ppljs-ppl-core/src/modelLoader.ts

ppljs network inference framework core module

import {ModelConfig,Model} from './interface/interface'

/*
*@author xusiyu@sensetime.com
*@remark asyn parse will be implement later;
*@breif  model includes model structure json files and model data(weight,bias...) binary file,
*        so we shoud parse the json model and read the model data
*@process
*@1.determine the path of model file and data file.we do this work in constructor()
*@2.parse the model json file according to the path.we do this work in parseJsonModel().
*@3.read the model data according to the path.we do this word in readModelData()
*/

export default  class ModelLoader {
    modelConfig_: ModelConfig;
    isLocalPath:  boolean;
    //模型路径可以为网络地址or本地，本地文件仅供测试使用
    constructor(modelConfig:ModelConfig){
        this.modelConfig_ = modelConfig;
        //if ModelConfig don't provide model name,we use default name
        if(modelConfig.modelName == undefined)
            this.modelConfig_.modelName = "model.json"
        if(modelConfig.binaryDataName == undefined)
            this.modelConfig_.binaryDataName = "model.dat"
        if (modelConfig.modelPath.charAt(modelConfig.modelPath.length - 1) != '/') {
            this.modelConfig_.modelPath = `${modelConfig.modelPath}/`;
        }
        this.isLocalPath = modelConfig.modelPath.indexOf('http') != 0;    
    }

    async LoadModel():Promise<Model> {
        var p1 = this.ParseJsonModel();
        var p2 = this.ReadModelData();
        return new Promise((resolve)=>{
            Promise.all([p1,p2]).then(res =>{
                this.AssignData(res[0],res[1]);
                resolve(res[0]);
            });
        });
    }

    readData(modelPath:string){ 
        return new Promise((resolve, reject) => {
            fetch(modelPath, {
                method: 'get',
            }).then(response =>{ resolve(response.arrayBuffer());})//可以获取blob或者arrayBuffer
            .then(err => reject(err));
        });
    }

    ParseJsonModel():Promise<Model> {
        const modelPath = this.modelConfig_.modelPath+this.modelConfig_.modelName;
        return new Promise((resolve, reject) => {
            fetch(modelPath, {
                method: 'get',
            }).then(response =>{resolve(response.json());})
            .then(err => reject(err));
        });
    }

    ReadModelData():Promise<Float32Array> {
        const binaryPath = this.modelConfig_.modelPath+this.modelConfig_.binaryDataName;
        return new Promise((resolve, reject) => {
            this.readData(binaryPath).then(response=> {
                var array = new Float32Array(<ArrayBuffer>response);
                //console.log("Float32Array length is ",array.length);
                //console.log("Float32Array data  is ",array);
                resolve(array);
            }).catch(err => reject(err));
        });
    }
    
     AssignData(opModel_:Model,dataArray: Float32Array) {
        //we put the binary data in dataArray
        opModel_.ops.forEach((op)=>{
            if(op.data!=undefined){
                var offset:number = op.hostDataOffset;
                //通过offset去读取响应的数据
                for (var item in op.data) {
                    var size_: number = op.data[item].size;
                    //we will read size_ length data from offset                    
                    //slice是pos 而非byte
                    op.data[item].hostData = dataArray.slice(offset,offset+size_);
                    offset +=(size_);
                }
            }
        })
    }
};