"use strict";

import tf from '@tensorflow/tfjs-node-gpu';

class AITrainer {
	constructor({ settings, GatewayRepo, DatasetFetcher }) {
		this.settings = settings;
		this.dataset_fetcher = DatasetFetcher;
		this.repo_gateway = GatewayRepo;
		this.model = this.generate_model();
	}
	
	generate_model() {
		let model = tf.sequential();
		model.add(tf.layers.dense({
			units: 256, // 256 nodes
			activation: "sigmoid", // Sigmoid activation function
			inputShape: [3], // 2 inputs - lat and long
		}))
		model.add(tf.layers.dense({
			units: 1, // 1 output value - RSSI
			activation: "sigmoid" // The example code uses softmax, but this is generally best used for classification tasks
		}));
		
		model.compile({
			optimizer: tf.train.adam(),
			loss: "absoluteDifference",
			metrics: [ "accuracy", "meanSquaredError" ]
		});
		
		return model;
	}
	
	async train_all() {
		for(let gateway of this.repo_gateway.iterate()) {
			await this.train_dataset(gateway.id);
		}
	}
	
	async train_gateway(gateway_id) {
		let dataset_input = tf.data.generator(
			this.dataset_fetcher.fetch_input.bind(null, gateway_id)
		);
		let dataset_output = tf.data.generator(
			this.dataset_fetcher.fetch_output.bind(null, gateway_id)
		);
		
		let dataset = tf.data.zip({
				xs: dataset_input,
				ys: dataset_output
			}).shuffle(this.settings.ai.batch_size)
			.batch(this.settings.ai.batch_size);
		
		let result = await this.model.fitDataset(dataset, {
			epochs: this.settings.ai.epochs,
			batchSize:  this.settings.ai.batch_size
		});
		console.log(result);
	}
}

export default AITrainer;