Skip to content

SDRClassifier: fix precision by using Real64 for PDF #667

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
breznak merged 11 commits into from
Sep 20, 2019
Merged

SDRClassifier: fix precision by using Real64 for PDF #667

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
ccf7e10
SDRClassifier: fix precision by using Real64 for PDF
breznak Sep 16, 2019
d437a6a
Classifier: more fixes
breznak Sep 16, 2019
d34d0cd
Classifier: make infer const
breznak Sep 16, 2019
bf5123b
Predictor: make infer const
breznak Sep 16, 2019
2609617
Merge branch 'master_community' into predictor_precision_fix
breznak Sep 17, 2019
22d6664
review: formatting
breznak Sep 17, 2019
4fd8fbb
Merge branch 'master_community' into predictor_precision_fix
breznak Sep 18, 2019
767f89a
Merge branch 'master_community' into predictor_precision_fix
breznak Sep 19, 2019
2aedac2
Predictor: use hashmap
breznak Sep 19, 2019
4d8708d
Classifier: simplify asserts
breznak Sep 19, 2019
df7ae4e
Predictor: remove numRecord arg, fixes
breznak Sep 19, 2019
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
68 changes: 34 additions & 34 deletions src/htm/algorithms/SDRClassifier.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,15 +44,11 @@ void Classifier::initialize(const Real alpha)
}


PDF Classifier::infer(const SDR & pattern)
{
// Check input dimensions, or if this is the first time the Classifier has
// been used then initialize it with the given SDR's dimensions.
PDF Classifier::infer(const SDR & pattern) const {
// Check input dimensions, or if this is the first time the Classifier is used and dimensions
// are unset, return zeroes.
if( dimensions_.empty() ) {
dimensions_ = pattern.dimensions;
while( weights_.size() < pattern.size ) {
weights_.push_back( vector<Real>( numCategories_, 0.0f ));
}
return PDF(numCategories_, 0.0f); //empty
} else if( pattern.dimensions != dimensions_ ) {
stringstream err_msg;
err_msg << "Classifier input SDR.dimensions mismatch: previously given SDR with dimensions ( ";
Expand Down Expand Up @@ -81,6 +77,15 @@ PDF Classifier::infer(const SDR & pattern)

void Classifier::learn(const SDR &pattern, const vector<UInt> &categoryIdxList)
{
// If this is the first time the Classifier is being used, weights are empty,
// so we set the dimensions to that of the input `pattern`
if( dimensions_.empty() ) {
dimensions_ = pattern.dimensions;
while( weights_.size() < pattern.size ) {
const auto initialEmptyWeights = PDF( numCategories_, 0.0f );
weights_.push_back( initialEmptyWeights );
}
}
// Check if this is a new category & resize the weights table to hold it.
const auto maxCategoryIdx = *max_element(categoryIdxList.begin(), categoryIdxList.end());
if( maxCategoryIdx >= numCategories_ ) {
Expand All @@ -93,7 +98,7 @@ void Classifier::learn(const SDR &pattern, const vector<UInt> &categoryIdxList)
}

// Compute errors and update weights.
const vector<Real> error = calculateError_(categoryIdxList, pattern);
const auto& error = calculateError_(categoryIdxList, pattern);
for( const auto& bit : pattern.getSparse() ) {
for(size_t i = 0u; i < numCategories_; i++) {
weights_[bit][i] += alpha_ * error[i];
Expand All @@ -103,9 +108,8 @@ void Classifier::learn(const SDR &pattern, const vector<UInt> &categoryIdxList)


// Helper function to compute the error signal in learning.
std::vector<Real> Classifier::calculateError_(
const std::vector<UInt> &categoryIdxList, const SDR &pattern)
{
std::vector<Real64> Classifier::calculateError_(const std::vector<UInt> &categoryIdxList,
const SDR &pattern) const {
// compute predicted likelihoods
auto likelihoods = infer(pattern);

Expand Down Expand Up @@ -165,13 +169,12 @@ void Predictor::reset() {
}


Predictions Predictor::infer(const UInt recordNum, const SDR &pattern)
{
updateHistory_( recordNum, pattern );
Predictions Predictor::infer(const UInt recordNum, const SDR &pattern) const {
checkMonotonic_(recordNum);
Copy link
Copy Markdown
Member Author

@breznak breznak Sep 16, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@ctrl-z-9000-times please have a look on the last 1,(2) commits, I intended to make infer const, as it imho should have been. For Classifier that was easy, for Predictor I had to remove the updateHistory_ from infer (no tests visibly broken).

Copy link
Copy Markdown
Collaborator

@ctrl-z-9000-times ctrl-z-9000-times Sep 17, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why does it matter if infer is constant? I doubt it will have any performance impact, and I don't think this will prevent any programming mistakes.

Consider the following chain of events:

infer( t, SDR-A ) -> PDF
learn( t+1, SDR-B, Labels )

The method updateHistory stores the given SDR inside of the predictor. Previously the call to learn would have associated SDR-A with Labels since that SDR was given to infer with. Now that will not happen.

Also, these changes allow the timestamps to the infer method to go backwards.

infer( t + 2, ... )
infer( t, ... )

I'm not saying the new behavior is wrong, just that it changed.

Copy link
Copy Markdown
Member Author

@breznak breznak Sep 17, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I doubt it will have any performance impact, and I don't think this will prevent any programming mistakes.

right, there won't be any performance gains, and the behavior has changed. I think it adds (expected) semantics, and better separates responsibilities of those 2 functions:

  • you know only what you learn()
  • and you can infer() const anytime without any worry of changing the state (comes with loosened requirement for monotonic timestamps)

Copy link
Copy Markdown
Collaborator

@ctrl-z-9000-times ctrl-z-9000-times Sep 17, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok, that makes sense +1


Predictions result;
for( const auto step : steps_ ) {
result[step] = classifiers_[step].infer( pattern );
result[step] = classifiers_.at(step).infer( pattern );
}
return result;
}
Expand All @@ -180,7 +183,18 @@ Predictions Predictor::infer(const UInt recordNum, const SDR &pattern)
void Predictor::learn(const UInt recordNum, const SDR &pattern,
const std::vector<UInt> &bucketIdxList)
{
updateHistory_( recordNum, pattern );
checkMonotonic_(recordNum);

// Update pattern history if this is a new record.
const UInt lastRecordNum = recordNumHistory_.empty() ? -1 : recordNumHistory_.back();
if (recordNumHistory_.size() == 0u || recordNum > lastRecordNum) {
patternHistory_.emplace_back( pattern );
recordNumHistory_.push_back(recordNum);
if (patternHistory_.size() > steps_.back() + 1u) {
patternHistory_.pop_front();
recordNumHistory_.pop_front();
}
}

// Iterate through all recently given inputs, starting from the furthest in the past.
auto pastPattern = patternHistory_.begin();
Expand All @@ -197,24 +211,10 @@ void Predictor::learn(const UInt recordNum, const SDR &pattern,
}


void Predictor::updateHistory_(const UInt recordNum, const SDR & pattern)
{
void Predictor::checkMonotonic_(const UInt recordNum) const {
// Ensure that recordNum increases monotonically.
UInt lastRecordNum = -1;
const UInt lastRecordNum = recordNumHistory_.empty() ? -1 : recordNumHistory_.back();
if( not recordNumHistory_.empty() ) {
lastRecordNum = recordNumHistory_.back();
if (recordNum < lastRecordNum) {
NTA_THROW << "The record number must increase monotonically.";
}
NTA_CHECK(recordNum >= lastRecordNum) << "The record number must increase monotonically.";
}

// Update pattern history if this is a new record.
if (recordNumHistory_.size() == 0u || recordNum > lastRecordNum) {
patternHistory_.emplace_back( pattern );
recordNumHistory_.push_back(recordNum);
if (patternHistory_.size() > steps_.back() + 1u) {
patternHistory_.pop_front();
recordNumHistory_.pop_front();
}
}
}
30 changes: 21 additions & 9 deletions src/htm/algorithms/SDRClassifier.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,13 @@

/** @file
* Definitions for the SDR Classifier & Predictor.
*
* `Classifier` learns mapping from SDR->input value (encoder's output). This is used when you need to "explain" the HTM network back to real-world, ie. mapping SDRs
* back to digits in MNIST digit classification task.
*
* `Predictor` has similar functionality for time-sequences where you want to "predict" N-steps ahead and then return real-world value.
* Internally it uses (several) Classifiers, and in nupic.core this used to be part for SDRClassifier, for htm.core this is a separate class Predictor.
*
*/

#ifndef NTA_SDR_CLASSIFIER_HPP
Expand All @@ -43,7 +50,7 @@ namespace htm {
*
* See also: https://en.wikipedia.org/wiki/Probability_distribution
*/
using PDF = std::vector<Real>;
using PDF = std::vector<Real64>; //Real64 (not Real/float) must be used here, otherwise precision is lost and Predictor never reaches sufficient results.

/**
* Returns the category with the greatest probablility.
Expand Down Expand Up @@ -115,7 +122,7 @@ class Classifier : public Serializable
* @returns: The Probablility Distribution Function (PDF) of the categories.
* This is indexed by the category label.
*/
PDF infer(const SDR & pattern);
PDF infer(const SDR & pattern) const;

/**
* Learn from example data.
Expand Down Expand Up @@ -147,12 +154,13 @@ class Classifier : public Serializable
/**
* 2D map used to store the data.
* Use as: weights_[ input-bit ][ category-index ]
* Real64 (not just Real) so the computations do not lose precision.
*/
std::vector<std::vector<Real>> weights_;
std::vector<std::vector<Real64>> weights_;

// Helper function to compute the error signal for learning.
std::vector<Real> calculateError_(const std::vector<UInt> &bucketIdxList,
const SDR &pattern);
std::vector<Real64> calculateError_(const std::vector<UInt> &bucketIdxList,
const SDR &pattern) const;
};

/**
Expand All @@ -179,9 +187,11 @@ using Predictions = std::map<UInt, PDF>;
* This class handles missing datapoints.
*
* Compatibility Note: This class is the replacement for the old SDRClassifier.
* It no longer provides estimates of the actual value.
* It no longer provides estimates of the actual value. Instead, users can get a rough estimate
* from bucket-index. If more precision is needed, use more buckets in the encoder.
*
* Example Usage:
* ```
* // Predict 1 and 2 time steps into the future.
* // Make a sequence of 4 random SDRs. Each SDR has 1000 bits and 2% sparsity.
* vector<SDR> sequence( 4, { 1000 } );
Expand All @@ -208,6 +218,7 @@ using Predictions = std::map<UInt, PDF>;
* Predictions B = pred.infer( 1, sequence[1] );
* argmax( B[1] ) -> labels[2]
* argmax( B[2] ) -> labels[3]
* ```
*/
class Predictor : public Serializable
{
Expand Down Expand Up @@ -242,7 +253,7 @@ class Predictor : public Serializable
*
* @returns: A mapping from prediction step to PDF.
*/
Predictions infer(UInt recordNum, const SDR &pattern);
Predictions infer(const UInt recordNum, const SDR &pattern) const;

/**
* Learn from example data.
Expand All @@ -252,7 +263,8 @@ class Predictor : public Serializable
* @param pattern: The active input SDR.
* @param bucketIdxList: Vector of the current value bucket indices or categories.
*/
void learn(UInt recordNum, const SDR &pattern,
void learn(const UInt recordNum,
const SDR &pattern,
const std::vector<UInt> &bucketIdxList);

CerealAdapter;
Expand All @@ -276,7 +288,7 @@ class Predictor : public Serializable
// Stores the input pattern history, starting with the previous input.
std::deque<SDR> patternHistory_;
std::deque<UInt> recordNumHistory_;
void updateHistory_(UInt recordNum, const SDR & pattern);
void checkMonotonic_(UInt recordNum) const;

// One per prediction step
std::map<UInt, Classifier> classifiers_;
Expand Down