如何在浏览器（客户端）中使用 ONNX 运行 u2net 模型？答案

【问题标题】：How to run u2net model with ONNX in browser (client side)?如何在浏览器（客户端）中使用 ONNX 运行 u2net 模型？
【发布时间】：2021-11-26 04:17:21
【问题描述】：

我正在尝试在浏览器中运行u2net 模型，我已将pytorch u2netp model 转换为ONNX model 并编写了以下代码来运行它，但结果很差。我遵循与 python 脚本相同的预处理步骤，但没有得到结果。我无法找到 onnx 函数来执行预处理，所以我使用 for 循环来更改每个通道中的值。

<!DOCTYPE html>
<html>
    <header>
        <title>ONNX Runtime JavaScript examples: Quick Start - Web (using script tag)</title>
        <input id="image-selector" type="file" style="top:10px;left:10px" >
        <button id="predict-button" class="btn btn-dark float-right" style="top:10px;left:70px" >Predict</button>
        <img id="selected-image"  src="" />
        <canvas id="canvas" width =320px height=320px ></canvas>
    </header>
    <body>
        <!-- import ONNXRuntime Web from CDN -->
        <script src="https://code.jquery.com/jquery-3.3.1.min.js"></script>
        <script src="https://cdn.jsdelivr.net/npm/onnxruntime-web/dist/ort.min.js"></script>
        <script>
      
            $("#image-selector").change(function () {
                let reader = new FileReader();
                reader.onload = function () {
                    let dataURL = reader.result;
                    $("#selected-image").attr("src", dataURL);
                }
                let file = $("#image-selector").prop("files")[0];
                reader.readAsDataURL(file);
            }); 
            
            $("#predict-button").click(async function (){
                
                    const canvas = document.getElementById("canvas");
                    const ctx = canvas.getContext("2d");


                    const session = await ort.InferenceSession.create('./u2netp.onnx').then(console.log("model loaded"));
                    
                    const inputNames = session.inputNames;
                    const outputNames = session.outputNames;
                    console.log('inputNames', inputNames)
                    console.log('outputNames', outputNames)
                   
                    let image = $("#selected-image").get(0);
    
                    console.log("image.naturalHeight", image.naturalHeight)
                    console.log("image.naturalWidth", image.naturalWidth)
                    
                    var oc = document.createElement('canvas'),
                    octx = oc.getContext('2d');
                    oc.width = 320;
                    oc.height = 320;
                    octx.drawImage(image, 0, 0, oc.width, oc.height);
                    var input_imageData = octx.getImageData(0, 0, 320, 320);
                  
                    var floatArr =  new Float32Array(320 * 320 * 3)
                    
                    var j = 0
                    for (let i = 1; i < input_imageData.data.length+1; i ++) {
                        if(i % 4 != 0){
                            floatArr[j] = (input_imageData.data[i-1].toFixed(2))/255;  // red   color
                            j = j + 1;
                        } 
                    } 
                    console.log("floatArr1", floatArr)
                    for (let i = 1; i < floatArr.length+1; i += 3) {
                        floatArr[i-1] = (floatArr[i-1] - 0.485)/0.229  // red   color
                        floatArr[i] = (floatArr[i] - 0.456)/0.224  // green color
                        floatArr[i+1] = (floatArr[i+1] - 0.406)/0.225  // blue  color
                    } 

                    console.log("floatArr2", floatArr)
                   
                    const input = new ort.Tensor('float32', floatArr,  [1, 3, 320, 320])
  
                    a = inputNames[0]
                    console.log("a", a)
                    const feeds = {"input.1": input};
                    console.log("feeds", feeds)
                    const results = await session.run(feeds).then();
                    const pred = Object.values(results)[0]
                    console.log('pred', pred)
                    
                    console.log('pred.data.length', pred.data.length)
                    console.log('pred.data[0]', Math.round(pred.data[0]*255))
                    
                    var myImageData = ctx.createImageData(320, 320);
                    for (let i = 0; i < pred.data.length*4; i += 4) {
                        var pixelIndex = i;
                        if(i != 0){
                            t = i/4;
                        }
                        else{
                            t = 0;
                        }
                        myImageData.data[pixelIndex    ] = Math.round(pred.data[t]*255);  // red   color
                        myImageData.data[pixelIndex + 1] = Math.round(pred.data[t]*255);  // green color
                        myImageData.data[pixelIndex + 2] = Math.round(pred.data[t]*255);  // blue  color
                        myImageData.data[pixelIndex + 3] = 255;
                    } 
                    ctx.putImageData(myImageData, 10, 10);
                    console.log("myImageData", myImageData)

                });
        </script>
    </body>
</html>

【问题讨论】：

标签： javascript deep-learning onnxruntime

【解决方案1】：

在将图像数据数组传递给模型之前需要对其进行转置。

<!DOCTYPE html>
<html>
    <header>
        <title>ONNX Runtime JavaScript examples: Quick Start - Web (using script tag)</title>
        <input id="image-selector" type="file" style="top:10px;left:10px" >
        <button id="predict-button" class="btn btn-dark float-right" style="top:10px;left:70px" >Predict</button>
        <img id="selected-image"  src="" />
        <canvas id="canvas" width =320px height=320px ></canvas>
    </header>
    <body>
        <!-- import ONNXRuntime Web from CDN -->
        <script src="https://code.jquery.com/jquery-3.3.1.min.js"></script>
        <script src="https://cdn.jsdelivr.net/npm/onnxruntime-web/dist/ort.min.js"></script>
        <script>
            $("#image-selector").change(function () {
                let reader = new FileReader();
                reader.onload = function () {
                    let dataURL = reader.result;
                    $("#selected-image").attr("src", dataURL);
                }
                let file = $("#image-selector").prop("files")[0];
                reader.readAsDataURL(file);
            }); 

            // async function main() {
            $("#predict-button").click(async function (){
                    const canvas = document.getElementById("canvas");
                    const ctx = canvas.getContext("2d");


                    const session = await ort.InferenceSession.create('./u2netp.onnx').then(console.log("model loaded"));
                    // input_name = getInputs();
                    const inputNames = session.inputNames;
                    const outputNames = session.outputNames;
                    let image = $("#selected-image").get(0);
                    var oc = document.createElement('canvas'),
                    octx = oc.getContext('2d');
                    oc.width = 320;
                    oc.height = 320;
                    octx.drawImage(image, 0, 0, oc.width, oc.height);
                    var input_imageData = octx.getImageData(0, 0, 320, 320);
                    console.log("input_imageData",input_imageData.data)
                    var floatArr =  new Float32Array(320 * 320 * 3)
                    var floatArr1 =  new Float32Array(320 * 320 * 3)
                    var floatArr2 =  new Float32Array(320 * 320 * 3)
                    
                    var j = 0
                    for (let i = 1; i < input_imageData.data.length+1; i ++) {
                        if(i % 4 != 0){
                            floatArr[j] = (input_imageData.data[i-1].toFixed(2))/255;  // red   color
                            j = j + 1;
                        } 
                    } 
                    console.log("floatArr", floatArr)
                    for (let i = 1; i < floatArr.length+1; i += 3) {
                        floatArr1[i-1] = (floatArr[i-1] - 0.485)/0.229  // red   color
                        floatArr1[i] = (floatArr[i] - 0.456)/0.224  // green color
                        floatArr1[i+1] = (floatArr[i+1] - 0.406)/0.225  // blue  color
                    } 
                    var k = 0
                    for (let i = 0; i < floatArr.length; i += 3) {
                        floatArr2[k] = floatArr[i]  // red   color
                        k = k + 1
                    } 
                    console.log("k", k)
                    var l = 102400 
                    for (let i = 1; i < floatArr.length; i += 3) {
                        floatArr2[l] = floatArr[i]  // red   color
                        l = l + 1
                    } 
                    console.log("l", l)
                    var m = 204800
                    for (let i = 2; i < floatArr.length; i += 3) {
                        floatArr2[m] = floatArr[i]  // red   color
                        m = m + 1
                    } 
                    const input = new ort.Tensor('float32', floatArr2,  [1, 3, 320, 320])
                    a = inputNames[0]
                    const feeds = {"input.1": input};
                    const results = await session.run(feeds).then();
                    const pred = Object.values(results)[0]
                    var myImageData = ctx.createImageData(320, 320);
                    for (let i = 0; i < pred.data.length*4; i += 4) {
                        var pixelIndex = i;
                        if(i != 0){
                            t = i/4;
                        }
                        else{
                            t = 0;
                        }
                        myImageData.data[pixelIndex    ] = Math.round(pred.data[t]*255);  // red   color
                        myImageData.data[pixelIndex + 1] = Math.round(pred.data[t]*255);  // green color
                        myImageData.data[pixelIndex + 2] = Math.round(pred.data[t]*255);  // blue  color
                        myImageData.data[pixelIndex + 3] = 255;
                    } 
                    ctx.putImageData(myImageData, 10, 10);
                });
        </script>
    </body>
</html>

【讨论】：