void test(
Net& model,
torch::Device device,
DataLoader& data_loader,
size_t dataset_size) {
torch::NoGradGuard no_grad;
model.eval();
double test_loss = 0;
int32_t correct = 0;
for (const auto& batch : data_loader) {
auto data = batch.data.to(device), targets = batch.target.to(device);
auto output = model.forward(data);
test_loss += torch::nll_loss(
output,
targets,
/*weight=*/{},
Reduction::Sum)
.template item<float>();
auto pred = output.argmax(1);
correct += pred.eq(targets).sum().template item<int64_t>();
}
test_loss /= dataset_size;
std::printf(
"\nTest set: Average loss: %.4f | Accuracy: %.3f\n",
test_loss,
static_cast<double>(correct) / dataset_size);
}
void TrainModelNN (_String* model, _String* matrix)
{
_String errMsg;
long modelIdx = modelNames.Find(model);
_Parameter verbI;
checkParameter (VerbosityLevelString, verbI, 0.0);
char buffer [128];
if (modelIdx < 0) {
errMsg = *model & " did not refer to an existring model";
} else {
_Variable* boundsMatrix = FetchVar (LocateVarByName (*matrix));
if (boundsMatrix && (boundsMatrix->ObjectClass() == MATRIX)) {
_Matrix * bmatrix = (_Matrix*) boundsMatrix->GetValue ();
if (bmatrix->IsAStringMatrix() && (bmatrix->GetVDim () == 3)) {
_Variable* modelMatrix = LocateVar (modelMatrixIndices.lData[modelIdx]);
_SimpleList modelVariableList;
{
_AVLList mvla (&modelVariableList);
modelMatrix->ScanForVariables (mvla, true);
mvla.ReorderList();
}
if (bmatrix->GetHDim () == modelVariableList.lLength) {
// now map model variables to bounds matrix
_SimpleList variableMap;
_String *myName;
for (long k = 0; k < modelVariableList.lLength; k++) {
myName = ((_FString*)bmatrix->GetFormula(k,0)->Compute())->theString;
long vID = LocateVarByName (*myName);
if (vID < 0) {
break;
}
vID = variableNames.GetXtra (vID);
vID = modelVariableList.Find(vID);
if (vID < 0) {
break;
}
variableMap << vID;
}
if (variableMap.lLength == modelVariableList.lLength) {
_Matrix vBounds (variableMap.lLength,2, false, true);
long k2 = 0;
for (; k2 < variableMap.lLength; k2++) {
_Parameter lb = ((_FString*)bmatrix->GetFormula(k2,1)->Compute())->theString->toNum(),
ub = ((_FString*)bmatrix->GetFormula(k2,2)->Compute())->theString->toNum();
if ( ub>lb || k2) {
vBounds.Store (k2,0,lb);
vBounds.Store (k2,1,ub);
if (ub<=lb && vBounds (k2-1,0) <= vBounds (k2-1,1) && (!CheckEqual(vBounds (k2-1,0),0.0) || !CheckEqual(vBounds (k2-1,1),1.0))) {
break;
}
}
}
if (k2 == modelVariableList.lLength) {
// set up the sampling now
_String fName = ProcessLiteralArgument (&ModelNNFile,nil);
FILE* nnFile = doFileOpen (fName.getStr(), "w");
if (nnFile) {
_Matrix* modelMatrix = (_Matrix*) LocateVar(modelMatrixIndices.lData[modelIdx])->GetValue();
_Parameter mainSteps,
checkSteps,
errorTerm,
loopMax,
hiddenNodes,
absError,
nn1,
nn2;
long fullDimension = modelMatrix->GetHDim() * modelMatrix->GetVDim();
checkParameter (ModelNNTrainingSteps, mainSteps, 10000.0);
checkParameter (ModelNNVerificationSample, checkSteps, 500.0);
checkParameter (ModelNNPrecision, errorTerm, 0.01);
checkParameter (ModelNNTrainingSteps, loopMax, 10);
checkParameter (ModelNNHiddenNodes, hiddenNodes, 5);
checkParameter (ModelNNLearningRate, nn1, .3);
checkParameter (ModelNNPersistenceRate, nn2, .1);
Net** matrixNet = new Net* [fullDimension] ;
for (long i = 0; i < fullDimension; i++) {
checkPointer (matrixNet [i] = new Net (variableMap.lLength,(long)hiddenNodes,1,errorTerm,nn1,nn2,100,200,true));
//matrixNet[i]->verbose = true;
}
checkPointer (matrixNet);
_List tIn,
tOut;
FILE* varSamples = doFileOpen ("variableSamples.out", "w");
fprintf (varSamples, "%s" ,LocateVar(modelVariableList.lData[0])->GetName()->getStr());
for (long vc = 1; vc < modelVariableList.lLength; vc++) {
fprintf (varSamples, ",%s" ,LocateVar(modelVariableList.lData[variableMap.lData[vc]])->GetName()->getStr());
}
fprintf (varSamples, "\n");
for (long itCount = 0; itCount < loopMax; itCount ++) {
if (verbI > 5) {
snprintf (buffer, sizeof(buffer), "\nNeural Network Pass %ld. Building a training set...\n", itCount);
BufferToConsole (buffer);
}
while (tIn.countitems() < mainSteps) {
NNMatrixSampler (0, vBounds, modelVariableList, variableMap, modelMatrix, tIn, tOut);
}
_Matrix inData (mainSteps, variableMap.lLength, false, true);
_Parameter *md = inData.theData;
for (long matrixC = 0; matrixC < mainSteps; matrixC++) {
_Parameter * ed = ((_Matrix*)tIn (matrixC))->theData;
fprintf (varSamples, "\n%g",*ed);
*md = *ed;
ed++;
md++;
for (long entryC = 1; entryC < variableMap.lLength; entryC++, ed++, md++) {
*md = *ed;
fprintf (varSamples, ",%g", *md);
}
}
tIn.Clear();
if (verbI > 5) {
BufferToConsole ( "Done Building Training Set. Training...\n");
}
long lastDone = 0;
for (long cellCount = 0; cellCount < fullDimension; cellCount++) {
Net* thisCell = matrixNet[cellCount];
_Matrix outVector (mainSteps, 1, false, true);
for (long oc = 0; oc < mainSteps; oc++) {
outVector.theData[oc] = ((_Matrix*)tOut(oc))->theData[cellCount];
}
thisCell->studyAll (inData.theData, outVector.theData, mainSteps);
long nowDone = (cellCount+1)*100./fullDimension;
if (nowDone > lastDone) {
snprintf (buffer, sizeof(buffer),"%ld%% done\n", lastDone = nowDone);
BufferToConsole (buffer);
}
}
tOut.Clear();
if (verbI > 5) {
BufferToConsole ( "Done Training. Resampling...\n");
}
_PMathObj tObj = _Constant(0).Time();
_Parameter time1 = tObj->Value(),
time2;
while (tIn.countitems() < checkSteps) {
NNMatrixSampler (0, vBounds, modelVariableList, variableMap, modelMatrix, tIn, tOut);
}
absError = 0.0;
DeleteObject (tObj);
tObj = _Constant(0).Time();
time2 = tObj->Value();
if (verbI > 5) {
snprintf (buffer, sizeof(buffer),"Done Resampling in %g seconds. Computing Error...\n", time2-time1);
BufferToConsole (buffer);
}
_Parameter maxValT,
maxValE;
for (long verCount = 0; verCount < checkSteps; verCount++) {
_Parameter* inData = ((_Matrix*)tIn(verCount))->theData,
* outData = ((_Matrix*)tOut(verCount))->theData;
for (long cellCount = 0; cellCount < fullDimension; cellCount++) {
Net *thisCell = matrixNet[cellCount];
_Parameter estVal = thisCell->eval(inData)[0],
trueVal = outData[cellCount],
localError;
localError = estVal-trueVal;
if (localError < 0) {
localError = -localError;
}
if (absError < localError) {
maxValT = trueVal;
maxValE = estVal;
absError = localError;
}
}
}
DeleteObject (tObj);
tObj = _Constant(0).Time();
time1 = tObj->Value();
DeleteObject (tObj);
if (verbI > 5) {
snprintf (buffer, sizeof(buffer), "Done Error Checking in %g seconds. Got max abs error %g on the pair %g %g\n", time1-time2, absError, maxValT, maxValE);
BufferToConsole (buffer);
}
if (absError <= errorTerm) {
break;
}
}
if (absError > errorTerm) {
ReportWarning (_String("Couldn't achive desired precision in TrainModelNN. Achieved error of ") & absError);
}
fclose (varSamples);
fprintf (nnFile,"{{\n\"%s\"", LocateVar(modelVariableList.lData[0])->GetName()->getStr());
_Matrix newBounds (modelVariableList.lLength, 2, false, true);
if (vBounds(0,0)>vBounds(0,1)) {
newBounds.Store (variableMap.lData[0],0,0.);
newBounds.Store (variableMap.lData[0],1,1.);
} else {
newBounds.Store (variableMap.lData[0],0,vBounds(0,0));
newBounds.Store (variableMap.lData[0],1,vBounds(0,1));
}
for (long varCounter = 1; varCounter < modelVariableList.lLength; varCounter ++) {
fprintf (nnFile,",\n\"%s\"", LocateVar(modelVariableList.lData[varCounter])->GetName()->getStr());
if (vBounds(varCounter,0)>vBounds(varCounter,1)) {
newBounds.Store (variableMap.lData[varCounter],0,0.);
newBounds.Store (variableMap.lData[varCounter],1,1.);
} else {
newBounds.Store (variableMap.lData[varCounter],0,vBounds(varCounter,0));
newBounds.Store (variableMap.lData[varCounter],1,vBounds(varCounter,1));
}
}
fprintf (nnFile,"\n}}\n");
newBounds.toFileStr (nnFile);
for (long i2 = 0; i2 < fullDimension; i2++) {
matrixNet[i2]->save(nnFile);
delete matrixNet [i2];
}
fclose (nnFile);
delete matrixNet;
} else {
errMsg = _String ("Failed to open ") & fName & " for writing";
}
} else {
errMsg = _String ("Invalid variable bounds in row ") & (k2+1) & " of the bounds matrix";
}
} else {
errMsg = *myName & " was not one of the model parameters";
}
} else {
errMsg = *matrix & " must be a have the same number of rows as the number of model parameters";
}
} else {
errMsg = *matrix & " must be a string matrix with 3 columns";
}
} else {
errMsg = *matrix & " was not the identifier of a valid matrix variable";
}
}
if (errMsg.sLength) {
errMsg = errMsg & _String(" in call to TrainModelNN.");
WarnError (errMsg);
}
}