使用onnxruntime、opencvsharp
onnxruntime利用animegan2 onnx（512x512 体积约8M）训练好的机器学习模型进行本机计算
同时使用opencv（opencvsharp）的L0smooth替代原来的nuget插件包
---
dotnet add package AnimeGANv2_Onnx_Sharp --version 1.0.0
源代码：
songshizhao/AnimeGANv2_Sharp (github.com)
-----------
代码参考：
public class OnnxHelper
{
private static InferenceSession _session;
private static SessionOptions _option;
private static List<NamedOnnxValue> container = new List<NamedOnnxValue>();
public static async Task Init()
{
var file = await StorageFile.GetFileFromApplicationUriAsync(new Uri(@"ms-appx:///face_paint_512_v2_0.onnx"));
var buffer = await FileIO.ReadBufferAsync(file);
_option = new SessionOptions
{
LogSeverityLevel = OrtLoggingLevel.ORT_LOGGING_LEVEL_ERROR
};
_option.AppendExecutionProvider_CPU(0);
//_option.AppendExecutionProvider_CoreML();
_session = new InferenceSession(buffer.ToArray(), _option);
}
public static async Task Init2()
{
var file = await StorageFile.GetFileFromApplicationUriAsync(new Uri(@"ms-appx:///face_paint_512_v2_0.onnx"));
var buffer = await FileIO.ReadBufferAsync(file);
_option = new SessionOptions
{
LogSeverityLevel = OrtLoggingLevel.ORT_LOGGING_LEVEL_ERROR
};
_option.AppendExecutionProvider_CPU(0);
//_option.AppendExecutionProvider_CoreML();
_session = new InferenceSession(buffer.ToArray(), _option);
}
private static DenseTensor<float> PreInput(Mat mat)
{
var mat512 = new Mat();
Cv2.Resize(mat, mat512, new Size(512, 512));
Mat m512 = new Mat();
Cv2.CvtColor(mat512, m512, ColorConversionCodes.BGR2RGB);
//输入Tensor
DenseTensor<float> input_tensor = new DenseTensor<float>(new[] { 1, 3, 512, 512 });
// 输入Tensor
for (int y = 0; y < m512.Height; y++)
{
for (int x = 0; x < m512.Width; x++)
{
input_tensor[0, 0, y, x] = (mat512.At<Vec3b>(y, x)[0] / 255f - 0.5f) / 0.5f;
input_tensor[0, 1, y, x] = (mat512.At<Vec3b>(y, x)[1] / 255f - 0.5f) / 0.5f;
input_tensor[0, 2, y, x] = (mat512.At<Vec3b>(y, x)[2] / 255f - 0.5f) / 0.5f;
}
}
return input_tensor;
}
public static async Task<Mat> Smooth(Mat sourceMat, double lambda = 2e-2, double kappa = 2.0)
{
Mat output = new Mat();//sourceMat.Width,sourceMat.Height,MatType.CV_8UC4
sourceMat = sourceMat.CvtColor(ColorConversionCodes.BGRA2BGR, 3);
await Task.Run(() =>
{
OpenCvSharp.XImgProc.CvXImgProc.L0Smooth(sourceMat, output, lambda, kappa);
});
return output;
}
//
public static Mat Bytes2Mat8UC3(int w, int h, byte[] bytes)
{
Mat mat = new Mat(h, w, MatType.CV_8UC4, bytes);
mat = mat.CvtColor(ColorConversionCodes.BGRA2BGR, 3);
return mat;
}
public static byte[] Mat2Bytes(Mat mat)
{
return mat.ToBytes();
}
public static async Task<Mat> RunB2M(int w, int h, byte[] bytes)
{
Mat mat = Bytes2Mat8UC3(w, h, bytes);
var inputTensor = PreInput(mat);//DenseTensor<float> inputTensor
Mat result_mat = new Mat();
await Task.Run(() =>
{
container.Clear();
container.Add(NamedOnnxValue.CreateFromTensor<float>("input_image", inputTensor));
var results = _session.Run(container);
var resultArrays = results.ToArray();
var result_tensors = results[0].AsTensor<float>();
var result_array = result_tensors.ToArray();
float[] temp_r = new float[512 * 512];
float[] temp_g = new float[512 * 512];
float[] temp_b = new float[512 * 512];
Array.Copy(result_array, temp_r, 512 * 512);
Array.Copy(result_array, 512 * 512, temp_g, 0, 512 * 512);
Array.Copy(result_array, 512 * 512 * 2, temp_b, 0, 512 * 512);
Mat rmat = new Mat(512, 512, MatType.CV_32FC1, temp_r);
Mat gmat = new Mat(512, 512, MatType.CV_32FC1, temp_g);
Mat bmat = new Mat(512, 512, MatType.CV_32FC1, temp_b);
rmat = (rmat + 1f) * 127.5f;
gmat = (gmat + 1f) * 127.5f;
bmat = (bmat + 1f) * 127.5f;
Cv2.Merge(new Mat[] { rmat, gmat, bmat }, result_mat);
});
return result_mat;
}
public static async Task<byte[]> RunB2B(int w, int h, byte[] bytes)
{
Mat result_mat = new Mat();
await Task.Run(() =>
{
Mat mat = Bytes2Mat8UC3(w, h, bytes);
var inputTensor = PreInput(mat);
container.Clear();
container.Add(NamedOnnxValue.CreateFromTensor<float>("input_image", inputTensor));
var results = _session.Run(container);
var resultArrays = results.ToArray();
var result_tensors = results[0].AsTensor<float>();
var result_array = result_tensors.ToArray();
float[] temp_r = new float[512 * 512];
float[] temp_g = new float[512 * 512];
float[] temp_b = new float[512 * 512];
Array.Copy(result_array, temp_r, 512 * 512);
Array.Copy(result_array, 512 * 512, temp_g, 0, 512 * 512);
Array.Copy(result_array, 512 * 512 * 2, temp_b, 0, 512 * 512);
Mat rmat = new Mat(512, 512, MatType.CV_32FC1, temp_r);
Mat gmat = new Mat(512, 512, MatType.CV_32FC1, temp_g);
Mat bmat = new Mat(512, 512, MatType.CV_32FC1, temp_b);
rmat = (rmat + 1f) * 127.5f;
gmat = (gmat + 1f) * 127.5f;
bmat = (bmat + 1f) * 127.5f;
Cv2.Merge(new Mat[] { rmat, gmat, bmat }, result_mat);
});
return result_mat.ToBytes();
}
public static async Task<Mat> RunM2M(Mat mat)
{
var inputTensor = PreInput(mat);//DenseTensor<float> inputTensor
Mat result_mat = new Mat();
await Task.Run(() =>
{
container.Clear();
container.Add(NamedOnnxValue.CreateFromTensor<float>("input_image", inputTensor));
var results = _session.Run(container);
var resultArrays = results.ToArray();
var result_tensors = results[0].AsTensor<float>();
var result_array = result_tensors.ToArray();
float[] temp_r = new float[512 * 512];
float[] temp_g = new float[512 * 512];
float[] temp_b = new float[512 * 512];
Array.Copy(result_array, temp_r, 512 * 512);
Array.Copy(result_array, 512 * 512, temp_g, 0, 512 * 512);
Array.Copy(result_array, 512 * 512 * 2, temp_b, 0, 512 * 512);
Mat rmat = new Mat(512, 512, MatType.CV_32FC1, temp_r);
Mat gmat = new Mat(512, 512, MatType.CV_32FC1, temp_g);
Mat bmat = new Mat(512, 512, MatType.CV_32FC1, temp_b);
rmat = (rmat + 1f) * 127.5f;
gmat = (gmat + 1f) * 127.5f;
bmat = (bmat + 1f) * 127.5f;
Cv2.Merge(new Mat[] { rmat, gmat, bmat }, result_mat);
});
return result_mat;
}
}
-----------
点评：
opencv自带L0smooth的功能，可比我自己C#手写的无速度优化版本速度快多了，完全替代了我写的垃圾nuget