/* eslint-disable @typescript-eslint/unbound-method */ import { layers, Sequential, sequential, Tensor, train, } from '@tensorflow/tfjs-node' import { ModelParameters } from './types' export class SimpleNN { model: Sequential constructor( private parameters: ModelParameters, private wordEmbeddingsMatrix: Tensor ) { this._createModel() } private _createModel(): void { this.model = sequential() this._addEmbeddingLayer() this._addLSTMLayer() this._addDenseLayer() this._compile() this.model.summary() } private _addEmbeddingLayer(): void { this.model.add( layers.embedding({ inputDim: this.wordEmbeddingsMatrix.shape[0], outputDim: this.wordEmbeddingsMatrix.shape[1], inputLength: this.parameters.maxSeqLen, trainable: this.parameters.trainableEmbeddings, weights: [this.wordEmbeddingsMatrix], }) ) } private _addLSTMLayer(): void { this.model.add( layers.lstm({ units: 128, dropout: 0.3, recurrentDropout: 0.3, }) ) } private _addDenseLayer(): void { this.model.add( layers.dense({ units: this.parameters.intentsCount, activation: 'softmax', }) ) } private _compile(): void { this.model.compile({ optimizer: train.adam(this.parameters.learningRate), loss: 'sparseCategoricalCrossentropy', metrics: ['accuracy'], }) } }