double ff::FBNN::nntest(const FMatrix& x, const FMatrix& y)
{
// std::cout << "start nntest" << std::endl;
FColumn labels;
nnpredict(x,y,labels);
FColumn expected = rowMaxIndexes(y);
std::vector<int> bad = findUnequalIndexes(labels,expected);
// std::cout << "end nntest" << std::endl;
return double(bad.size()) / x.rows();//Haven't return bad vector.(nntest.m does)
}
double nneval(NN* nn, int numdata, const double** test_x, const double** test_y)
{
int i, num_err=0;
for(i=0; i< numdata; i++){
if ( nnpredict(nn, i, test_x) != indOfMaxVal_double( test_y[i], nn->layer[ nn->n -2 ].units )){
num_err ++;
}
}
return (double) num_err / numdata ;
}
int main(int argc, char* argv[]){
int c, i;
char *inputImageFile = NULL;
char *inputLabelFile = NULL;
char *inputModel = "noname.model";
char *outputResult = "result.txt";
while( (c=getopt(argc, argv, "i:y:m:o:h")) != -1 ){
switch(c)
{
case 'h':
printf("usage: \n-i input image file\n-m inputModel\n-o outputResult\n-y (optional)input label file\n\tif provided, it calculate accuracy immediately.\n");
break;
case 'i':
inputImageFile = optarg;
break;
case 'y':
inputLabelFile = optarg;
break;
case 'm':
inputModel = optarg;
break;
case 'o':
outputResult = optarg;
break;
case '?':
printf("Illegal option\n");
printf("usage: \n-i input image file\n-m inputModel\n-o outputResult\n-y (optional)input label file\n\tif provided, it calculate accuracy immediately.\n");
exit(1);
break;
default:
printf("usage: \n-i input image file\n-m inputModel\n-o outputResult\n-y (optional)input label file\n\tif provided, it calculate accuracy immediately.\n");
exit(1);
break;
}
}
int test_num ;
double **test_x = mnist_load_data(inputImageFile, &test_num);
int dim = 28 * 28; // supposed data dimension
// normalization parameters
double* mean = (double*) malloc( dim * sizeof(double));
double* sigma = (double*) malloc( dim * sizeof(double));
// load from model
NN* mm = importModel(mean, sigma, inputModel);
normalize_zscore_apply(test_x, test_num, mean, sigma, dim);
if(inputLabelFile != NULL){
double **test_y = mnist_load_labels(inputLabelFile, &test_num);
double errorRate = nneval(mm, test_num, (const double**)test_x, (const double**) test_y);
printf("Error rate=%f\n", errorRate);
}
// write result
FILE* fp = fopen(outputResult, "w");
for(i=0; i< test_num; i++){
fprintf(fp, "%d\n", nnpredict(mm, i, (const double**)test_x) );
}
fclose(fp);
return 0;
}
