nnrbm.h

00001 #ifndef __NNRBM_
00002 #define __NNRBM_
00003 
00004 #include "StreamOutput.h"
00005 #include "Framework.h"
00006 
00007 #include <math.h>
00008 #include <assert.h>
00009 #include <sstream>
00010 #include <vector>
00011 
00012 using namespace std;
00013 
00039 class NNRBM : public Framework
00040 {
00041 public:
00042     NNRBM();
00043     ~NNRBM();
00044 
00045     // input/output data definition methods
00046     void setNrTargets ( int n );
00047     void setNrInputs ( int n );
00048     void setNrExamplesTrain ( int n );
00049     void setNrExamplesProbe ( int n );
00050     void setTrainInputs ( REAL* inputs );
00051     void setTrainTargets ( REAL* targets );
00052     void setProbeInputs ( REAL* inputs );
00053     void setProbeTargets ( REAL* targets );
00054 
00055     // learn parameters
00056     void setInitWeightFactor ( REAL factor );
00057     void setLearnrate ( REAL learnrate );
00058     void setLearnrateMinimum ( REAL learnrateMin );
00059     void setLearnrateSubtractionValueAfterEverySample ( REAL learnrateDecreaseRate );
00060     void setLearnrateSubtractionValueAfterEveryEpoch ( REAL learnrateDecreaseRate );
00061     void setMomentum ( REAL momentum );
00062     void setWeightDecay ( REAL weightDecay );
00063     void setBatchSize ( int size );
00064     void setMinUpdateErrorBound ( REAL minUpdateBound );
00065     void setMaxEpochs ( int epochs );
00066     void setRPROPPosNeg ( REAL etaPos, REAL etaNeg );
00067     void setRPROPMinMaxUpdate ( REAL min, REAL max );
00068     void setL1Regularization ( bool en );
00069     void initNNWeights ( time_t seed );
00070     void enableErrorFunctionMAE ( bool en );
00071     void setRBMLearnParams ( REAL learnrateWeights, REAL learnrateBiasVis, REAL learnrateBiasHid, REAL weightDecay, int maxEpoch );
00072 
00073     // set net inner stucture
00074     void setNNStructure ( int nrLayer, int* neuronsPerLayer, bool lastLinearLayer = false, int* layerType = 0 );
00075 
00076     // all around training
00077     void printLearnrate();
00078     void setScaleOffset ( REAL scale, REAL offset );
00079     void setNormalTrainStopping ( bool en );
00080     void setGlobalEpochs ( int e );
00081     void enableRPROP ( bool en );
00082     void useBLASforTraining ( bool enable );
00083     void trainOneEpoch ( bool* updateLayer = 0 );
00084     int trainNN();
00085     REAL getRMSETrain();
00086     REAL getRMSEProbe();
00087     void predictSingleInput ( REAL* input, REAL* output );
00088     REAL* getWeightPtr();
00089     void setWeights ( REAL* w );
00090     int getNrWeights();
00091     double m_sumSquaredError;
00092     double m_sumSquaredErrorSamples;
00093 
00094     int getWeightIndex ( int layer, int neuron, int weight );
00095     int getBiasIndex ( int layer, int neuron );
00096     int getOutputIndex ( int layer, int neuron );
00097 
00098     void rbmPretraining ( REAL* input, REAL* target, int nSamples, int nTarget, int firstNLayer = -1, int crossRun = -1, int nFinetuningEpochs = -1, double finetuningLearnRate = 0.0 );
00099     vector<int> getEncoder();
00100     
00101     // tmp methods
00102     void printMiddleLayerToFile(string fname, REAL* input, REAL* target, int nSamples, int nTarget);
00103     void printAutoencoderWeightsToJavascript(string fname);
00104 
00105     NNRBM* m_nnAuto;
00106 
00107 private:
00108     // training
00109     void saveWeights();
00110     REAL calcRMSE ( REAL* inputs, REAL* targets, int examples );
00111     void forwardCalculation ( REAL* input );
00112     void forwardCalculationBLAS ( REAL* input );
00113     void backpropBLAS ( REAL* input, REAL* target, bool* updateLayer = 0 );
00114     void backprop ( REAL* input, REAL* target, bool* updateLayer = 0 );
00115 
00116     // weight init
00117     REAL getInitWeight ( int fanIn );
00118 
00119     // data description
00120     int m_nrTargets;
00121     int m_nrInputs;
00122     int m_nrExamplesTrain;
00123     int m_nrExamplesProbe;
00124     REAL* m_inputsTrain;
00125     REAL* m_inputsProbe;
00126     REAL* m_targetsTrain;
00127     REAL* m_targetsProbe;
00128 
00129     // learn params
00130     REAL m_initWeightFactor;
00131     int m_globalEpochs;
00132     REAL m_RPROP_etaPos;
00133     REAL m_RPROP_etaNeg;
00134     REAL m_RPROP_updateMin;
00135     REAL m_RPROP_updateMax;
00136     REAL m_learnRate;
00137     REAL m_learnRateMin;
00138     REAL m_learnrateDecreaseRate;
00139     REAL m_learnrateDecreaseRateEpoch;
00140     REAL m_momentum;
00141     REAL m_weightDecay;
00142     REAL m_minUpdateBound;
00143     int m_batchSize;
00144     REAL m_rbmLearnrateWeights;
00145     REAL m_rbmLearnrateBiasVis;
00146     REAL m_rbmLearnrateBiasHid;
00147     REAL m_rbmWeightDecay;
00148     int m_rbmMaxEpochs;
00149 
00150     // training
00151     REAL m_scaleOutputs;
00152     REAL m_offsetOutputs;
00153     int m_maxEpochs;
00154     bool m_useBLAS;
00155     bool m_enableRPROP;
00156     bool m_normalTrainStopping;
00157     bool m_enableL1Regularization;
00158     bool m_errorFunctionMAE;
00159 
00160     // net description, inner structure
00161     int m_nrLayer;
00162     int* m_neuronsPerLayer;
00163     int m_nrWeights;
00164     int m_nrOutputs;
00165     int* m_nrLayWeights;
00166     int* m_nrLayWeightOffsets;
00167     REAL* m_outputs;
00168     REAL* m_outputsTmp;
00169     REAL* m_derivates;
00170     REAL* m_d1;
00171     REAL* m_weights;
00172     REAL* m_weightsTmp0;
00173     REAL* m_weightsTmp1;
00174     REAL* m_weightsTmp2;
00175     REAL* m_weightsBatchUpdate;
00176     REAL* m_weightsOld;
00177     REAL* m_weightsOldOld;
00178     REAL* m_deltaW;
00179     REAL* m_deltaWOld;
00180     REAL* m_adaptiveRPROPlRate;
00181     int* m_activationFunctionPerLayer;
00182     bool m_enableAutoencoder;
00183 
00184 };
00185 
00186 #endif