PyEstimator Class Reference

Class to invoke a pretrained python estimator that follows the sklearn interface. More...

#include <PyEstimator.h>

Public Member Functions
	PyEstimator (const std::string &pickleFileName)
	Construct the Estimator.
double	predict (const std::vector< double > &inputVariables)
	Call the predict method of the estimator.
double	predict (const std::vector< NamedFloatTuple * > &floatTuples)
	Call the predict method of the estimator.
double	predict (boost::python::object &array)
	Call the predict method of the estimator.

Private Member Functions
void	unpickleEstimator (const std::string &pickleFileName)
	Load the estimator object from the pickled file.
void	expand (size_t nVars)
	Reserves space for at least n variable in the input array.

Private Attributes
std::string	m_pickleFileName
	File name of the pickle file that contains the trained estimator.
boost::python::object	m_estimator
	Retrained python estimator object.
boost::python::object	m_predict
	Python bound prediction method - cached to avoid repeated lookup.
boost::python::object	m_array
	Array to be served to the estimator.
size_t	m_nCurrent = 0
	Cache for the current length of the input array.
bool	m_is_binary_classification = false
	Internal flag to keep track whether a binary classification with predict_proba is evaluated.

Detailed Description

Class to invoke a pretrained python estimator that follows the sklearn interface.

Definition at line 23 of file PyEstimator.h.

Constructor & Destructor Documentation

PyEstimator()

PyEstimator ( const std::string &  pickleFileName )

explicit

Construct the Estimator.

Parameters

pickleFileName

Name of the file that contains the pickled estimator object eg from sklearn

Definition at line 22 of file PyEstimator.cc.

  : m_pickleFileName(pickleFileName)
{
  try {
    // Construct an array with one entry
    // Expand it once the number of variables is known.
    boost::python::object numpy = boost::python::import("numpy");
    boost::python::list initValues;
    initValues.append(0.0);
    m_array = numpy.attr("array")(initValues);
    m_nCurrent = boost::python::len(m_array);
    // boost::python::object array = boost::python::import("array");
    // m_array = array.attr("array")("d");
    unpickleEstimator(pickleFileName);
  } catch (const boost::python::error_already_set&) {
    PyErr_Print();
    B2ERROR("Could not construct PyEstimator from " << pickleFileName);
  }
}

Member Function Documentation

expand()

void expand ( size_t  nVars )

private

Reserves space for at least n variable in the input array.

Definition at line 124 of file PyEstimator.cc.

{
  m_nCurrent = boost::python::len(m_array);
  if (nVars == m_nCurrent) return;
  try {
    boost::python::object numpy = boost::python::import("numpy");
    boost::python::object shape = boost::python::make_tuple(1, nVars); // one sample with nVars columns
    m_array = numpy.attr("resize")(m_array, shape);
  } catch (const boost::python::error_already_set&) {
    PyErr_Print();
    B2ERROR("Resize failed");
  }
}

predict() [1/3]

double predict

(

boost::python::object &

array

)

Call the predict method of the estimator.

Definition at line 71 of file PyEstimator.cc.

{
  boost::python::object predictions;
  try {
    predictions = m_predict(array);
    if (m_is_binary_classification) {
      // In case of a binary classification we take the signal probability
      boost::python::object prediction = predictions[0];
      return boost::python::extract<double>(prediction[1]);
    } else {
      // In case of regression we take the regression value
      boost::python::object prediction = predictions[0];
      return boost::python::extract<double>(prediction);
    }
  } catch (const boost::python::error_already_set&) {
    PyErr_Print();
    B2ERROR("Estimation failed in python object");
  }
  return NAN;
}

predict() [2/3]

double predict

(

const std::vector< double > &

inputVariables

)

Call the predict method of the estimator.

Definition at line 42 of file PyEstimator.cc.

{
  size_t nVars = inputVariables.size();
  expand(nVars);
 
  for (size_t iVar = 0; iVar < nVars; ++iVar) {
    m_array[boost::python::make_tuple(0, iVar)] = inputVariables[iVar];
  }
  return predict(m_array);
}

predict() [3/3]

double predict

(

const std::vector< NamedFloatTuple * > &

floatTuples

)

Call the predict method of the estimator.

Definition at line 53 of file PyEstimator.cc.

{
  size_t nVars = 0;
  for (NamedFloatTuple* floatTuple : floatTuples) {
    nVars += floatTuple->size();
  }
  expand(nVars);
  size_t iVar = 0;
 
  for (NamedFloatTuple* floatTuple : floatTuples) {
    for (size_t iTuple = 0; iTuple < floatTuple->size(); ++iTuple) {
      m_array[boost::python::make_tuple(0, iVar)] = floatTuple->get(iTuple);
      ++iVar;
    }
  }
  return predict(m_array);
}

unpickleEstimator()

void unpickleEstimator ( const std::string &  pickleFileName )

private

Load the estimator object from the pickled file.

Definition at line 92 of file PyEstimator.cc.

{
  try {
    std::string absPickleFilePath = FileSystem::findFile(pickleFileName);
    boost::python::object io = boost::python::import("io");
    boost::python::object pickle = boost::python::import("pickle");
    boost::python::object pickleFile = io.attr("open")(absPickleFilePath, "rb");
    boost::python::object estimator = pickle.attr("load")(pickleFile);
    m_estimator = estimator;
  } catch (const boost::python::error_already_set&) {
    PyErr_Print();
    B2ERROR("Could not open pickle file " << pickleFileName);
  }
 
  try {
    m_predict = m_estimator.attr("predict_proba");
    m_is_binary_classification = true;
  } catch (const boost::python::error_already_set&) {
    // AttributeError occured, but this is allowed to fail
    // Clear the exception and carry on.
    PyErr_Clear();
    B2INFO("Estimator in " << m_pickleFileName << " is not a binary classifier. Trying as regressor");
    try {
      m_predict = m_estimator.attr("predict");
      m_is_binary_classification = false;
    } catch (const boost::python::error_already_set&) {
      PyErr_Print();
      B2ERROR("Could neither find 'predict' not 'predict_proba' in python estimator from file " << pickleFileName);
    }
  }
}

Member Data Documentation

m_array

boost::python::object m_array

private

Array to be served to the estimator.

Definition at line 57 of file PyEstimator.h.

m_estimator

boost::python::object m_estimator

private

Retrained python estimator object.

Definition at line 51 of file PyEstimator.h.

m_is_binary_classification

bool m_is_binary_classification = false

private

Internal flag to keep track whether a binary classification with predict_proba is evaluated.

Definition at line 63 of file PyEstimator.h.

m_nCurrent

size_t m_nCurrent = 0

private

Cache for the current length of the input array.

Definition at line 60 of file PyEstimator.h.

m_pickleFileName

std::string m_pickleFileName

private

File name of the pickle file that contains the trained estimator.

Definition at line 48 of file PyEstimator.h.

m_predict

boost::python::object m_predict

private

Python bound prediction method - cached to avoid repeated lookup.

Definition at line 54 of file PyEstimator.h.

---

The documentation for this class was generated from the following files:

• tracking/trackFindingCDC/mva/include/PyEstimator.h
• tracking/trackFindingCDC/mva/src/PyEstimator.cc