void ActGraph::trainModel(const std::string featureFile)
{
// features
cv::Mat stat; // mean_x, mean_y, var_x, var_y
cv::Mat shapeXY; // track shapes
cv::Mat mbhX; // MBHx
cv::Mat mbhY; // MBHy
// import features from file
readFeatures(featureFile, stat, shapeXY, mbhX, mbhY);
// GMM-clustering
trainModel(stat, shapeXY, mbhX, mbhY);
}
int main(int argc, char** argv)
{
if (argc < 3)
{
fprintf(stderr, "Usage %s input output\n", argv[0]);
exit (-1);
}
featureValues* head = retriveFeatures(argv[1]);
dumpFeatureInfo(head);
aggInstFeature af[INST_TYPE_NUM];
readFeatures(af, argv[1]);
int i;
for (i=0; i<INST_TYPE_NUM; i++)
{
printf("%s:%u\n", InstTypeName[i], af[i].count);
}
return 0;
}
int
main (int argc, char** argv)
{
// const char trace = 0;
// const unsigned debug = 0;
const char *me = "main";
// parse and validate the command line options
struct paramsStruct params;
parseCommandLine(argc, argv, ¶ms);
// create results directory in ANALYSIS directory
// permissions are read, write for the owner
char resultsDirectoryName[256] = "";
makeResultsDirectory(resultsDirectoryName,
sizeof(resultsDirectoryName),
¶ms);
// start logging
char logFilePath[256] = "";
{
const int outputLength = snprintf(logFilePath,
sizeof(logFilePath),
"%s/run.log",
resultsDirectoryName);
if (outputLength > (int) sizeof(logFilePath)) {
fprintf(stderr, "%s: logFilePath too small", me);
exit(1);
}
}
Log_T log = Log_new(logFilePath, stderr);
// log the command line parameters
LOG(log,"started log file %s\n", logFilePath);
LOG(log,"params: algo=%s\n", params.algo);
LOG(log," : obs=%s\n", params.obs);
LOG(log," : radius=%d\n", params.radius);
LOG(log," : which=%s\n", params.which);
// check the command line parameters
assert(strcmp(params.algo, "knn") == 0);
assert(strcmp(params.obs, "1A") == 0);
// read the input files
const unsigned nObservations = 217376; // adjust of OBS != 1A
const unsigned nFeatures = 55;
double *apns = readCsvNoHeader(nObservations, "aps.csv");
double *dates = readCsvNoHeader(nObservations, "date.csv");
char *featuresHeaderP;
double *features = readFeatures(nObservations,
nFeatures,
&featuresHeaderP);
double *prices = readCsvNoHeader(nObservations, "SALE-AMOUNT-log.csv");
// convert dates to days past the epoch
unsigned dayStdColumn = 5; // the 6th column contains the standardized day value
assert(columnHeaderEqual(featuresHeaderP, dayStdColumn, "day-std"));
double *days = convertDatesToDays(nObservations, dates);
free(dates);
double mean;
double stdv;
determineMeanStdv(nObservations, days, &mean, &stdv);
double *daysStd = standardize(nObservations, days, mean, stdv);
replaceDay(nObservations, nFeatures, features, daysStd, dayStdColumn);
free(days);
free(daysStd);
// generate one set of estimates
FILE *resultFile;
{
char resultFilePath[256];
const int outputLength =
snprintf(resultFilePath,
sizeof(resultFilePath),
"%s/estimates-laufer.csv",
resultsDirectoryName);
if (outputLength > (int) sizeof(resultFilePath)) {
fprintf(stderr, "%s: resultFilePath too small", me);
exit(1);
}
LOG(log, " result file path: %s\n", resultFilePath);
resultFile = fopen(resultFilePath, "w");
}
assert(resultFile);
if (strcmp(params.which, "laufer"))
createLaufer(nObservations, nFeatures,
apns, dates, features, prices,
log,
resultFile);
else
assert(NULL != "logic error");
// OLD CODE BELOW THIS LINE
#if 0
double **pricesHatP = NULL;
if (params.useCache)
pricesHatP = readCache(nObservations, params.obs, log, kMax);
// determine estimated prices for any missing entries in the cache
// this operation could be fast or very slow
// MAYBE: write out cache periodically
const unsigned cacheMutated = completeCache(nObservations,
pricesHatP,
params.obs,
log,
kMax,
pricesP,
debug);
if (params.useCache && cacheMutated)
writeCache(nObservations, pricesHatP, params.obs, log, kMax);
// select which set of estimates to create
if (paramsP->whichIsLaufer)
createEstimatesLaufer(nObservations, nFeatures,
features, dates, prices);
else
assert(false); // should never get here
// pricesHatP[i][k] is
// the estimate priced of transaction indexed i for k nearest neighbors
// for each value of k, determine RMSE overall all the test transactions
// determine kArgMin, the k providing the lowest RMSE
// write CSV containing <k, rmse> values
char resultFilePath[256];
{
const int outputLength =
snprintf(resultFilePath,
sizeof(resultFilePath),
"%s/k-rmse.csv",
directoryName);
if (outputLength > (int) sizeof(resultFilePath)) {
fprintf(stderr, "%s: resultFilePath too small", me);
exit(1);
}
LOG(log, " result file path: %s\n", resultFilePath);
}
FILE *resultFile = fopen(resultFilePath, "w");
assert(resultFile);
// log best k for random sample of test observations
bestK(0.01, nObservations, pricesHatP, pricesP, log, kMax);
// write CSV header
fprintf(resultFile, "k,rmse\n");
unsigned kArgMin = 0;
double lowestRMSE = DBL_MAX;
for (unsigned hpK = 0; hpK < kMax; hpK++) {
// determine rmse for this k
const double rmse = determineRmse(nObservations, pricesHatP, pricesP, hpK);
// check if we have a new best k
LOG(log, "hpK %u rmse %f\n", hpK + 1, rmse);
fprintf(resultFile, "%u,%f\n", hpK + 1, rmse);
if (rmse < lowestRMSE) {
lowestRMSE = rmse;
kArgMin = hpK;
}
}
#endif
//
LOG(log, "%s\n", "finished");
exit(0);
}
