double NeuralNet::trainNet(const std::vector<double>& data, const std::vector<double>& trainingOutput, const unsigned int outType)
{
int outputLayer = m_layers.size();
int inputLayer = 0;
std::vector<double> output,error,delta,prevOut,nextDeltas;
double cost = 0.0;
// run net forward
output = runNet( data );
error = computeError( output, trainingOutput );
switch(outType)
{
case SCALAR:
cost = computeMSE( error );
break;
case PROB:
cost = logloss( output, trainingOutput );
break;
}
// propagate error backward through layers
for(int i=outputLayer; i>0; i--)
{
// calculate initial gradient
if( i==outputLayer )
delta = error;
else
delta = m_layers[i-1]->computeDeltas(error,m_layers[i]->retrieveWeights());
if( i > 1)
prevOut = m_layers[ i-1-1 ]->retrieveOutputs();
if( i <= 1)
prevOut = data;
m_layers[i-1]->updateWeights( prevOut,delta,m_learningRate,m_momentum,m_weightDecay );
error = delta;
}
return cost;
}
void bestFit()
{
initHF();
while(1)
{
// uint32_t start_time = getClockTime();
// read in the samples one by one.
// for each sample, simultaneously calculate
// the incremental inner product across all the hF polynomials.
RxAndComputeInnerProducts();
// At this point you have the projections. Calculate
// MSE for each projection..
computeMSE();
// At this point, we have the best SI.
double SI = best_sigma_index;
write_float64("out0_data", SI);
// write_float64("out1_data", dotP0[0]);
int I;
write_float64("out1_data", dotP0[best_sigma_index]);
write_float64("out1_data", dotP1[best_sigma_index]);
write_float64("out1_data", dotP2[best_sigma_index]);
write_float64("out1_data", dotP3[best_sigma_index]);
write_float64("out1_data", dotP4[best_sigma_index]);
write_float64("out1_data", dotP5[best_sigma_index]);
// }
// uint32_t end_time = getClockTime();
// write_uint32("elapsed_time_pipe", (end_time - start_time));
initFit();
}
}
void runPmfBatchGradientDescentOMP(vector<users> &allUsers,
vector<business> &allBusiness) {
/* Switch off the dynamic thread setting. */
omp_set_dynamic(0);
/* Set the number of threads. */
omp_set_num_threads(NO_OF_TRIALS);
/* Output file object. */
ofstream fout;
/* Input file object. */
ifstream fin;
#if !FORCE_INPUT
editInputBatchTextForGradientDescent();
#endif
/* Open input batch file mentioning different models. */
fin.open(BATCH_INPUT_TEXT);
/* Read number of models. */
unsigned int n;
fin >> n;
/* Read the model parameters. */
unsigned int *latentSpace = new unsigned int[n];
unsigned int *maxIterations = new unsigned int[n];
double *lambdaU = new double[n];
double *lambdaV = new double[n];
for (unsigned int i = 0; i < n; i++) {
fin >> latentSpace[i];
fin >> maxIterations[i];
fin >> lambdaU[i];
fin >> lambdaV[i];
}
fin.close();
/* Log file name. */
string logFileName;
/* Get the log file name. */
logFileName = getLogFileName(LOG_BATCH_RESULTS, 0, 0);
/* Open the log file. */
fout.open(logFileName.c_str());
/* Now run each of these models and test them against the validation and
* test data sets. */
for (unsigned int i = 0; i < n; i++) {
/* Declare the model. */
collaborativeFiltering collabFilteringModel[NO_OF_TRIALS];
/* Initialize the root mean square errors to 0.*/
double rmseTraining[NO_OF_TRIALS] = { 0.0 };
double rmseValidation[NO_OF_TRIALS] = { 0.0 };
double rmseTest[NO_OF_TRIALS] = { 0.0 };
double rmsTrainError = 0.0;
double rmsValidationError = 0.0;
double rmsTestError = 0.0;
/* Train the model for some fixed number of times and accumulate the
* error results and log its average values. */
/* Launch as per given thread ID. */
cout << "Running PMF Algorithm with K = " << latentSpace[i] << " for "
<< maxIterations[i] << " iterations, lambda U = " << lambdaU[i]
<< ", lambda V = " << lambdaV[i] << endl << endl;
collaborativeFiltering collabFilteringTest;
/* Initialize the model. */
initCollabFilteringModel(collabFilteringTest, allUsers, allBusiness,
latentSpace[0], maxIterations[0], lambdaU[0], lambdaV[0], true);
vector<review> testReviews = collabFilteringTest.testReviews;
vector<users> trainUsers = collabFilteringTest.trainUsers;
vector<business> trainBusiness = collabFilteringTest.trainBusiness;
deinitCollabFilteringModel(collabFilteringTest);
#pragma omp parallel
{
/* Get the thread ID. */
unsigned int j = omp_get_thread_num();
/* Initialize the model. */
collabFilteringModel[j].testReviews = testReviews;
initCollabFilteringModel(collabFilteringModel[j], trainUsers,
trainBusiness, latentSpace[i], maxIterations[i], lambdaU[i],
lambdaV[i], false);
/* Train the model. */
probablisticMatrixFactorizationGradientDescent(
collabFilteringModel[j]);
#if LOG_FEATURES
cout << "Logging the computed features." << endl;
/* Save the estimated user and business data. */
logUserBusinessFeatures(collabFilteringModel[j]);
cout << "Validating the computed features." << endl;
/* Validate the computed features with existing reviews on training
* data. */
validateAndLogReviews(collabFilteringModel[j], TRAINING_DATA);
/* Validate the computed features with existing reviews on
* validation data. */
validateAndLogReviews(collabFilteringModel[j], VALIDATION_DATA);
/* Validate the computed features with existing reviews on testing
* data. */
validateAndLogReviews(collabFilteringModel[j], TESTING_DATA);
#endif
#if LOG_MSE
cout << "Logging the mean square error after every iteration."
<< endl;
/* Log the mean square error. */
logMsePerIteration(collabFilteringModel[j]);
#endif
/* Compute the error on training data set. */
rmseTraining[j] = computeMSE(collabFilteringModel[j],
TRAINING_DATA);
/* Compute the error on validation data set. */
rmseValidation[j] = computeMSE(collabFilteringModel[j],
VALIDATION_DATA);
/* Compute the error on testing data set. */
rmseTest[j] = computeMSE(collabFilteringModel[j], TESTING_DATA);
/* De-initialize the model. */
deinitCollabFilteringModel(collabFilteringModel[j]);
}
for (unsigned int j = 0; j < NO_OF_TRIALS; j++) {
rmsTestError += rmseTest[j];
rmsValidationError += rmseValidation[j];
rmsTrainError += rmseTraining[j];
}
/* Log the findings along with the model specifications. */
fout << "Model " << setw(7) << i + 1 << ", ";
fout << setw(13) << setprecision(5) << latentSpace[i] << ", ";
fout << setw(13) << setprecision(5) << maxIterations[i] << ", ";
fout << setw(13) << setprecision(5) << lambdaU[i] << ", ";
fout << setw(13) << setprecision(5) << lambdaV[i] << ", ";
fout << setw(13) << setprecision(5) << rmsTrainError / NO_OF_TRIALS
<< ", ";
fout << setw(13) << setprecision(5) << rmsValidationError / NO_OF_TRIALS
<< ", ";
fout << setw(13) << setprecision(5) << rmsTestError / NO_OF_TRIALS
<< ", ";
fout << endl;
}
/* Close the log file.*/
fout.close();
/* Switch on the dynamic thread setting. */
omp_set_dynamic(1);
delete latentSpace;
delete maxIterations;
delete lambdaU;
delete lambdaV;
}