之前项目中用到C#调用python库，要不就是脚本调用，要不就是rpc服务调用，脚本调用无法传递复杂参数，rpc服务调用又觉得太麻烦，两边都需要制定协议。之前尝试过使用这个机器学习的库，是要通过后端调用python实现的，就试着去深入了解了一些这个东西，直到发现了pythonnet这个很🐂×的库，就尝试它来调用python上的sklearn。
1.准备数据
首先在python上下载手写字符的数据，这里下载很慢，在我给的git项目中会提供。
from sklearn.datasets import fetch_openmlimport numpy as npmnist = fetch_openml('mnist_784', version=1, as_frame=False,cache=True)mnist.target = mnist.target.astype(np.uint8)mnist.data.shapenp.save("mnist_data",mnist.data)np.save("mnist_target",mnist.target)
2.创建新项目
这里使用vs创建一个netcore3.1项目，然后安装3.8.5（这里面自带pythonnet依赖），这里后台安装的是python3.8.10，安装好对应的依赖库。
public class TSNEByScikit: Base{PyObject PyInstance = null;public TSNEByScikit(){if (!PythonEngine.IsInitialized)PythonEngine.Initialize();using (Py.GIL()){dynamic manifold = Py.Import("sklearn.manifold");//这里很重要，不能带括号，否则后面FitTransform方法会报奇怪的错误dynamic tsne = manifold.TSNE;var ksargs = GetPrammeters();var pyargs = ToTuple(new object[]{new PyInt(2)});PyInstance = tsne.Invoke(pyargs, ksargs);}}public NDarray FitTransform(NDarray X){using (Py.GIL()){var obj = X.ToPython();PyTuple pyTuple = new PyTuple(new PyObject[] {X.PyObject});var kwargs = new PyDict();kwargs["X"] = X.PyObject;dynamic np = Py.Import("numpy");dynamic npX= np.array(new float[] {1,2,3});dynamic manifold = Py.Import("sklearn.linear_model");NDarray result= new NDarray(this.PyInstance.InvokeMethod("fit_transform", pyTuple));//float[] netResult = result.GetData<float>();return result;}}private PyDict GetPrammeters(){PyDict result = new PyDict();//result["n_components"] =new PyInt(2);result["random_state"] = new PyInt(42);return result;}}
3.调用tsne算法
先设置好参与训练的数据个数，点击读取数据，再点击tsne转换。
public partial class FrmMain : Form{public FrmMain(){InitializeComponent();}NDarray X = null;NDarray y = null;private void btnLoad_Click(object sender, EventArgs e){var X=np.load("mnist_data.npy");lstLog.Items.Add(X.shape);var y = np.load("mnist_target.npy");lstLog.Items.Add(y.shape);int count = (int)numScale.Value;this.X = X[$":{count},:"].astype(np.float16);this.y = y[$":{count}"]; ;}private void btnTSNE_Click(object sender, EventArgs e){Stopwatch sw = new Stopwatch();sw.Start();TSNEByScikit tsne = new TSNEByScikit();var result=tsne.FitTransform(X);sw.Stop();lstLog.Items.Add($"{this.X.shape[0]}条数据，耗时{sw.ElapsedMilliseconds}ms");var xs = result[":,0"].GetData<float>().Select(t => (double)t).ToArray(); var ys = result[":,1"].GetData<float>().Select(t => (double)t).ToArray();var t = y.GetData<byte>();var splitSamples = SplitSamples(xs, ys, t);pltResult.Plot.Clear();var colors = GetColors();foreach (var item in splitSamples){int colorData = item.target;pltResult.Plot.AddScatter(new double[] {item.xs}, new double[] {item.ys }, colors[colorData]);}pltResult.Refresh();}private List<(int target, double xs, double ys)> SplitSamples(double[] xs,double[] ys,byte[] target){List<(int, double, double)> datas = new List<(int, double, double)>();for (int i = 0; i < xs.Length; i++){datas.Add((target[i], xs[i], ys[i]));}return datas;}private Color[] GetColors(){return new Color[] {Color.Red, Color.Yellow, Color.Gray, Color.Gold, Color.HotPink, Color.LemonChiffon, Color.Black, Color.Black, Color.Brown,Color.Cyan };}}
仓库链接
