#ifndef TRAIN_IMAGE_INC_CPP

#define TRAIN_IMAGE_INC_CPP

#include <set>

#include "train.inc.cpp"

#include "layer.simple.inc.cpp"

class TrainerImage: public Trainer {

protected:

  std::vector<unsigned char> data;

  std::vector<unsigned char> tmpdata;

  std::vector<int> shuffle;

  std::vector<int> shuffle2;

  Layout pbl;

  Layout::List flist, blist;

  FILE *f;

  size_t imgsize;

  int count;

  int workCount;

public:

  int pad;

  const char *datafile;

  const char *outfile;

  Layer *dataLayer;

  TrainerImage(): f(), imgsize(), count(), workCount(), pad(), datafile(), outfile(), dataLayer() { }

protected:

  bool prepare() override {

    assert(datafile);   

    assert(fl->layout.getD() == 3);

    #ifndef NDEBUG

    Layer *dl = dataLayer ? dataLayer : fl;

    assert(dl->layout.getW() == bl->layout.getW());

    assert(dl->layout.getH() == bl->layout.getH());

    assert(dl->layout.getD() == bl->layout.getD());

    #endif

    imgsize = fl->layout.getActiveCount();

    fl->layout.split(flist, threadsCount);

    bl->layout.split(blist, threadsCount);

    pbl = bl->layout;

    pbl.padXY(pad);

    f = fopen(datafile, "rb");

    if (!f) return false;

    fseeko64(f, 0, SEEK_END);

    long long size = ftello64(f);

    count = size/imgsize;

    if (count < 1) return fclose(f), f = nullptr, false;

    workCount = itersPerBlock > count ? count : itersPerBlock;

    printf("allocated size: %lld\n", (long long)(imgsize*workCount));

    data.resize(workCount*imgsize);

    shuffle.resize(count);

    for(int i = 0; i < count; ++i)

      shuffle[i] = i;

    shuffle2.resize(workCount);

    for(int i = 0; i < workCount; ++i)

      shuffle2[i] = i;

    return loadBlocks();

    //return true;

  }

  void finish() override

    { if (f) fclose(f), f = nullptr; }

  bool loadBlocks() {

    for(int i = 0; i < workCount; ++i) {

      int j = rand()%count;

      if (i != j) std::swap(shuffle[i], shuffle[j]);

    }

    typedef std::pair<int, int> Pair;

    typedef std::set<Pair> Set;

    Set set;

    for(int i = 0; i < workCount; ++i)

      set.insert(Pair(shuffle[i], i));

    for(Set::iterator i = set.begin(); i != set.end(); ++i) {

      fseeko64(f, i->first*imgsize, SEEK_SET);

      if (!fread(data.data() + i->second*imgsize, imgsize, 1, f))

        return fclose(f), f = nullptr, false;

    }

    return true;

  }

  bool prepareBlock() override {

    for(int i = 0; i < workCount; ++i) {

      int j = rand()%workCount;

      if (i != j) std::swap(shuffle2[i], shuffle2[j]);

    }

    //return loadBlocks();

    return true;

  }

  void finishBlock() override {

    if (outfile && !dataLayer) {

      std::string outfile0(outfile);

      std::string outfile1 = outfile0 + ".1.tga";

      outfile0 += ".0.tga";

      unsigned char *id0 = data.data() + shuffle2[(itersPerBlock-1)%workCount]*imgsize;

      tgaSave(outfile0.c_str(), id0, fl->layout.getW(), fl->layout.getH(), fl->layout.getD());

      struct I: public Iter {

        typedef unsigned char* DataType;

        static inline void iter4(Neuron &n, DataType d, DataAccumType&) { *d = n.v < 0 ? 0 : n.v > 1 ? 255 : (unsigned char)(n.v*255.999); }

      };

      tmpdata.resize(imgsize);

      unsigned char *id1 = tmpdata.data();

      iterateNeurons2(bl->layout, bl->layout, bl->neurons, id1);

      tgaSave(outfile1.c_str(), id1, bl->layout.getW(), bl->layout.getH(), bl->layout.getD());

    }

  }

  void loadData(Barrier &barrier, int, int iter) override {

    struct I: public Iter {

      typedef const unsigned char* DataType;

      static inline void iter4(Neuron &n, DataType d, DataAccumType&) { n.v = *d/(NeuronReal)255; }

    };

    const unsigned char *id = data.data() + shuffle2[iter%workCount]*imgsize;

    iterateNeurons2(flist[barrier.tid], fl->layout, fl->neurons, id);

  }

  Quality verifyData(Barrier &barrier, int, int iter) override {

    Layout l = blist[barrier.tid];

    Layout dl = bl->layout;

    Layout pl = pbl;

    int d = l.getD();

    int w = l.getW();

    int dx = l.sz - d;

    int dy = (l.sx - w)*l.sz;

    int ddx = dl.getD();

    int ddy = (dl.getW() - w)*ddx;

    AccumReal aq = 0;

    NeuronReal ratio = this->ratio;

    Neuron *in = bl->neurons + (l.y0*l.sx + l.x0)*l.sz + l.z0;

    const unsigned char *id = data.data() + shuffle2[iter%workCount]*imgsize + ((l.y0-dl.y0)*l.sx + l.x0-dl.x0)*l.sz + l.z0-dl.z0;

    for(int y = l.y0; y < l.y1; ++y, in += dy, id += ddy) {

      bool outside = y < pl.y0 || y >= pl.y1;

      for(int x = l.x0; x < l.x1; ++x, in += dx, id += ddx) {

        if (outside || x < pl.x0 || x >= pl.x1) {

          for(Neuron *e = in + d; in < e; ++in) in->d = 0;

        } else {

          const unsigned char *iid = id;

          for(Neuron *e = in + d; in < e; ++in, ++iid) {

            NeuronReal v1 = *iid/(NeuronReal)255;

            NeuronReal v0 = in->v;

            NeuronReal diff = v1 - v0;

            in->d *= diff*ratio;

            aq += diff*diff;

          }

        }

      }

    }

    return Quality( sqrt(aq/pbl.getActiveCount()) );

  }

};

#endif