static void Apply(HMMType& hmm, void* /* extraInfo */)
{
// Load observations.
const string inputFile = CLI::GetParam<string>("input_file");
mat dataSeq;
data::Load(inputFile, dataSeq, true);
// See if transposing the data could make it the right dimensionality.
if ((dataSeq.n_cols == 1) && (hmm.Emission()[0].Dimensionality() == 1))
{
Log::Info << "Data sequence appears to be transposed; correcting."
<< endl;
dataSeq = dataSeq.t();
}
// Verify correct dimensionality.
if (dataSeq.n_rows != hmm.Emission()[0].Dimensionality())
Log::Fatal << "Observation dimensionality (" << dataSeq.n_rows << ") "
<< "does not match HMM Gaussian dimensionality ("
<< hmm.Emission()[0].Dimensionality() << ")!" << endl;
arma::Col<size_t> sequence;
hmm.Predict(dataSeq, sequence);
// Save output.
const string outputFile = CLI::GetParam<string>("output_file");
data::Save(outputFile, sequence, true);
}
static void Apply(HMMType& hmm, vector<mat>* trainSeqPtr)
{
const bool batch = CLI::HasParam("batch");
const double tolerance = CLI::GetParam<double>("tolerance");
// Do we need to replace the tolerance?
if (CLI::HasParam("tolerance"))
hmm.Tolerance() = tolerance;
const string labelsFile = CLI::GetParam<string>("labels_file");
// Verify that the dimensionality of our observations is the same as the
// dimensionality of our HMM's emissions.
vector<mat>& trainSeq = *trainSeqPtr;
for (size_t i = 0; i < trainSeq.size(); ++i)
if (trainSeq[i].n_rows != hmm.Emission()[0].Dimensionality())
Log::Fatal << "Dimensionality of training sequence " << i << " ("
<< trainSeq[i].n_rows << ") is not equal to the dimensionality of "
<< "the HMM (" << hmm.Emission()[0].Dimensionality() << ")!"
<< endl;
vector<arma::Row<size_t>> labelSeq; // May be empty.
if (labelsFile != "")
{
// Do we have multiple label files to load?
char lineBuf[1024];
if (batch)
{
fstream f(labelsFile);
if (!f.is_open())
Log::Fatal << "Could not open '" << labelsFile << "' for reading."
<< endl;
// Now read each line in.
f.getline(lineBuf, 1024, '\n');
while (!f.eof())
{
Log::Info << "Adding training sequence labels from '" << lineBuf
<< "'." << endl;
// Now read the matrix.
Mat<size_t> label;
data::Load(lineBuf, label, true); // Fatal on failure.
// Ensure that matrix only has one row.
if (label.n_cols == 1)
label = trans(label);
if (label.n_rows > 1)
Log::Fatal << "Invalid labels; must be one-dimensional." << endl;
// Check all of the labels.
for (size_t i = 0; i < label.n_cols; ++i)
{
if (label[i] >= hmm.Transition().n_cols)
{
Log::Fatal << "HMM has " << hmm.Transition().n_cols << " hidden "
<< "states, but label on line " << i << " of '" << lineBuf
<< "' is " << label[i] << " (should be between 0 and "
<< (hmm.Transition().n_cols - 1) << ")!" << endl;
}
}
labelSeq.push_back(label.row(0));
f.getline(lineBuf, 1024, '\n');
}
f.close();
}
else
{
Mat<size_t> label;
data::Load(labelsFile, label, true);
// Ensure that matrix only has one row.
if (label.n_cols == 1)
label = trans(label);
if (label.n_rows > 1)
Log::Fatal << "Invalid labels; must be one-dimensional." << endl;
// Verify the same number of observations as the data.
if (label.n_elem != trainSeq[labelSeq.size()].n_cols)
Log::Fatal << "Label sequence " << labelSeq.size() << " does not have"
<< " the same number of points as observation sequence "
<< labelSeq.size() << "!" << endl;
// Check all of the labels.
for (size_t i = 0; i < label.n_cols; ++i)
{
if (label[i] >= hmm.Transition().n_cols)
{
Log::Fatal << "HMM has " << hmm.Transition().n_cols << " hidden "
<< "states, but label on line " << i << " of '" << labelsFile
<< "' is " << label[i] << " (should be between 0 and "
<< (hmm.Transition().n_cols - 1) << ")!" << endl;
}
}
labelSeq.push_back(label.row(0));
}
// Now perform the training with labels.
hmm.Train(trainSeq, labelSeq);
}
else
{
// Perform unsupervised training.
hmm.Train(trainSeq);
}
// Save the model.
if (CLI::HasParam("output_model_file"))
{
const string modelFile = CLI::GetParam<string>("output_model_file");
SaveHMM(hmm, modelFile);
}
}