder_, l1t::HGCalTriggerCell::setCompressedCharge(), l1t::HGCalTriggerCell::setMipPt(), reco::LeafCandidate::setP4(), l1t::HGCalTriggerCell::setPosition(), l1t::HGCalTriggerCell::setUncompressedCharge(), triggerTools_, parallelization::uint, HGCalTriggerGeometryBase::validTriggerCell(), HGCalWaferIndex::waferU(), HGCalWaferIndex::waferV(), and zeroSuppresionThreshold_.

                                                                                                            {
   std::array<double, nTriggerCells_> mipPt;
   std::array<double, nTriggerCells_> uncompressedCharge;
   std::array<double, nTriggerCells_> compressedCharge;
   std::array<double, maxAEInputSize_> ae_inputArray;
   std::array<double, nTriggerCells_> ae_outputArray;
 
   //reset inputs to 0 to account for zero suppressed trigger cells
   mipPt.fill(0);
   uncompressedCharge.fill(0);
   compressedCharge.fill(0);
   ae_inputArray.fill(0);
   ae_outputArray.fill(0);
 
   double modSum = 0;
 
   int bitsPerOutput = outputBitsPerLink_.at(nLinks);
 
   // largest expected input and output values, used for bit truncation
   // values of -1 for the number of bits used to keep full precision, in which case the MaxIntSize variables are not used
   double inputMaxIntSize = 1;
   if (bitsPerInput_ > 0)
     inputMaxIntSize = 1 << bitsPerInput_;
   double outputMaxIntSize = 1;
   if (bitsPerOutput > 0)
     outputMaxIntSize = 1 << bitsPerOutput;
   double outputMaxIntSizeGlobal = 1;
   if (maxBitsPerOutput_ > 0)
     outputMaxIntSizeGlobal = 1 << maxBitsPerOutput_;
 
   for (const auto& trigCell : trigCellVecInput) {
     if (triggerTools_.isScintillator(trigCell.detId()))
       return;  //currently, only silicon modules are setup to work (mapping of scinillators would be different, and needs to be investigated)
 
     HGCalTriggerDetId id(trigCell.detId());
     uint cellu = id.triggerCellU();
     uint cellv = id.triggerCellV();
     int inputIndex = cellUVremap_[cellu][cellv];
     if (inputIndex < 0) {
       throw cms::Exception("BadInitialization")
           << "Invalid index provided for trigger cell u=" << cellu << " v=" << cellv << " in cellUVRemap[" << cellu
           << "][" << cellv << "]";
     }
 
     mipPt[inputIndex] = trigCell.mipPt();
     uncompressedCharge[inputIndex] = trigCell.uncompressedCharge();
     compressedCharge[inputIndex] = trigCell.compressedCharge();
 
     modSum += trigCell.mipPt();
   }
 
   double normalization = modSum;
   if (modSum > 0) {
     //Use a bit shift normalization like will be implemented in ECON, rather than floating point sum
     //Normalizes to the MSB value of the module sum
     if (bitShiftNormalization_) {
       int msb = int(log2(modSum));
       normalization = pow(2, msb);
     }
 
     //normalize inputs to module sum
     for (uint i = 0; i < nInputs_; i++) {
       int remapIndex = cellRemap_[i];
       if (remapIndex < 0)
         continue;
       ae_inputArray[i] = mipPt[remapIndex] / normalization;
       //round to precision of input, if bitsPerInput_ is -1 keep full precision
       if (bitsPerInput_ > 0) {
         ae_inputArray[i] = std::round(ae_inputArray[i] * inputMaxIntSize) / inputMaxIntSize;
       }
     }
   }
 
   tensorflow::Tensor encoder_input(tensorflow::DT_FLOAT,
                                    {encoderShape_[0], encoderShape_[1], encoderShape_[2], encoderShape_[3]});
 
   float* d = encoder_input.flat<float>().data();
   for (uint i = 0; i < nInputs_; i++, d++) {
     *d = ae_inputArray[i];
   }
 
   int graphIndex = linkToGraphMap_.at(nLinks);
 
   std::vector<tensorflow::Tensor> encoder_outputs;
   tensorflow::run(session_encoder_.at(graphIndex).get(),
                   {{inputTensorName_encoder_, encoder_input}},
                   {outputTensorName_encoder_},
                   &encoder_outputs);
 
   if (encoder_outputs.empty()) {
     throw cms::Exception("BadInitialization") << "Autoencoder graph returning empty output vector";
   }
 
   d = encoder_outputs[0].flat<float>().data();
   for (int i = 0; i < encoder_outputs[0].NumElements(); i++, d++) {
     ae_encodedLayer_[i] = *d;
     //truncate the encoded layer bits
     if (bitsPerOutput > 0 && maxBitsPerOutput_ > 0) {
       ae_encodedLayer_[i] = std::round(ae_encodedLayer_[i] * outputMaxIntSize) / outputMaxIntSize;
     }
   }
 
   tensorflow::Tensor decoder_input(tensorflow::DT_FLOAT, {decoderShape_[0], decoderShape_[1]});
   d = decoder_input.flat<float>().data();
   for (int i = 0; i < nEncodedLayerNodes_; i++, d++) {
     *d = ae_encodedLayer_[i];
   }
 
   std::vector<tensorflow::Tensor> decoder_outputs;
   tensorflow::run(session_decoder_.at(graphIndex).get(),
                   {{inputTensorName_decoder_, decoder_input}},
                   {outputTensorName_decoder_},
                   &decoder_outputs);
 
   double outputSum = 0.;
 
   d = decoder_outputs[0].flat<float>().data();
   for (uint i = 0; i < nInputs_; i++, d++) {
     int remapIndex = cellRemapNoDuplicates_[i];
     if (remapIndex < 0)
       continue;
     outputSum += *d * normalization;
     ae_outputArray[remapIndex] = *d;
   }
 
   double renormalizationFactor = 1.;
   if (preserveModuleSum_) {
     renormalizationFactor = modSum / outputSum;
   }
 
   // Add data back into trigger cells
   if (modSum > 0) {
     //get detID for everything but cell, take first entry detID and subtract off cellU and cellV contribution
     HGCalTriggerDetId id(trigCellVecInput.at(0).detId());
     int subdet = id.subdet();
     int zp = id.zside();
     int type = id.type();
     int layer = id.layer();
     int waferU = id.waferU();
     int waferV = id.waferV();
 
     //use first TC to find mipPt conversions to Et and ADC
     float mipPtToEt_conv = trigCellVecInput[0].et() / trigCellVecInput[0].mipPt();
     float mipToADC_conv = trigCellVecInput[0].hwPt() / (trigCellVecInput[0].mipPt() * cosh(trigCellVecInput[0].eta()));
 
     for (int i = 0; i < nTriggerCells_; i++) {
       if (ae_outputArray[i] > 0) {
         int cellU = ae_outputCellU_[i];
         int cellV = ae_outputCellV_[i];
 
         HGCalTriggerDetId id(subdet, zp, type, layer, waferU, waferV, cellU, cellV);
 
         if (!triggerTools_.getTriggerGeometry()->validTriggerCell(id))
           continue;
 
         GlobalPoint point = triggerTools_.getTCPosition(id);
 
         double mipPt = ae_outputArray[i] * normalization * renormalizationFactor;
         double adc = mipPt * cosh(point.eta()) * mipToADC_conv;
         double et = mipPt * mipPtToEt_conv;
 
         if (mipPt < zeroSuppresionThreshold_)
           continue;
 
         l1t::HGCalTriggerCell triggerCell(reco::LeafCandidate::LorentzVector(), adc, 0, 0, 0, id);
         //Keep the pre-autoencoder charge for this cell
         triggerCell.setUncompressedCharge(uncompressedCharge[i]);
         triggerCell.setCompressedCharge(compressedCharge[i]);
         triggerCell.setMipPt(mipPt);
 
         math::PtEtaPhiMLorentzVector p4(et, point.eta(), point.phi(), 0.);
 
         triggerCell.setP4(p4);
         triggerCell.setPosition(point);
 
         trigCellVecOutput.push_back(triggerCell);
       }
     }
 
     if (saveEncodedValues_) {
       id = HGCalTriggerDetId(subdet, zp, type, layer, waferU, waferV, 0, 0);
       for (int i = 0; i < nEncodedLayerNodes_; i++) {
         l1t::HGCalConcentratorData encodedLayerData(ae_encodedLayer_[i] * outputMaxIntSizeGlobal, i, id);
         ae_encodedLayer_Output.push_back(encodedLayerData);
       }
     }
   }
 }

◆ setGeometry()

void HGCalConcentratorAutoEncoderImpl::setGeometry ( const HGCalTriggerGeometryBase *const geom )

inline

Definition at line 23 of file HGCalConcentratorAutoEncoderImpl.h.

References relativeConstraints::geom, HGCalTriggerTools::setGeometry(), and triggerTools_.

23 { triggerTools_.setGeometry(geom); }

HGCalTriggerTools::setGeometry

void setGeometry(const HGCalTriggerGeometryBase *const)

Definition: HGCalTriggerTools.cc:40

HGCalConcentratorAutoEncoderImpl::triggerTools_

HGCalTriggerTools triggerTools_

Definition: HGCalConcentratorAutoEncoderImpl.h:85

relativeConstraints.geom

geom

Definition: relativeConstraints.py:72

Member Data Documentation

◆ ae_encodedLayer_

std::array<double, nEncodedLayerNodes_> HGCalConcentratorAutoEncoderImpl::ae_encodedLayer_

private

Definition at line 83 of file HGCalConcentratorAutoEncoderImpl.h.

Referenced by select().

◆ ae_outputCellU_

constexpr int HGCalConcentratorAutoEncoderImpl::ae_outputCellU_[nTriggerCells_]

staticprivate

Initial value:

= {7, 6, 5, 4, 7, 6, 5, 4, 7, 6, 5, 4, 7, 6, 5, 4,
                                                          1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7,
                                                          3, 3, 3, 3, 2, 2, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0}

Definition at line 48 of file HGCalConcentratorAutoEncoderImpl.h.

Referenced by select().

◆ ae_outputCellV_

constexpr int HGCalConcentratorAutoEncoderImpl::ae_outputCellV_[nTriggerCells_]

staticprivate

Initial value:

= {4, 4, 4, 4, 5, 5, 5, 5, 6, 6, 6, 6, 7, 7, 7, 7,
                                                          0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3,
                                                          6, 5, 4, 3, 5, 4, 3, 2, 4, 3, 2, 1, 3, 2, 1, 0}

Definition at line 51 of file HGCalConcentratorAutoEncoderImpl.h.

Referenced by select().

◆ bitShiftNormalization_

bool HGCalConcentratorAutoEncoderImpl::bitShiftNormalization_

private

Definition at line 79 of file HGCalConcentratorAutoEncoderImpl.h.

Referenced by select().

◆ bitsPerInput_

int HGCalConcentratorAutoEncoderImpl::bitsPerInput_

private

Definition at line 60 of file HGCalConcentratorAutoEncoderImpl.h.

Referenced by select().

◆ cellRemap_

std::vector<int> HGCalConcentratorAutoEncoderImpl::cellRemap_

private

Definition at line 56 of file HGCalConcentratorAutoEncoderImpl.h.

Referenced by HGCalConcentratorAutoEncoderImpl(), and select().

◆ cellRemapNoDuplicates_

std::vector<int> HGCalConcentratorAutoEncoderImpl::cellRemapNoDuplicates_

private

Definition at line 57 of file HGCalConcentratorAutoEncoderImpl.h.

Referenced by HGCalConcentratorAutoEncoderImpl(), and select().

◆ cellUVremap_

constexpr int HGCalConcentratorAutoEncoderImpl::cellUVremap_[cellUVSize_][cellUVSize_]

staticprivate

Initial value:

= {{47, 46, 45, 44, -1, -1, -1, -1},
                                                                 {16, 43, 42, 41, 40, -1, -1, -1},
                                                                 {20, 17, 39, 38, 37, 36, -1, -1},
                                                                 {24, 21, 18, 35, 34, 33, 32, -1},
                                                                 {28, 25, 22, 19, 3, 7, 11, 15},
                                                                 {-1, 29, 26, 23, 2, 6, 10, 14},
                                                                 {-1, -1, 30, 27, 1, 5, 9, 13},
                                                                 {-1, -1, -1, 31, 0, 4, 8, 12}}

Definition at line 40 of file HGCalConcentratorAutoEncoderImpl.h.

Referenced by select().