Skip to content

[ML] Scale regularisers for final train #1755

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Feb 23, 2021
Merged

[ML] Scale regularisers for final train #1755

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
7c43b8c
Scale regularisers for final train
tveasey Feb 19, 2021
2c23e24
Test thresholds
tveasey Feb 22, 2021
691b36e
Docs
tveasey Feb 22, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions docs/CHANGELOG.asciidoc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,6 +45,10 @@

* Speed up training of regression and classification model training for data sets
with many features. (See {ml-pull}1746[#1746].)
* Avoid overfitting in final training by scaling regularizers to account for the
difference in the number of training examples. This results in a better match
between train and test error for classification and regression and often slightly
improved test errors. (See {ml-pull}1755[#1755].)

== {es} version 7.12.0

Expand Down
3 changes: 3 additions & 0 deletions include/maths/CBoostedTreeImpl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -311,6 +311,9 @@ class MATHS_EXPORT CBoostedTreeImpl final {
//! Set the hyperparamaters from the best recorded.
void restoreBestHyperparameters();

//! Scale the regulariser multipliers by \p scale.
void scaleRegularizers(double scale);

//! Check invariants which are assumed to hold after restoring.
void checkRestoredInvariants() const;

Expand Down
29 changes: 10 additions & 19 deletions lib/maths/CBoostedTreeFactory.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -769,32 +769,23 @@ void CBoostedTreeFactory::initializeUnsetDownsampleFactor(core::CDataFrame& fram
(logMinDownsampleFactor + logMaxDownsampleFactor) / 2.0};
LOG_TRACE(<< "mean log downsample factor = " << meanLogDownSampleFactor);

double previousDownsampleFactor{m_TreeImpl->m_DownsampleFactor};
double previousDepthPenaltyMultiplier{
double initialDownsampleFactor{m_TreeImpl->m_DownsampleFactor};
double initialDepthPenaltyMultiplier{
m_TreeImpl->m_Regularization.depthPenaltyMultiplier()};
double previousTreeSizePenaltyMultiplier{
double initialTreeSizePenaltyMultiplier{
m_TreeImpl->m_Regularization.treeSizePenaltyMultiplier()};
double previousLeafWeightPenaltyMultiplier{
double initialLeafWeightPenaltyMultiplier{
m_TreeImpl->m_Regularization.leafWeightPenaltyMultiplier()};

// We need to scale the regularisation terms to account for the difference
// in the downsample factor compared to the value used in the line search.
auto scaleRegularizers = [&](CBoostedTreeImpl& tree, double downsampleFactor) {
double scale{previousDownsampleFactor / downsampleFactor};
if (tree.m_RegularizationOverride.depthPenaltyMultiplier() == boost::none) {
tree.m_Regularization.depthPenaltyMultiplier(
scale * previousDepthPenaltyMultiplier);
}
if (tree.m_RegularizationOverride.treeSizePenaltyMultiplier() ==
boost::none) {
tree.m_Regularization.treeSizePenaltyMultiplier(
scale * previousTreeSizePenaltyMultiplier);
}
if (tree.m_RegularizationOverride.leafWeightPenaltyMultiplier() ==
boost::none) {
tree.m_Regularization.leafWeightPenaltyMultiplier(
scale * previousLeafWeightPenaltyMultiplier);
}
double scale{initialDownsampleFactor / downsampleFactor};
tree.m_Regularization.depthPenaltyMultiplier(initialDepthPenaltyMultiplier);
tree.m_Regularization.treeSizePenaltyMultiplier(initialTreeSizePenaltyMultiplier);
tree.m_Regularization.leafWeightPenaltyMultiplier(
initialLeafWeightPenaltyMultiplier);
tree.scaleRegularizers(scale);
return scale;
};

Expand Down
17 changes: 17 additions & 0 deletions lib/maths/CBoostedTreeImpl.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -261,6 +261,8 @@ void CBoostedTreeImpl::train(core::CDataFrame& frame,
LOG_TRACE(<< "Test loss = " << m_BestForestTestLoss);

this->restoreBestHyperparameters();
this->scaleRegularizers(allTrainingRowsMask.manhattan() /
m_TrainingRowMasks[0].manhattan());
this->startProgressMonitoringFinalTrain();
std::tie(m_BestForest, std::ignore, std::ignore) = this->trainForest(
frame, allTrainingRowsMask, allTrainingRowsMask, m_TrainingProgress);
Expand Down Expand Up @@ -1452,6 +1454,21 @@ void CBoostedTreeImpl::restoreBestHyperparameters() {
<< ", feature bag fraction* = " << m_FeatureBagFraction);
}

void CBoostedTreeImpl::scaleRegularizers(double scale) {
if (m_RegularizationOverride.depthPenaltyMultiplier() == boost::none) {
m_Regularization.depthPenaltyMultiplier(
scale * m_Regularization.depthPenaltyMultiplier());
}
if (m_RegularizationOverride.treeSizePenaltyMultiplier() == boost::none) {
m_Regularization.treeSizePenaltyMultiplier(
scale * m_Regularization.treeSizePenaltyMultiplier());
}
if (m_RegularizationOverride.leafWeightPenaltyMultiplier() == boost::none) {
m_Regularization.leafWeightPenaltyMultiplier(
scale * m_Regularization.leafWeightPenaltyMultiplier());
}
}

std::size_t CBoostedTreeImpl::numberHyperparametersToTune() const {
return m_RegularizationOverride.countNotSet() +
(m_DownsampleFactorOverride != boost::none ? 0 : 1) +
Expand Down
8 changes: 4 additions & 4 deletions lib/maths/unittest/CBoostedTreeTest.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -882,7 +882,7 @@ BOOST_AUTO_TEST_CASE(testCategoricalRegressors) {
LOG_DEBUG(<< "bias = " << modelBias);
LOG_DEBUG(<< " R^2 = " << modelRSquared);
BOOST_REQUIRE_CLOSE_ABSOLUTE(0.0, modelBias, 0.16);
BOOST_TEST_REQUIRE(modelRSquared > 0.97);
BOOST_TEST_REQUIRE(modelRSquared > 0.95);
}

BOOST_AUTO_TEST_CASE(testFeatureBags) {
Expand Down Expand Up @@ -1301,13 +1301,13 @@ BOOST_AUTO_TEST_CASE(testBinomialLogisticRegression) {
LOG_DEBUG(<< "log relative error = "
<< maths::CBasicStatistics::mean(logRelativeError));

BOOST_TEST_REQUIRE(maths::CBasicStatistics::mean(logRelativeError) < 0.681);
BOOST_TEST_REQUIRE(maths::CBasicStatistics::mean(logRelativeError) < 0.69);
meanLogRelativeError.add(maths::CBasicStatistics::mean(logRelativeError));
}

LOG_DEBUG(<< "mean log relative error = "
<< maths::CBasicStatistics::mean(meanLogRelativeError));
BOOST_TEST_REQUIRE(maths::CBasicStatistics::mean(meanLogRelativeError) < 0.51);
BOOST_TEST_REQUIRE(maths::CBasicStatistics::mean(meanLogRelativeError) < 0.52);
}

BOOST_AUTO_TEST_CASE(testImbalancedClasses) {
Expand Down Expand Up @@ -1389,7 +1389,7 @@ BOOST_AUTO_TEST_CASE(testImbalancedClasses) {
LOG_DEBUG(<< "recalls = " << core::CContainerPrinter::print(recalls));

BOOST_TEST_REQUIRE(std::fabs(precisions[0] - precisions[1]) < 0.1);
BOOST_TEST_REQUIRE(std::fabs(recalls[0] - recalls[1]) < 0.11);
BOOST_TEST_REQUIRE(std::fabs(recalls[0] - recalls[1]) < 0.13);
}

BOOST_AUTO_TEST_CASE(testClassificationWeightsOverride) {
Expand Down