int main(int argc, char** argv)
{
glutInit(&argc, argv);
myInit(argv[0]);
create_nurbs();
setRand();
glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
glutReshapeFunc(myReshape);
glutDisplayFunc(display);
glutIdleFunc(idle);
glutMainLoop();
return(0);
}
void model(int dim, float alpha1, float alpha2, float alpha3,
float beta6, float beta7, float beta8)
{
cout << "begin initialization: " << endl;
loadRating(DIR_PATH,rateMatrix); //初始化完成
mean = setMeanRating(rateMatrix); //求平均值，求bu和bi的值
int i,u,j,k;
//对bu，bi进行初始化,bu,bi的初始化的方法是求平均值，然后减去mean，
//在计算的过程中必须要记得所有的值，包括所有的打分总数和分数总和
int tmpIndex = 0;
for(i = 1; i < USER_NUM+1; ++i){
int vSize = rateMatrix[i].size();
for(j=0; j < vSize; ++j) {
bu[i] += rateMatrix[i][j].rate;
buNum[i] += 1;
bi[rateMatrix[i][j].item] += rateMatrix[i][j].rate;
biNum[rateMatrix[i][j].item] += 1;
}
}
//以下过程求平均值
for(i = 1; i < USER_NUM+1; ++i) {
if(buNum[i]>=1)bu[i] = (bu[i]/buNum[i]) - mean;
else bu[i] = 0.0;
//log<<i<<'\t'<<bu[i]<<endl;
//buBase[i] = bu[i];
}
for(i = 1; i < ITEM_NUM+1; ++i) {
if(biNum[i] >=1)bi[i] = (bi[i]/biNum[i]) - mean;
else bi[i] = 0.0;
//biBase[i] = bi[i];
//logbi<<i<<'\t'<<bi[i]<<endl;
}
//@todo 不知道是否能针对初始化的过程做一些优化
//对w ,c进行初始化，初始化的方法是随机函数，不知道这种方法是否好，是否会影响结果？？？？？？？
for(int i = 1; i < ITEM_NUM+1; ++i){
setRand(w[i],ITEM_NUM+1,0); //初始化q[i]
setRand(c[i],ITEM_NUM+1,0); //初始化x[i]
}
for(int i = 1; i < ITEM_NUM+1; ++i){
setRand(q[i],K_NUM+1,0); //初始化q[i]
setRand(y[i],K_NUM+1,0); //初始化p[i]
}
for(int i = 1; i < USER_NUM+1; ++i){
setRand(p[i],K_NUM+1,0); //初始化p[i]
}
cout <<"initialization end!"<<endl<< "begin iteration: " << endl;
float pui = 0.0 ; // 预测的u对i的打分
double preRmse = 1000000000000.0; //用于记录上一个rmse，作为终止条件的一种，如果rmse上升了，则停止
double nowRmse = 0.0;
cout <<"begin testRMSEProbe(): " << endl;
RMSEProbe();
//main loop
for(int step = 0; step < 70; ++step){ //只迭代35次
long double rmse = 0.0;
int n = 0;
for( u = 1; u < USER_NUM+1; ++u) { //循环处理每一个用户
int RuNum = rateMatrix[u].size(); //用户u打过分的item数目
float sqrtRuNum = 0.0;
if(RuNum>1) sqrtRuNum = (1.0/sqrt(RuNum));
//cacluate puTemp
for( k=1; k<K_NUM+1; ++k) {
double sumy = 0.0;
for(i=0; i < RuNum; ++i) {// 循环处理u打分过的每一个item
int itemI = rateMatrix[u][i].item;
sumy += y[itemI][k];
}
puTemp[u][k] = p[u][k]+ sqrtRuNum * sumy;
}
double sumQE[K_NUM+1] = {0.0};
//迭代处理
for(i=0; i < RuNum; ++i) {// 循环处理u打分过的每一个item
int itemI = rateMatrix[u][i].item;
short rui = rateMatrix[u][i].rate; //实际的打分
double bui = mean + bu[u] + bi[itemI];
pui = predictRate(u,itemI);
//cout<<u<<'\t'<<i<<'\t'<<pui<<'\t'<<rui<<endl;
float eui = rui - pui;
if( isnan(eui) ) {// fabs(eui) >= 4.2 ||
cout<<u<<'\t'<<i<<'\t'<<pui<<'\t'<<rui<<" "<<bu[u]<<" "<<bi[itemI]<<" "<<mean<<endl;
exit(1);
}
rmse += eui * eui; ++n;
if(n % 10000000 == 0)cout<<"step:"<<step<<" n:"<<n<<" dealed!"<<endl;
bu[u] += alpha1 * (eui - beta6 * bu[u]);
bi[itemI] += alpha1 * (eui - beta6 * bi[itemI]);
for( k=1; k< K_NUM+1; ++k) {
p[u][k] += alpha2 * (eui*q[itemI][k] - beta7*p[u][k]);
q[itemI][k] += alpha2 * (eui*puTemp[u][k] - beta7*q[itemI][k]);
sumQE[K_NUM+1] = eui * q[itemI][k];
}
for( j=0; j< RuNum; ++j) {
int itemJ = rateMatrix[u][j].item;
double ruj = (double)rateMatrix[u][j].rate;
w[itemI][itemJ] += alpha3 * (sqrtRuNum*eui*(ruj - mean -bu[u]-bi[itemJ]) - beta8*w[itemI][itemJ]);
c[itemI][itemJ] += alpha3 * (sqrtRuNum*eui - beta8*c[itemI][itemJ]);
}
}
for(j=0; j < RuNum; ++j) {// 循环处理u打分过的每一个item
int itemJ = rateMatrix[u][j].item;
for( k=1; k< K_NUM+1; ++k) {
y[itemJ][k] += alpha2 * (sqrtRuNum * sumQE[k] - beta7*y[itemJ][k]);
}
}
}
nowRmse = sqrt( rmse / n);
if( nowRmse >= preRmse && step >= 3) break; //如果rmse已经开始上升了，则跳出循环
else
preRmse = nowRmse;
cout << step << "\t" << nowRmse <<'\t'<< preRmse<<" n:"<<n<<endl;
//alpha1 *= 0.9;
//alpha2 *= 0.9;
//alpha3 *= 0.9;
RMSEProbe(); // 检查训练集情况
//alpha *= 0.992; //逐步减小学习速率
//RMSEProbe();
}
//ratingAll(data); //预测所有的结果
RMSEProbe(); // 检查训练集情况
return;
}
void model(int dim, float alpha, float beta)
{
cout << "begin initialization: " << endl;
loadRating(DIR_PATH,rateMatrix); //初始化完成
mean = setMeanRating(rateMatrix); //求平均值，求bu和bi的值
int i,u,j,k;
//对bu，bi进行初始化,bu,bi的初始化的方法是求平均值，然后减去mean，
//在计算的过程中必须要记得所有的值，包括所有的打分总数和分数总和
int tmpIndex = 0;
for(i = 1; i < USER_NUM+1; ++i){
int vSize = rateMatrix[i].size();
for(j=0; j < vSize; ++j) {
bi[rateMatrix[i][j].item] += (rateMatrix[i][j].rate - mean);
biNum[rateMatrix[i][j].item] += 1;
}
}
//ofstream logbi("bi.txt");
for(i = 1; i < ITEM_NUM+1; ++i) {
if(biNum[i] >=1)bi[i] = bi[i]/(biNum[i]+25);
else bi[i] = 0.0;
}
for(i = 1; i < USER_NUM+1; ++i){
int vSize = rateMatrix[i].size();
for(j=0; j < vSize; ++j) {
bu[i] += (rateMatrix[i][j].rate - mean - bi[rateMatrix[i][j].item]);
buNum[i] += 1;
}
}
//以下过程求平均值
for(i = 1; i < USER_NUM+1; ++i) {
if(buNum[i]>=1)bu[i] = bu[i]/(buNum[i]+10);
else bu[i] = 0.0;
//logbi<<i<<" "<<buNum[i]<<" "<<bu[i]<<endl;
}
//logbi.close();
//@todo 不知道是否能针对初始化的过程做一些优化
//对w进行初始化，初始化的方法是随机函数，不知道这种方法是否好，是否会影响结果？？？？？？？
for(int i = 1; i < ITEM_NUM+1; ++i){
setRand(q[i],K_NUM+1,mean); //初始化q[i]
}
for(int i = 1; i < USER_NUM+1; ++i){
setRand(p[i],K_NUM+1,mean); //初始化p[i]
}
cout <<"initialization end!"<<endl<< "begin iteration: " << endl;
float pui = 0.0 ; // 预测的u对i的打分
double preRmse = 1000000000000.0; //用于记录上一个rmse，作为终止条件的一种，如果rmse上升了，则停止
double nowRmse = 0.0;
cout <<"begin testRMSEProbe(): " << endl;
RMSEProbe();
//main loop
for(int step = 0; step < 90; ++step){ //只迭代60次
long double rmse = 0.0;
int n = 0;
for( u = 1; u < USER_NUM+1; ++u) { //循环处理每一个用户
int RuNum = rateMatrix[u].size(); //用户u打过分的item数目
float sqrtRuNum = 0.0;
if(RuNum>1) sqrtRuNum = (1.0/sqrt(RuNum));
//迭代处理
for(i=0; i < RuNum; ++i) {// 循环处理u打分过的每一个item
int itemI = rateMatrix[u][i].item;
short rui = rateMatrix[u][i].rate; //实际的打分
pui = predictRate(u,itemI);
float eui = rui - pui;
if( isnan(eui) ) {// fabs(eui) >= 4.2 ||
cout<<u<<'\t'<<i<<'\t'<<pui<<'\t'<<rui<<" "<<bu[u]<<" "<<bi[itemI]<<" "<<mean<<endl;
//printArray(q[itemI],p[u],K_NUM+1);
exit(1);
}
rmse += eui * eui; ++n;
if(n % 10000000 == 0)cout<<"step:"<<step<<" n:"<<n<<" dealed!"<<endl;
bu[u] += alpha * (eui - beta * bu[u]);
bi[itemI] += alpha * (eui - beta * bi[itemI]);
for( k=1; k< K_NUM+1; ++k) {
p[u][k] += alpha * (eui*q[itemI][k] - beta*p[u][k]);
q[itemI][k] += alpha * (eui*p[u][k] - beta*q[itemI][k]);
}
}
}
nowRmse = sqrt( rmse / n);
cout << step << "\t" << nowRmse <<'\t'<< preRmse<<" n:"<<n<<endl;
if( nowRmse >= preRmse && step >= 3) break; //如果rmse已经开始上升了，则跳出循环
else
preRmse = nowRmse;
RMSEProbe(); // 检查训练集情况
//alpha *= 0.999; //逐步减小学习速率
//RMSEProbe();
}
RMSEProbe(); // 检查训练集情况
return;
}
void jen::Strat(int minR,int maxR,int minC,int maxC, SitRep sitrep)
{
dir = lt;
if(rank == crown)
{
r = ((maxR -minR)/2 + minR);
c = ((maxC -minC)/2 + minC);
return;
}
if(rank == knight)
{
if(sitrep.thing[(((maxR -minR)/2 + minR))+1][((maxC -minC)/2 + minC)].rank!=knight)
{
r = (((maxR -minR)/2 + minR))+1;
c = ((maxC -minC)/2 + minC);
return;
}
else if(sitrep.thing[(((maxR -minR)/2 + minR))-1][((maxC -minC)/2 + minC)].rank!=knight)
{
r = (((maxR -minR)/2 + minR))-1;
c = ((maxC -minC)/2 + minC);
return;
}
else if(sitrep.thing[(((maxR -minR)/2 + minR))][(((maxC -minC)/2 + minC))+1].rank!=knight)
{
r = (((maxR -minR)/2 + minR));
c = (((maxC -minC)/2 + minC))+1;
return;
}
else if(sitrep.thing[(((maxR -minR)/2 + minR))][(((maxC -minC)/2 + minC))-1].rank!=knight)
{
r = (((maxR -minR)/2 + minR));
c = (((maxC -minC)/2 + minC))-1;
return;
}
else
{
setRand(minR, maxR, minC, maxC, sitrep);
return;
}
}
else if(rank == archer)
{
if(sitrep.thing[(((maxR -minR)/2 + minR))+1][(((maxC -minC)/2 + minC))+1].rank!=archer)
{
r = (((maxR -minR)/2 + minR))+1;
c = (((maxC -minC)/2 + minC))+1;
return;
}
else if(sitrep.thing[(((maxR -minR)/2 + minR))+1][(((maxC -minC)/2 + minC))-1].rank!=archer)
{
r = (((maxR -minR)/2 + minR))+1;
c = (((maxC -minC)/2 + minC))-1;
return;
}
else if(sitrep.thing[(((maxR -minR)/2 + minR))-1][(((maxC -minC)/2 + minC))-1].rank!=archer)
{
r = (((maxR -minR)/2 + minR))-1;
c = (((maxC -minC)/2 + minC))-1;
return;
}
else if(sitrep.thing[(((maxR -minR)/2 + minR))-1][(((maxC -minC)/2 + minC))+1].rank!=archer)
{
r = (((maxR -minR)/2 + minR))-1;
c = (((maxC -minC)/2 + minC))+1;
return;
}
if(sitrep.thing[(((maxR -minR)/2 + minR))+2][(((maxC -minC)/2 + minC))+1].rank!=archer)
{
r = (((maxR -minR)/2 + minR))+2;
c = (((maxC -minC)/2 + minC))+1;
return;
}
else if(sitrep.thing[(((maxR -minR)/2 + minR))+2][(((maxC -minC)/2 + minC))-1].rank!=archer)
{
r = (((maxR -minR)/2 + minR))+2;
c = (((maxC -minC)/2 + minC))-1;
return;
}
else if(sitrep.thing[(((maxR -minR)/2 + minR))-2][(((maxC -minC)/2 + minC))-1].rank!=archer)
{
r = (((maxR -minR)/2 + minR))-2;
c = (((maxC -minC)/2 + minC))-1;
return;
}
else if(sitrep.thing[(((maxR -minR)/2 + minR))-2][(((maxC -minC)/2 + minC))-1].rank!=archer)
{
r = (((maxR -minR)/2 + minR))-2;
c = (((maxC -minC)/2 + minC))+1;
return;
}
else
{
setRand(minR, maxR, minC, maxC, sitrep);
return;
}
}
else if (rank==infantry)
{
sideOfBoard loca = findSide(minR, maxR, minC, maxC);
if(loca==brt)
{
int i, j;
for(j=minC; j<maxC;j++)
{
for(i=minR; i<maxR; i++)
{
if(sitrep.thing[i][j].what==space)
{
r=i;
c=j;
return;
}
}
}
}
if(loca==blt)
{
int i, j;
for(j=maxC; j>minC;j--)
{
for(i=maxR; i>minR; i--)
{
if(sitrep.thing[i][j].what==space)
{
r=i;
c=j;
return;
}
}
}
}
}
}
void model(int dim, double & alpha, double & beta)
{
vector<Rating> data;
loadRating("u1.base",data); //初始化
//1、首先搞清楚整个过程
//（1）求平均值
//2、一步一步实现每个过程
//3、什么时候弄完，什么时候回去
mean = setMeanRating(data); //求平均值，求bu和bi的值
//中间删除一段代码，是xlvector计算平均打分写论文用的
double cc = 1 / (double)(dim);
//cout << cc << endl;exit(1);
//初始化bu，就是user的bias（偏置）
for(int i = 1; i < USER_NUM+1; ++i){
//cout<< bu[i] << endl;
//if(bu[i] < 0) continue; //这里bu[i]小于0永远不成立，这个语句永远不会执行
p[i].clear(); //清空所有的p[i]
setRand(p[i],dim,cc); //初始化p[i]
}
cout << bi.size() << endl;
for(int i = 1; i < ITEM_NUM+1; ++i){
//if(bi[i] < 0) continue;
//cout << i << endl;
q[i].clear();
setRand(q[i],dim,cc);
}
cout << "begin : " << endl;
for(int step = 0; step < 25; ++step){ //只迭代25次
double rmse = 0;
double n = 0;
for(int ii = 0; ii < data.size(); ++ii){ //利用每一个打分调优结果
int i = ii;
//下面两句不知道做什么用？
//if(data[i].value() > 3) continue;
//if(data[i].test > test_level) continue;
int user = data[i].user;
int item = data[i].item;
double rui = data[i].value();
double pui = rating(data[i]); //预测打分，向量点乘
double err = rui - pui;
rmse += err * err; ++n;
//更新数据
bu[user] += alpha * (err - beta * bu[user]);
bi[item] += alpha * (err - beta * bi[item]);
for(int k = 0; k < dim; ++k){
p[user][k] += alpha * (err * q[item][k] - beta * p[user][k]);
q[item][k] += alpha * (err * p[user][k] - beta * q[item][k]);
}
}
cout << step << "\t" << sqrt(rmse / n) << endl;
alpha *= 0.98; //逐步减小学习速率
}
ratingAll(data); //预测所有的结果
RMSEProbe(); // 检查训练集情况
//输出bu的值
ofstream outputbu("bu.txt");
for(int i=1; i < USER_NUM+1; i++)
{
outputbu << i << "\t" << bu[i] << endl;
}
outputbu.close();
//输出bi的值
ofstream outputbi("bi.txt");
for(int i=1; i < ITEM_NUM+1; i++)
{
outputbi << i << "\t"<< bi[i] << endl;
}
outputbu.close();
}
void model(int dim, double & alpha, double & beta)
{
vector<Rating> data;
loadRating("u1.base",data); //初始化
/* 检查初始化是否正确
for(int i=0; i<data.size();i++){
cout << data[i].user<<" "<<data[i].item<<" "<<data[i].rate<<" "<<endl;
}
cout<<data.size()<<endl;
exit(1);
*/
//1、首先搞清楚整个过程
//（1）求平均值
//2、一步一步实现每个过程
//3、什么时候弄完，什么时候回去
setMeanRating(data); //求平均值
map<int, int> ni; //？
map<int, map<string, int> > itr, itn; //一个存储某部电影每月的积分总和，一个存储某部电影当月打分的总数量
for(int i = 0; i < data.size(); ++i){
int user = data[i].user;
int item = data[i].item;
bu[user] = 0; //svd++，用户偏差
bi[item] = 0; //item偏差
if(data[i].test == 0){ //test应该是将一批数据随机赋值为1，
if(i % 1000000 == 0) cout << i << endl; //每一百万个数据打印一下，以前共100M的打分
ostringstream oss;
oss << data[i].year() << "-" << data[i].month() << "-" << 1;//data[i].mday();
itr[item][oss.str()] += data[i].real; //存储每月的打分总和值
itn[item][oss.str()] += 1; //存储每月的数量
ni[item]++; //该电影打分总次数
}
}
//out主要记录月份对item打分的总的影响
ofstream out("netflix-bit.csv");
out << endl;
for(map<int, map<string, int> >::iterator i = itr.begin(); i != itr.end(); ++i){
for(map<string, int>::iterator t = i->second.begin(); t != i->second.end(); ++t){
out << i->first << "," << t->first << "," ;
out << (double)(t->second) / (double)(itn[i->first][t->first]) << "," ;
out << itn[i->first][t->first] << "," << ni[i->first] << endl;
//itemId,2002-7-1,itemId在2002-7的平均打分，itemId在2002-7的打分总数，该电影打分总次数
}
}
out.close();
//return; 这个地方是原来有的，估计是xlvector测试用的，可以不管
double cc = 1 / (double)(dim);
//初始化bu，就是user的bias（偏置）
for(int i = 0; i < bu.size(); ++i){
if(bu[i] < 0) continue; //这里bu[i]小于0永远不成立，这个语句永远不会执行
p[i].clear(); //清空所有的p[i]
setRand(p[i],dim,cc); //初始话p[i]
}
for(int i = 0; i < bi.size(); ++i){
if(bi[i] < 0) continue;
q[i].clear();
setRand(q[i],dim,cc);
}
cout << "begin : " << endl;
for(int step = 0; step < 25; ++step){ //只迭代25次
double rmse = 0;
double n = 0;
for(int ii = 0; ii < data.size(); ++ii){ //利用每一个打分调优结果
int i = ii;
//下面两句不知道做什么用？
if(data[i].value() > 3) continue;
if(data[i].test > test_level) continue;
int user = data[i].user;
int item = data[i].item;
double rui = data[i].value();
double pui = rating(data[i]); //预测打分，向量点乘
double err = rui - pui;
rmse += err * err; ++n;
//更新数据
bu[user] += alpha * (err - beta * bu[user]);
bi[item] += alpha * (err - beta * bi[item]);
for(int k = 0; k < dim; ++k){
p[user][k] += alpha * (err * q[item][k] - beta * p[user][k]);
q[item][k] += alpha * (err * p[user][k] - beta * q[item][k]);
}
}
cout << step << "\t" << sqrt(rmse / n) << "\t";
RMSEProbe(data); // 检查训练集情况
alpha *= 0.92; //逐步减小学习速率
}
ratingAll(data); //预测所有的结果
if(test_level == 0){
ostringstream oss;
oss << "pb-svd-" << dim << ".txt";
ratingProbe(oss.str().c_str(), data);
saveRating("../data/xlvector-svd20.data0", data);
}
if(test_level == 1){
ostringstream oss1;
oss1 << "qf-svd-" << dim << ".txt";
ratingQuiz(oss1.str().c_str(), data);
}
}
void model(int dim, double & alpha, double & beta)
{
vector<Rating> data;
loadRating("../data/netflix3.data",data); //初始化
setMeanRating(data); //求平均值
map<int, int> ni;
map<int, map<string, int> > itr, itn;
for(int i = 0; i < data.size(); ++i){
int user = data[i].user;
int item = data[i].item;
bu[user] = 0;
bi[item] = 0;
if(data[i].test == 0){
if(i % 1000000 == 0) cout << i << endl;
ostringstream oss;
oss << data[i].year() << "-" << data[i].month() << "-" << 1;//data[i].mday();
itr[item][oss.str()] += data[i].real;
itn[item][oss.str()] += 1;
ni[item]++;
}
}
ofstream out("netflix-bit.csv");
out << endl;
for(map<int, map<string, int> >::iterator i = itr.begin(); i != itr.end(); ++i){
for(map<string, int>::iterator t = i->second.begin(); t != i->second.end(); ++t){
out << i->first << "," << t->first << "," << (double)(t->second) / (double)(itn[i->first][t->first]) << "," << itn[i->first][t->first] << "," << ni[i->first] << endl;
}
}
out.close();
return;
double cc = 1 / (double)(dim);
for(int i = 0; i < bu.size(); ++i){
if(bu[i] < 0) continue;
p[i].clear();
setRand(p[i],dim,cc);
}
for(int i = 0; i < bi.size(); ++i){
if(bi[i] < 0) continue;
q[i].clear();
setRand(q[i],dim,cc);
}
cout << "begin : " << endl;
for(int step = 0; step < 25; ++step){
double rmse = 0;
double n = 0;
for(int ii = 0; ii < data.size(); ++ii){
int i = ii;
if(data[i].value() > 3) continue;
if(data[i].test > test_level) continue;
int user = data[i].user;
int item = data[i].item;
double rui = data[i].value();
double pui = rating(data[i]);
double err = rui - pui;
rmse += err * err; ++n;
bu[user] += alpha * (err - beta * bu[user]);
bi[item] += alpha * (err - beta * bi[item]);
for(int k = 0; k < dim; ++k){
p[user][k] += alpha * (err * q[item][k] - beta * p[user][k]);
q[item][k] += alpha * (err * p[user][k] - beta * q[item][k]);
}
}
cout << step << "\t" << sqrt(rmse / n) << "\t";
RMSEProbe(data);
alpha *= 0.92;
}
ratingAll(data);
if(test_level == 0){
ostringstream oss;
oss << "pb-svd-" << dim << ".txt";
ratingProbe(oss.str().c_str(), data);
saveRating("../data/xlvector-svd20.data0", data);
}
if(test_level == 1){
ostringstream oss1;
oss1 << "qf-svd-" << dim << ".txt";
ratingQuiz(oss1.str().c_str(), data);
}
}