UWP通过机器学习加载ONNX进行表情识别
首先我们先来说说这个ONNX
ONNX是一种针对机器学习所设计的开放式的文件格式,用于存储训练好的模型。它使得不同的人工智能框架(如Pytorch, MXNet)可以采用相同格式存储模型数据并交互。 ONNX的规范及代码主要由微软,亚马逊 ,Facebook 和 IBM 等公司共同开发,以开放源代码的方式托管在Github上。目前官方支持加载ONNX模型并进行推理的深度学习框架有: Caffe2, PyTorch, MXNet,ML.NET,TensorRT 和 Microsoft CNTK,并且 TensorFlow 也非官方的支持ONNX。---维基百科
看了上面的引用 大家应该知道了 这个其实是个文件格式用来存储训练好的模型,所以我这篇帖子既然是做表情识别那肯定是需要有个能识别表情的模型。有了这个模型我们就可以根据图片上的人物,进行表情的识别判断了。
刚好微软对机器学习这块也挺上心的,所以我也趁着疫情比较闲,就来学习学习了。UWP的机器学习的api微软已经切成正式了,所以大家可以放心使用。
这就是uwp api文档 开头就是AI的
我其实是个小白 所以我就直接拿官方的一个demo的简化版来进行讲解了,官方的demo演示如下。
这个app就是通过摄像头读取每一帧 进行和模型匹配得出结果的
下面是机器学习的微软的github地址
Emoji8的git地址
我今天要说的就是这个demo的简化代码大致运行流程
下面是项目结构图
我把官方项目简化了 所以只留下了识别后的文本移除了一些依赖的库
核心代码在IntelligenceService类里的Current_SoftwareBitmapFrameCaptured方法里
private async void Current_SoftwareBitmapFrameCaptured(object sender, SoftwareBitmapEventArgs e)
{
Debug.WriteLine("FrameCaptured");
Debug.WriteLine($"Frame evaluation started {DateTime.Now}" );
if (e.SoftwareBitmap != null)
{
BitmapPixelFormat bpf = e.SoftwareBitmap.BitmapPixelFormat;
var uncroppedBitmap = SoftwareBitmap.Convert(e.SoftwareBitmap, BitmapPixelFormat.Nv12);
var faces = await _faceDetector.DetectFacesAsync(uncroppedBitmap);
if (faces.Count > 0)
{
//crop image to focus on face portion
var faceBox = faces[0].FaceBox;
VideoFrame inputFrame = VideoFrame.CreateWithSoftwareBitmap(e.SoftwareBitmap);
VideoFrame tmp = null;
tmp = new VideoFrame(e.SoftwareBitmap.BitmapPixelFormat, (int)(faceBox.Width + faceBox.Width % 2) - 2, (int)(faceBox.Height + faceBox.Height % 2) - 2);
await inputFrame.CopyToAsync(tmp, faceBox, null);
//crop image to fit model input requirements
VideoFrame croppedInputImage = new VideoFrame(BitmapPixelFormat.Gray8, (int)_inputImageDescriptor.Shape[3], (int)_inputImageDescriptor.Shape[2]);
var srcBounds = GetCropBounds(
tmp.SoftwareBitmap.PixelWidth,
tmp.SoftwareBitmap.PixelHeight,
croppedInputImage.SoftwareBitmap.PixelWidth,
croppedInputImage.SoftwareBitmap.PixelHeight);
await tmp.CopyToAsync(croppedInputImage, srcBounds, null);
ImageFeatureValue imageTensor = ImageFeatureValue.CreateFromVideoFrame(croppedInputImage);
_binding = new LearningModelBinding(_session);
TensorFloat outputTensor = TensorFloat.Create(_outputTensorDescriptor.Shape);
List<float> _outputVariableList = new List<float>();
// Bind inputs + outputs
_binding.Bind(_inputImageDescriptor.Name, imageTensor);
_binding.Bind(_outputTensorDescriptor.Name, outputTensor);
// Evaluate results
var results = await _session.EvaluateAsync(_binding, new Guid().ToString());
Debug.WriteLine("ResultsEvaluated: " + results.ToString());
var outputTensorList = outputTensor.GetAsVectorView();
var resultsList = new List<float>(outputTensorList.Count);
for (int i = 0; i < outputTensorList.Count; i++)
{
resultsList.Add(outputTensorList[i]);
}
var softMaxexOutputs = SoftMax(resultsList);
double maxProb = 0;
int maxIndex = 0;
// Comb through the evaluation results
for (int i = 0; i < Constants.POTENTIAL_EMOJI_NAME_LIST.Count(); i++)
{
// Record the dominant emotion probability & its location
if (softMaxexOutputs[i] > maxProb)
{
maxIndex = i;
maxProb = softMaxexOutputs[i];
}
}
Debug.WriteLine($"Probability = {maxProb}, Threshold set to = {Constants.CLASSIFICATION_CERTAINTY_THRESHOLD}, Emotion = {Constants.POTENTIAL_EMOJI_NAME_LIST[maxIndex]}");
// For evaluations run on the MainPage, update the emoji carousel
if (maxProb >= Constants.CLASSIFICATION_CERTAINTY_THRESHOLD)
{
Debug.WriteLine("first page emoji should start to update");
IntelligenceServiceEmotionClassified?.Invoke(this, new ClassifiedEmojiEventArgs(CurrentEmojis._emojis.Emojis[maxIndex]));
}
// Dispose of resources
if (e.SoftwareBitmap != null)
{
e.SoftwareBitmap.Dispose();
e.SoftwareBitmap = null;
}
}
}
IntelligenceServiceProcessingCompleted?.Invoke(this, null);
Debug.WriteLine($"Frame evaluation finished {DateTime.Now}");
}
//WinML team function
private List<float> SoftMax(List<float> inputs)
{
List<float> inputsExp = new List<float>();
float inputsExpSum = 0;
for (int i = 0; i < inputs.Count; i++)
{
var input = inputs[i];
inputsExp.Add((float)Math.Exp(input));
inputsExpSum += inputsExp[i];
}
inputsExpSum = inputsExpSum == 0 ? 1 : inputsExpSum;
for (int i = 0; i < inputs.Count; i++)
{
inputsExp[i] /= inputsExpSum;
}
return inputsExp;
}
public static BitmapBounds GetCropBounds(int srcWidth, int srcHeight, int targetWidth, int targetHeight)
{
var modelHeight = targetHeight;
var modelWidth = targetWidth;
BitmapBounds bounds = new BitmapBounds();
// we need to recalculate the crop bounds in order to correctly center-crop the input image
float flRequiredAspectRatio = (float)modelWidth / modelHeight;
if (flRequiredAspectRatio * srcHeight > (float)srcWidth)
{
// clip on the y axis
bounds.Height = (uint)Math.Min((srcWidth / flRequiredAspectRatio + 0.5f), srcHeight);
bounds.Width = (uint)srcWidth;
bounds.X = 0;
bounds.Y = (uint)(srcHeight - bounds.Height) / 2;
}
else // clip on the x axis
{
bounds.Width = (uint)Math.Min((flRequiredAspectRatio * srcHeight + 0.5f), srcWidth);
bounds.Height = (uint)srcHeight;
bounds.X = (uint)(srcWidth - bounds.Width) / 2; ;
bounds.Y = 0;
}
return bounds;
}