This package is intended for Microsoft use only. It is not designed to be consumed as an independent package.
@microsoft/bf-dispatcher
'bf-dispatcher' is a generic NLP intent classification package/library.
It currently can do:
Sampling
> Bootstrap resampling
> Auto active learning down sampling
> Stratified down sampling
Evaluation and Reporting
> Cross Validation against a training set
> Test again a test set
To support these features, it internally contains a Softmax Regression (MaxEnt) learner that can consume sparse text features and train models to support auto-active-learning down-sampling and cross validation. It can also produce model quality reports.
Currently, 'bf-dispatcher' can consume two forms of input sources: LU and TSV columnar files. It uses the bf-lu package (https://github.com/microsoft/botframework-cli/tree/master/packages/lu) to parse and load a .lu file as input.
To demonstrate the auto active learning process, there are some example functions implemented in
"src/model/supervised/classifier/auto_active_learning/AppAutoActiveLearner.ts" that can do the following:
a) consume a LU or columnar TSV file,
b) use a bootstrap resampler to select training instances based on a prior label/instance (intent/utterance)
distribition,
c) iterate through batches of input utterance/label records, and train models to select most relevant utterance/intent
pairs through an auto active learning process, and
d) use a stratifier sampler to limit the training size.
Below are some examples of using the AutoActiveLearner class.
/**
* This function can read a LU file with intent and utterance data and run through 3 steps of
* sampling processes:
* 0) Bootstrap Resampling
* 1) Auto Active Learning Sampling
* 2) Stratified Sampling
*
* @param luContent - a .lu file content in string form as input.
* @param doBootstrapResampling - boolean flag to activate bootstrap resampling (BRS) logic or not.
* @param brsDistribution - explicit distribution to control bootstrap resampling process
* @param doAutoActiveLearning - boolean flag to activate auto active leaning (AAL) process or not.
* @param aalLimitInitialNumberOfInstancesPerCategory - AAL initial number of instances per category/label/intent.
* @param aalNumberOfInstancesPerIteration - AAL number of instances processed per iterations.
* @param aalInstanceSelectionThreshold - AAL threshold to pick a tested instance for training in next iteration
* @param learnerParameterEpochs - AAL Softmax Regression learner parameter - number of epochs.
* @param learnerParameterMiniBatchSize - AAL Softmax Regression learner parameter - mini-batch size.
* @param learnerParameterL1Regularization - AAL Softmax Regression learner parameter - L1 regularization.
* @param learnerParameterL2Regularization - AAL Softmax Regression learner parameter - L2 regularization.
* @param learnerParameterLossEarlyStopRatio - AAL Softmax Regression learner parameter - early stop ratio.
* @param learnerParameterLearningRate - AAL Softmax Regression learner parameter - learning rate.
* @param learnerParameterToCalculateOverallLossAfterEpoch - AAL Softmax Regression learner parameter - flag
* @param limitingSampleSize - sample size controled by a final stratified sampling process.
*/
public static async mainAutoActiveLearnerWithLuContent(
luContent: string,
doBootstrapResampling: boolean =
AppAutoActiveLearner.defaultDoBootstrapResampling,
brsDistribution: TMapStringKeyGenericValue<number> =
DictionaryMapUtility.newTMapStringKeyGenericValue<number>(),
doAutoActiveLearning: boolean =
AutoActiveLearner.defaultDoAutoActiveLearning,
aalLimitInitialNumberOfInstancesPerCategory: number =
AutoActiveLearner.defaultAalLimitInitialNumberOfInstancesPerCategory,
aalNumberOfInstancesPerIteration: number =
AutoActiveLearner.defaultAalNumberOfInstancesPerIteration,
aalInstanceSelectionThreshold: number =
AutoActiveLearner.defaultAalInstanceSelectionThreshold,
learnerParameterEpochs: number =
AppSoftmaxRegressionSparse.defaultEpochs,
learnerParameterMiniBatchSize: number =
AppSoftmaxRegressionSparse.defaultMiniBatchSize,
learnerParameterL1Regularization: number =
AppSoftmaxRegressionSparse.defaultL1Regularization,
learnerParameterL2Regularization: number =
AppSoftmaxRegressionSparse.defaultL2Regularization,
learnerParameterLossEarlyStopRatio: number =
AppSoftmaxRegressionSparse.defaultLossEarlyStopRatio,
learnerParameterLearningRate: number =
AppSoftmaxRegressionSparse.defaultLearningRate,
learnerParameterToCalculateOverallLossAfterEpoch: boolean =
true,
limitingSampleSize: number =
DefaultLimitingSampleSize): Promise<{
"newLuData": LuData,
"learner": SoftmaxRegressionSparse,
"seedingInstanceIndexArray": number[],
"seedingInstanceIndexArrayInitial": number[],
}> {
// -----------------------------------------------------------------------
let luData: LuData =
await LuData.createLuData(
luContent,
new NgramSubwordFeaturizer(),
true);
// -----------------------------------------------------------------------
if (doBootstrapResampling) {
const bootstrapSamplerKeyMap: BootstrapSamplerKeyMapDistribution<number> =
new BootstrapSamplerKeyMapDistribution<number>(
brsDistribution,
luData.getIntentInstanceIndexMapArray());
// ---- NOTE-FOR-REFERENCE ---- const bootstrapSamplerKeyMap: BootstrapSamplerKeyMap<number> =
// ---- NOTE-FOR-REFERENCE ---- new BootstrapSamplerKeyMap(data.getIntentInstanceIndexMapArray());
Utility.debuggingLog(`luData.getIntentInstanceIndexMapArray()=` +
`${Utility.mapToJsonSerialization(luData.getIntentInstanceIndexMapArray())}`);
Utility.debuggingLog(`bootstrapSamplerKeyMap.computeSamplingNumberInstancesPerLabel()=` +
`${bootstrapSamplerKeyMap.computeSamplingNumberInstancesPerLabel()}`);
// ---- NOTE-FOR-DEBUGGING ---- const samplingIndexArrayGenerator =
// ---- NOTE-FOR-DEBUGGING ---- bootstrapSamplerKeyMap.sampleInstances();
// ---- NOTE-FOR-DEBUGGING ---- for (const element of samplingIndexArrayGenerator) {
// ---- NOTE-FOR-DEBUGGING ---- Utility.debuggingLog(`element of samplingIndexArrayGenerator=` +
// ---- NOTE-FOR-DEBUGGING ---- `${element}`);
// ---- NOTE-FOR-DEBUGGING ---- }
const samplingIndexArray: number[] =
[...bootstrapSamplerKeyMap.sampleInstances()];
Utility.debuggingLog(`samplingIndexArray.length=` +
`${samplingIndexArray.length}`);
const luDataBootstrapSampled: Data =
await luData.createDataFromSamplingExistingDataUtterances(
luData,
-1, // ---- NOTE-NO-NEED-FOR-LuData ---- labelColumnIndex,
-1, // ---- NOTE-NO-NEED-FOR-LuData ---- textColumnIndex,
-1, // ---- NOTE-NO-NEED-FOR-LuData ---- linesToSkip,
samplingIndexArray,
false);
luData = luDataBootstrapSampled as LuData;
}
// -----------------------------------------------------------------------
const results =
luData.collectSmallUtteranceIndexSetCoveringAllIntentEntityLabels();
const smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels: Map<string, Set<number>> =
results.smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels;
const smallUtteranceIndexEntityTypeMapCoveringAllIntentEntityLabels: Map<string, Set<number>> =
results.smallUtteranceIndexEntityTypeMapCoveringAllIntentEntityLabels;
const smallUtteranceIndexSetCoveringAllIntentEntityLabels: Set<number> =
results.smallUtteranceIndexSetCoveringAllIntentEntityLabels;
const remainingUtteranceIndexSet: Set<number> =
results.remainingUtteranceIndexSet;
Utility.debuggingLog(`smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels=` +
`${Utility.stringMapSetToJson(smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels)}`);
Utility.debuggingLog(`smallUtteranceIndexEntityTypeMapCoveringAllIntentEntityLabels=` +
`${Utility.stringMapSetToJson(smallUtteranceIndexEntityTypeMapCoveringAllIntentEntityLabels)}`);
Utility.debuggingLog(`smallUtteranceIndexSetCoveringAllIntentEntityLabels=` +
`${Utility.setToJsonSerialization(smallUtteranceIndexSetCoveringAllIntentEntityLabels)}`);
Utility.debuggingLog(`remainingUtteranceIndexSet=` +
`${Utility.setToJsonSerialization(remainingUtteranceIndexSet)}`);
Utility.debuggingLog(`smallUtteranceIndexSetCoveringAllIntentEntityLabels.size=` +
`${smallUtteranceIndexSetCoveringAllIntentEntityLabels.size}`);
Utility.debuggingLog(`remainingUtteranceIndexSet.size=` +
`${remainingUtteranceIndexSet.size}`);
// -------------------------------------------------------------------
if (!doAutoActiveLearning) {
aalLimitInitialNumberOfInstancesPerCategory = -1;
}
const resultsInitialSampling: {
"seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels": Map<string, Set<number>>,
"candidateUtteranceIndexSetSampled": Set<number>,
"candidateUtteranceIndexSetRemaining": Set<number>,
} = luData.collectUtteranceIndexSetSeedingIntentTrainingSet(
smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels,
remainingUtteranceIndexSet,
aalLimitInitialNumberOfInstancesPerCategory);
const seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels: Map<string, Set<number>> =
resultsInitialSampling.seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels;
const candidateUtteranceIndexSetSampled: Set<number> =
resultsInitialSampling.candidateUtteranceIndexSetSampled;
const candidateUtteranceIndexSetRemaining: Set<number> =
resultsInitialSampling.candidateUtteranceIndexSetRemaining;
Utility.debuggingLog(`seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels=` +
`${Utility.stringMapSetToJson(seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels)}`);
Utility.debuggingLog(`candidateUtteranceIndexSetSampled=` +
`${Utility.setToJsonSerialization(candidateUtteranceIndexSetSampled)}`);
Utility.debuggingLog(`candidateUtteranceIndexSetRemaining=` +
`${Utility.setToJsonSerialization(candidateUtteranceIndexSetRemaining)}`);
Utility.debuggingLog(`candidateUtteranceIndexSetSampled.size=` +
`${candidateUtteranceIndexSetSampled.size}`);
Utility.debuggingLog(`candidateUtteranceIndexSetRemaining.size=` +
`${candidateUtteranceIndexSetRemaining.size}`);
const countSeedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels: number =
[...seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels].reduce(
(accumulation: number, entry: [string, Set<number>]) =>
accumulation + entry[1].size, 0);
Utility.debuggingLog(`countSeedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels=` +
`${countSeedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels}`);
// -------------------------------------------------------------------
const seedingUtteranceIndexArray: number[] =
[...seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels].reduce(
(accumulation: number[], entry: [string, Set<number>]) =>
accumulation.concat(Array.from(entry[1])), []);
Utility.debuggingLog(`seedingUtteranceIndexArray.length=` +
`${seedingUtteranceIndexArray.length}`);
// -------------------------------------------------------------------
const seedingInstanceIndexArray: number[] =
Utility.cloneArray(seedingUtteranceIndexArray);
const intentLabelIndexArray: number[] =
luData.getIntentLabelIndexArray();
const utteranceFeatureIndexArrays: number[][] =
luData.getUtteranceFeatureIndexArrays();
const autoActiveLearner: AutoActiveLearner =
new AutoActiveLearner(
doAutoActiveLearning,
aalLimitInitialNumberOfInstancesPerCategory,
aalNumberOfInstancesPerIteration,
aalInstanceSelectionThreshold,
learnerParameterEpochs,
learnerParameterMiniBatchSize,
learnerParameterL1Regularization,
learnerParameterL2Regularization,
learnerParameterLossEarlyStopRatio,
learnerParameterLearningRate,
learnerParameterToCalculateOverallLossAfterEpoch);
const learned: {
"seedingInstanceIndexArray": number[],
"learner": SoftmaxRegressionSparse,
} = autoActiveLearner.learn(
luData.getFeaturizerLabels(),
luData.getFeaturizerLabelMap(),
luData.getFeaturizer().getNumberLabels(),
luData.getFeaturizer().getNumberFeatures(),
intentLabelIndexArray,
utteranceFeatureIndexArrays,
seedingInstanceIndexArray,
Array.from(candidateUtteranceIndexSetRemaining));
let aalSampledInstanceIndexArray: number[] =
learned.seedingInstanceIndexArray;
const learner: SoftmaxRegressionSparse =
learned.learner;
// -----------------------------------------------------------------------
const numberInstancesPreSelected: number =
seedingUtteranceIndexArray.length;
if (limitingSampleSize > numberInstancesPreSelected) {
limitingSampleSize -= numberInstancesPreSelected;
const reservoirArraySampler: ReservoirArraySampler<number> = new ReservoirArraySampler(
aalSampledInstanceIndexArray,
numberInstancesPreSelected);
aalSampledInstanceIndexArray =
[...reservoirArraySampler.sampleInstances(limitingSampleSize)];
}
// -----------------------------------------------------------------------
const newLuData: LuData = await LuData.createLuDataFromFilteringExistingLuDataUtterances(
luData,
new Set<number>(aalSampledInstanceIndexArray),
false);
return {
newLuData,
learner,
seedingInstanceIndexArray: aalSampledInstanceIndexArray,
seedingInstanceIndexArrayInitial: seedingUtteranceIndexArray };
// -----------------------------------------------------------------------
}
/**
* This function can read a TSV columnar file with label and text data and run through 3 steps of
* sampling processes:
* 0) Bootstrap Resampling
* 1) Auto Active Learning Sampling
* 2) Stratified Sampling
*
* @param columnarContent - content of a TSV columnar file in string form as input.
* @param labelColumnIndex - label/intent column index.
* @param textColumnIndex - text/utterace column index.
* @param linesToSkip - number of header lines skipped before processing each line as an instance record.
* @param doBootstrapResampling - boolean flag to activate bootstrap resampling (BRS) logic or not.
* @param brsDistribution - explicit distribution to control bootstrap resampling process
* @param doAutoActiveLearning - boolean flag to activate auto active leaning (AAL) process or not.
* @param aalLimitInitialNumberOfInstancesPerCategory - AAL initial number of instances per category/label/intent.
* @param aalNumberOfInstancesPerIteration - AAL number of instances processed per iterations.
* @param aalInstanceSelectionThreshold - AAL threshold to pick a tested instance for training in next iteration
* @param learnerParameterEpochs - AAL Softmax Regression learner parameter - number of epochs.
* @param learnerParameterMiniBatchSize - AAL Softmax Regression learner parameter - mini-batch size.
* @param learnerParameterL1Regularization - AAL Softmax Regression learner parameter - L1 regularization.
* @param learnerParameterL2Regularization - AAL Softmax Regression learner parameter - L2 regularization.
* @param learnerParameterLossEarlyStopRatio - AAL Softmax Regression learner parameter - early stop ratio.
* @param learnerParameterLearningRate - AAL Softmax Regression learner parameter - learning rate.
* @param learnerParameterToCalculateOverallLossAfterEpoch - AAL Softmax Regression learner parameter - flag
* @param limitingSampleSize - sample size controled by a final stratified sampling process.
*/
public static async mainAutoActiveLearnerWithColumnarContent(
columnarContent: string,
labelColumnIndex: number,
textColumnIndex: number,
linesToSkip: number,
doBootstrapResampling: boolean =
AppAutoActiveLearner.defaultDoBootstrapResampling,
brsDistribution: TMapStringKeyGenericValue<number> =
DictionaryMapUtility.newTMapStringKeyGenericValue<number>(),
doAutoActiveLearning: boolean =
AutoActiveLearner.defaultDoAutoActiveLearning,
aalLimitInitialNumberOfInstancesPerCategory: number =
AutoActiveLearner.defaultAalLimitInitialNumberOfInstancesPerCategory,
aalNumberOfInstancesPerIteration: number =
AutoActiveLearner.defaultAalNumberOfInstancesPerIteration,
aalInstanceSelectionThreshold: number =
AutoActiveLearner.defaultAalInstanceSelectionThreshold,
learnerParameterEpochs: number =
AppSoftmaxRegressionSparse.defaultEpochs,
learnerParameterMiniBatchSize: number =
AppSoftmaxRegressionSparse.defaultMiniBatchSize,
learnerParameterL1Regularization: number =
AppSoftmaxRegressionSparse.defaultL1Regularization,
learnerParameterL2Regularization: number =
AppSoftmaxRegressionSparse.defaultL2Regularization,
learnerParameterLossEarlyStopRatio: number =
AppSoftmaxRegressionSparse.defaultLossEarlyStopRatio,
learnerParameterLearningRate: number =
AppSoftmaxRegressionSparse.defaultLearningRate,
learnerParameterToCalculateOverallLossAfterEpoch: boolean =
true,
limitingSampleSize: number =
DefaultLimitingSampleSize): Promise<{
"newColumnarData": ColumnarData,
"learner": SoftmaxRegressionSparse,
"seedingInstanceIndexArray": number[],
"seedingInstanceIndexArrayInitial": number[],
}> {
// -----------------------------------------------------------------------
let columnarData: ColumnarData =
ColumnarData.createColumnarData(
columnarContent,
new NgramSubwordFeaturizer(),
labelColumnIndex,
textColumnIndex,
linesToSkip,
true);
// -----------------------------------------------------------------------
if (doBootstrapResampling) {
const bootstrapSamplerKeyMap: BootstrapSamplerKeyMapDistribution<number> =
new BootstrapSamplerKeyMapDistribution<number>(
brsDistribution,
columnarData.getIntentInstanceIndexMapArray());
// ---- NOTE-FOR-REFERENCE ---- const bootstrapSamplerKeyMap: BootstrapSamplerKeyMap<number> =
// ---- NOTE-FOR-REFERENCE ---- new BootstrapSamplerKeyMap(data.getIntentInstanceIndexMapArray());
Utility.debuggingLog(`columnarData.getIntentInstanceIndexMapArray()=` +
`${Utility.mapToJsonSerialization(columnarData.getIntentInstanceIndexMapArray())}`);
Utility.debuggingLog(`bootstrapSamplerKeyMap.computeSamplingNumberInstancesPerLabel()=` +
`${bootstrapSamplerKeyMap.computeSamplingNumberInstancesPerLabel()}`);
// ---- NOTE-FOR-DEBUGGING ---- const samplingIndexArrayGenerator =
// ---- NOTE-FOR-DEBUGGING ---- bootstrapSamplerKeyMap.sampleInstances();
// ---- NOTE-FOR-DEBUGGING ---- for (const element of samplingIndexArrayGenerator) {
// ---- NOTE-FOR-DEBUGGING ---- Utility.debuggingLog(`element of samplingIndexArrayGenerator=` +
// ---- NOTE-FOR-DEBUGGING ---- `${element}`);
// ---- NOTE-FOR-DEBUGGING ---- }
const samplingIndexArray: number[] =
[...bootstrapSamplerKeyMap.sampleInstances()];
Utility.debuggingLog(`samplingIndexArray.length=` +
`${samplingIndexArray.length}`);
const columnarDataBootstrapSampled: Data =
await columnarData.createDataFromSamplingExistingDataUtterances(
columnarData,
labelColumnIndex,
textColumnIndex,
linesToSkip,
samplingIndexArray,
false);
columnarData = columnarDataBootstrapSampled as ColumnarData;
}
// -----------------------------------------------------------------------
const results =
columnarData.collectSmallUtteranceIndexSetCoveringAllIntentEntityLabels();
const smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels: Map<string, Set<number>> =
results.smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels;
const smallUtteranceIndexEntityTypeMapCoveringAllIntentEntityLabels: Map<string, Set<number>> =
results.smallUtteranceIndexEntityTypeMapCoveringAllIntentEntityLabels;
const smallUtteranceIndexSetCoveringAllIntentEntityLabels: Set<number> =
results.smallUtteranceIndexSetCoveringAllIntentEntityLabels;
const remainingUtteranceIndexSet: Set<number> =
results.remainingUtteranceIndexSet;
Utility.debuggingLog(`smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels=` +
`${Utility.stringMapSetToJson(smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels)}`);
Utility.debuggingLog(`smallUtteranceIndexEntityTypeMapCoveringAllIntentEntityLabels=` +
`${Utility.stringMapSetToJson(smallUtteranceIndexEntityTypeMapCoveringAllIntentEntityLabels)}`);
Utility.debuggingLog(`smallUtteranceIndexSetCoveringAllIntentEntityLabels=` +
`${Utility.setToJsonSerialization(smallUtteranceIndexSetCoveringAllIntentEntityLabels)}`);
Utility.debuggingLog(`remainingUtteranceIndexSet=` +
`${Utility.setToJsonSerialization(remainingUtteranceIndexSet)}`);
Utility.debuggingLog(`smallUtteranceIndexSetCoveringAllIntentEntityLabels.size=` +
`${smallUtteranceIndexSetCoveringAllIntentEntityLabels.size}`);
Utility.debuggingLog(`remainingUtteranceIndexSet.size=` +
`${remainingUtteranceIndexSet.size}`);
// -------------------------------------------------------------------
if (!doAutoActiveLearning) {
aalLimitInitialNumberOfInstancesPerCategory = -1;
}
const resultsInitialSampling: {
"seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels": Map<string, Set<number>>,
"candidateUtteranceIndexSetSampled": Set<number>,
"candidateUtteranceIndexSetRemaining": Set<number>,
} = columnarData.collectUtteranceIndexSetSeedingIntentTrainingSet(
smallUtteranceIndexIntentMapCoveringAllIntentEntityLabels,
remainingUtteranceIndexSet,
aalLimitInitialNumberOfInstancesPerCategory);
const seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels: Map<string, Set<number>> =
resultsInitialSampling.seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels;
const candidateUtteranceIndexSetSampled: Set<number> =
resultsInitialSampling.candidateUtteranceIndexSetSampled;
const candidateUtteranceIndexSetRemaining: Set<number> =
resultsInitialSampling.candidateUtteranceIndexSetRemaining;
Utility.debuggingLog(`seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels=` +
`${Utility.stringMapSetToJson(seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels)}`);
Utility.debuggingLog(`candidateUtteranceIndexSetSampled=` +
`${Utility.setToJsonSerialization(candidateUtteranceIndexSetSampled)}`);
Utility.debuggingLog(`candidateUtteranceIndexSetRemaining=` +
`${Utility.setToJsonSerialization(candidateUtteranceIndexSetRemaining)}`);
Utility.debuggingLog(`candidateUtteranceIndexSetSampled.size=` +
`${candidateUtteranceIndexSetSampled.size}`);
Utility.debuggingLog(`candidateUtteranceIndexSetRemaining.size=` +
`${candidateUtteranceIndexSetRemaining.size}`);
const countSeedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels: number =
[...seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels].reduce(
(accumulation: number, entry: [string, Set<number>]) =>
accumulation + entry[1].size, 0);
Utility.debuggingLog(`countSeedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels=` +
`${countSeedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels}`);
// -------------------------------------------------------------------
const seedingUtteranceIndexArray: number[] =
[...seedingUtteranceIndexIntentMapCoveringAllIntentEntityLabels].reduce(
(accumulation: number[], entry: [string, Set<number>]) =>
accumulation.concat(Array.from(entry[1])), []);
Utility.debuggingLog(`seedingUtteranceIndexArray.length=` +
`${seedingUtteranceIndexArray.length}`);
// -------------------------------------------------------------------
const seedingInstanceIndexArray: number[] =
Utility.cloneArray(seedingUtteranceIndexArray);
const intentLabelIndexArray: number[] =
columnarData.getIntentLabelIndexArray();
const utteranceFeatureIndexArrays: number[][] =
columnarData.getUtteranceFeatureIndexArrays();
const autoActiveLearner: AutoActiveLearner =
new AutoActiveLearner(
doAutoActiveLearning,
aalLimitInitialNumberOfInstancesPerCategory,
aalNumberOfInstancesPerIteration,
aalInstanceSelectionThreshold,
learnerParameterEpochs,
learnerParameterMiniBatchSize,
learnerParameterL1Regularization,
learnerParameterL2Regularization,
learnerParameterLossEarlyStopRatio,
learnerParameterLearningRate,
learnerParameterToCalculateOverallLossAfterEpoch);
const learned: {
"seedingInstanceIndexArray": number[],
"learner": SoftmaxRegressionSparse,
} = autoActiveLearner.learn(
columnarData.getFeaturizerLabels(),
columnarData.getFeaturizerLabelMap(),
columnarData.getFeaturizer().getNumberLabels(),
columnarData.getFeaturizer().getNumberFeatures(),
intentLabelIndexArray,
utteranceFeatureIndexArrays,
seedingInstanceIndexArray,
Array.from(candidateUtteranceIndexSetRemaining));
let aalSampledInstanceIndexArray: number[] =
learned.seedingInstanceIndexArray;
const learner: SoftmaxRegressionSparse =
learned.learner;
// -----------------------------------------------------------------------
const numberInstancesPreSelected: number =
seedingUtteranceIndexArray.length;
if (limitingSampleSize > numberInstancesPreSelected) {
limitingSampleSize -= numberInstancesPreSelected;
const reservoirArraySampler: ReservoirArraySampler<number> = new ReservoirArraySampler(
aalSampledInstanceIndexArray,
numberInstancesPreSelected);
aalSampledInstanceIndexArray =
[...reservoirArraySampler.sampleInstances(limitingSampleSize)];
}
// -----------------------------------------------------------------------
const newColumnarData: ColumnarData =
ColumnarData.createColumnarDataFromFilteringExistingColumnarDataUtterances(
columnarData,
labelColumnIndex,
textColumnIndex,
linesToSkip,
new Set<number>(aalSampledInstanceIndexArray),
false);
return {
newColumnarData,
learner,
seedingInstanceIndexArray: aalSampledInstanceIndexArray,
seedingInstanceIndexArrayInitial: seedingUtteranceIndexArray };
// -----------------------------------------------------------------------
}
In "src/model/evaluation/cross_validation/AppCrossValidator.ts", there are some example functions that
demonstrates how to use 'bf-dispatcher' to run cross validation and evaluate model performance.
/**
* This function consumes a LU file content as input and run cross validation (CV) to evaluate models trained from
* the input label/text (intent/utterance) instance set.
*
* @param luContent - input LU file content as input.
* @param numberOfCrossValidationFolds - number of cross validation (CV) folds.
* @param learnerParameterEpochs - CV Softmax Regression Learner parameter - number of epochs
* @param learnerParameterMiniBatchSize - CV Softmax Regression learner parameter - mini-batch size.
* @param learnerParameterL1Regularization - CV Softmax Regression learner parameter - L1 regularization.
* @param learnerParameterL2Regularization - CV Softmax Regression learner parameter - L2 regularization.
* @param learnerParameterLossEarlyStopRatio - CV Softmax Regression learner parameter - early stop ratio.
* @param learnerParameterLearningRate - CV Softmax Regression learner parameter - learning rate.
* @param learnerParameterToCalculateOverallLossAfterEpoch - CV Softmax Regression learner parameter - flag
*/
export async function mainCrossValidatorWithLuContent(
luContent: string,
numberOfCrossValidationFolds: number =
CrossValidator.defaultNumberOfCrossValidationFolds,
learnerParameterEpochs: number =
AppSoftmaxRegressionSparse.defaultEpochs,
learnerParameterMiniBatchSize: number =
AppSoftmaxRegressionSparse.defaultMiniBatchSize,
learnerParameterL1Regularization: number =
AppSoftmaxRegressionSparse.defaultL1Regularization,
learnerParameterL2Regularization: number =
AppSoftmaxRegressionSparse.defaultL2Regularization,
learnerParameterLossEarlyStopRatio: number =
AppSoftmaxRegressionSparse.defaultLossEarlyStopRatio,
learnerParameterLearningRate: number =
AppSoftmaxRegressionSparse.defaultLearningRate,
learnerParameterToCalculateOverallLossAfterEpoch: boolean =
true): Promise<CrossValidator> {
// -----------------------------------------------------------------------
const luData: LuData =
await LuData.createLuData(
luContent,
new NgramSubwordFeaturizer(),
true);
// -----------------------------------------------------------------------
if (!numberOfCrossValidationFolds) {
numberOfCrossValidationFolds = CrossValidator.defaultNumberOfCrossValidationFolds;
}
// -------------------------------------------------------------------
const intents: string[] =
luData.getIntents();
const utterances: string[] =
luData.getUtterances();
const intentLabelIndexArray: number[] =
luData.getIntentLabelIndexArray();
const utteranceFeatureIndexArrays: number[][] =
luData.getUtteranceFeatureIndexArrays();
assert(intentLabelIndexArray, "intentLabelIndexArray is undefined.");
assert(utteranceFeatureIndexArrays, "utteranceFeatureIndexArrays is undefined.");
const crossValidator: CrossValidator =
new CrossValidator(
luData.getFeaturizerLabels(),
luData.getFeaturizerLabelMap(),
luData.getFeaturizer().getNumberLabels(),
luData.getFeaturizer().getNumberFeatures(),
intents,
utterances,
intentLabelIndexArray,
utteranceFeatureIndexArrays,
luData.getIntentInstanceIndexMapArray(),
numberOfCrossValidationFolds,
learnerParameterEpochs,
learnerParameterMiniBatchSize,
learnerParameterL1Regularization,
learnerParameterL2Regularization,
learnerParameterLossEarlyStopRatio,
learnerParameterLearningRate,
learnerParameterToCalculateOverallLossAfterEpoch);
return crossValidator;
// -----------------------------------------------------------------------
}
/**
* This function consumes a columnar TSV file content as input and run cross validation (CV) to
* evaluate models trained from the input label/text (intent/utterance) instance set.
*
* @param columnarContent - content of a TSV columnar file in string form as input.
* @param labelColumnIndex - label/intent column index.
* @param textColumnIndex - text/utterace column index.
* @param linesToSkip - number of header lines skipped before processing each line as an instance record.
* @param numberOfCrossValidationFolds - number of cross validation (CV) folds.
* @param learnerParameterEpochs - CV Softmax Regression Learner parameter - number of epochs
* @param learnerParameterMiniBatchSize - CV Softmax Regression learner parameter - mini-batch size.
* @param learnerParameterL1Regularization - CV Softmax Regression learner parameter - L1 regularization.
* @param learnerParameterL2Regularization - CV Softmax Regression learner parameter - L2 regularization.
* @param learnerParameterLossEarlyStopRatio - CV Softmax Regression learner parameter - early stop ratio.
* @param learnerParameterLearningRate - CV Softmax Regression learner parameter - learning rate.
* @param learnerParameterToCalculateOverallLossAfterEpoch - CV Softmax Regression learner parameter - flag
*/
export function mainCrossValidatorWithColumnarContent(
columnarContent: string,
labelColumnIndex: number,
textColumnIndex: number,
linesToSkip: number,
numberOfCrossValidationFolds: number =
CrossValidator.defaultNumberOfCrossValidationFolds,
learnerParameterEpochs: number =
AppSoftmaxRegressionSparse.defaultEpochs,
learnerParameterMiniBatchSize: number =
AppSoftmaxRegressionSparse.defaultMiniBatchSize,
learnerParameterL1Regularization: number =
AppSoftmaxRegressionSparse.defaultL1Regularization,
learnerParameterL2Regularization: number =
AppSoftmaxRegressionSparse.defaultL2Regularization,
learnerParameterLossEarlyStopRatio: number =
AppSoftmaxRegressionSparse.defaultLossEarlyStopRatio,
learnerParameterLearningRate: number =
AppSoftmaxRegressionSparse.defaultLearningRate,
learnerParameterToCalculateOverallLossAfterEpoch: boolean =
true): CrossValidator {
// -----------------------------------------------------------------------
const columnarData: ColumnarData =
ColumnarData.createColumnarData(
columnarContent,
new NgramSubwordFeaturizer(),
labelColumnIndex,
textColumnIndex,
linesToSkip,
true);
// -----------------------------------------------------------------------
if (!numberOfCrossValidationFolds) {
numberOfCrossValidationFolds = CrossValidator.defaultNumberOfCrossValidationFolds;
}
// -------------------------------------------------------------------
const intents: string[] =
columnarData.getIntents();
const utterances: string[] =
columnarData.getUtterances();
const intentLabelIndexArray: number[] =
columnarData.getIntentLabelIndexArray();
const utteranceFeatureIndexArrays: number[][] =
columnarData.getUtteranceFeatureIndexArrays();
assert(intentLabelIndexArray, "intentLabelIndexArray is undefined.");
assert(utteranceFeatureIndexArrays, "utteranceFeatureIndexArrays is undefined.");
const crossValidator: CrossValidator =
new CrossValidator(
columnarData.getFeaturizerLabels(),
columnarData.getFeaturizerLabelMap(),
columnarData.getFeaturizer().getNumberLabels(),
columnarData.getFeaturizer().getNumberFeatures(),
intents,
utterances,
intentLabelIndexArray,
utteranceFeatureIndexArrays,
columnarData.getIntentInstanceIndexMapArray(),
numberOfCrossValidationFolds,
learnerParameterEpochs,
learnerParameterMiniBatchSize,
learnerParameterL1Regularization,
learnerParameterL2Regularization,
learnerParameterLossEarlyStopRatio,
learnerParameterLearningRate,
learnerParameterToCalculateOverallLossAfterEpoch);
return crossValidator;
// -----------------------------------------------------------------------
}
At last, 'bf-dispatcher' also contains several model performance reporter classes in "src/model/evaluation/report" For now, there are three report classes:
a) DataProfileReporter.ts: consume a data file and report label distribution and feature distribution. Some example functions are implemented in AppDataProfileReporter.ts.
b) ModelMetaDataProfileReporter.ts: load a model previously trained and generated and report its parameters. Some example functions are implemented in AppModelMetaDataProfileReporter.ts.
c) ThresholdReporter.ts: load a model and a test file, then report the model performance. Some example functions are implemented in AppThresholdReporter.ts.