casia.cpp

#include "SparseConvNet.h"
#include "NetworkArchitectures.h"
#include "SpatiallySparseDatasetCasiaOLHWDB.h"
#include "OnlineHandwritingPicture.h"

int epoch=0;
int cudaDevice=-1; //PCI bus ID, -1 for default GPU
int batchSize=100;

Picture* OnlineHandwritingPicture::distort(RNG& rng, batchType type) {
  OnlineHandwritingPicture* pic=new OnlineHandwritingPicture(*this);
  if (type==TRAINBATCH) {
    arma::mat aff=arma::eye(2,2);
    aff(0,0)+=rng.uniform(-0.3,0.3); //x stretch
    aff(1,1)+=rng.uniform(-0.3,0.3); //y stretch
    int r=rng.randint(3);
    float alpha=rng.uniform(-0.3,0.3);
    arma::mat x=arma::eye(2,2);
    if (r==0) x(0,1)=alpha;
    if (r==1) x(1,0)=alpha;
    if (r==2) {x(0,0)=cos(alpha);x(0,1)=-sin(alpha);x(1,0)=sin(alpha);x(1,1)=cos(alpha);}
    aff=aff*x;
    arma::mat y(1,2);
    y(0,0)=renderSize*rng.uniform(-0.1875,0.1875);
    y(0,1)=renderSize*rng.uniform(-0.1875,0.1875);
    for (int i=0;i<ops.size();++i)
      pic->ops[i]=pic->ops[i]*aff+arma::repmat(y,pic->ops[i].n_rows,1);
    pic->offset3d=rng.uniform(-0.1,0.1);
  }
  return pic;
}

class CNN : public SparseConvNet {
public:
  CNN (int dimension, int l, int k, ActivationFunction fn, int nInputFeatures, int nClasses, float p=0.0f, int cudaDevice=-1, int nTop=1)
    : SparseConvNet(dimension,nInputFeatures, nClasses, cudaDevice, nTop){
    for (int i=0;i<=l;i++)
      addLeNetLayerMP((i+1)*k*0+32<<i,
		      2+(i==0),
		      1,
		      (i<l)?3:1,
		      (i<l)?2:1,
		      fn,
		      p*i*1.0f/l);
    addSoftmaxLayer();
  }
};
int main() {
  std::string baseName="weights/casia";

  SpatiallySparseDataset trainSet=CasiaOLHWDB11TrainSet(40,Simple);
  SpatiallySparseDataset testSet=CasiaOLHWDB11TestSet(40,Simple);

  trainSet.summary();
  testSet.summary();
  CNN cnn(2,4,32,VLEAKYRELU,trainSet.nFeatures,trainSet.nClasses,0.0f,cudaDevice);
  cnn.calculateInputRegularizingConstants(trainSet);

  if (epoch>0)
    cnn.loadWeights(baseName,epoch);
  for (epoch++;;epoch++) {
    std::cout <<"epoch:" << epoch << ": " << std::flush;
    cnn.processDataset(trainSet, batchSize,0.003*exp(-epoch*0.1));
    cnn.saveWeights(baseName,epoch);
    if (epoch%2==0) {
      cnn.processDataset(testSet,  batchSize);
    }
  }
}