import { LayersModel, loadLayersModel, Sequential, Tensor, } from '@tensorflow/tfjs-node' import { accuracyScore } from 'machinelearn/metrics' import { WordEmbeddingsManager } from './embeddings/word-embeddings-manager' import { SimpleNN } from './simple-nn' import { EncodedPrediction, InputSet, ModelParameters, ModelTemplatesType, OutputSet, TrainingParameters, WordEmbeddingsConfig, } from './types' export class ModelManager { protected constructor(readonly model: Sequential | LayersModel) {} static fromModel(model: Sequential | LayersModel): ModelManager { return new ModelManager(model) } static async fromModelPath(path: string): Promise { const model = await loadLayersModel(`file://${path}`) return new ModelManager(model) } static async fromModelTemplate( template: ModelTemplatesType, wordEmbeddingsConfig: WordEmbeddingsConfig, parameters: ModelParameters ): Promise { const wordEmbeddingsManager = await WordEmbeddingsManager.withConfig( wordEmbeddingsConfig ) const wordEmbeddingsMatrix = wordEmbeddingsManager.matrix switch (template) { case ModelTemplatesType.SIMPLE_NN: default: return new ModelManager( new SimpleNN(parameters, wordEmbeddingsMatrix).model ) } } async train(params: TrainingParameters): Promise { await this.model.fit(params.x, params.y, { epochs: params.epochs, batchSize: params.batchSize, validationSplit: params.validationSplit, }) } predictProbabilities(input: Tensor): EncodedPrediction[] { const confidences = Array.from( (this.model.predict(input) as Tensor).dataSync() ) return confidences.map((confidence: number, intentId: number) => ({ intentId, confidence, })) } predict(input: Tensor): number { const probabilities = (this.model.predict( input ) as Tensor).arraySync() as number[][] const intentId = probabilities[0].indexOf(Math.max(...probabilities[0])) return intentId } evaluate(x: InputSet, y: OutputSet): number { const prediction = (this.model.predict( x ) as Tensor).arraySync() as number[][] const yPred: number[] = prediction.map(confidences => confidences.indexOf(Math.max(...confidences)) ) const yTrue = y.arraySync() as number[] const accuracy = accuracyScore(yTrue, yPred) return accuracy } async saveModel(modelDir: string): Promise { await this.model.save(`file://${modelDir}`) } }