YoloSharp

Train Yolo model in C# with TorchSharp.
With the help of this project you won't have to transform .pt model to onnx, and can train your own model in C#.

Feature

• Written in C# only.
• Train and predict your own model.
• Support Yolov5, Yolov5u, Yolov8, Yolov11 and Yolov12 now.
• Support Predict and Segment now.
• Support n/s/m/l/x size.
• Support LetterBox and Mosaic4 method for preprocessing images.
• Support NMS with GPU.
• Support Load PreTrained models from ultralytics/yolov5/yolov8/yolo11 and yolov12(converted).
• Support .Net6 or higher.

Models

You can download yolov5/yolov8 pre-trained models here.

Prediction Checkpoints

model	n	s	m	l	x
yolov5	yolov5n	yolov5s	yolov5m	yolov5l	yolov5x
yolov5	yolov5nu	yolov5su	yolov5mu	yolov5lu	yolov5xu
yolov8	yolov8n	yolov8s	yolov8m	yolov8l	yolov8x
yolov11	yolov11n	yolov11s	yolov11m	yolov11l	yolov11x

Segmention Checkpoints

model	n	s	m	l	x
yolov8	yolov8n	yolov8s	yolov8m	yolov8l	yolov8x
yolov11	yolov11n	yolov11s	yolov11m	yolov11l	yolov11x

How to use

You can download the code or add it from nuget.

dotnet add package IntptrMax.YoloSharp

Note

Please add one of libtorch-cpu, libtorch-cuda-12.1, libtorch-cuda-12.1-win-x64 or libtorch-cuda-12.1-linux-x64 version 2.5.1.0 to execute.

In your code you can use it as below.

Predict

You can use it with the code below:

MagickImage predictImage = new MagickImage(predictImagePath);

// Create predictor
Predictor predictor = new Predictor(sortCount, yoloType: yoloType, deviceType: deviceType, yoloSize: yoloSize, dtype: dtype);

// Train model
predictor.LoadModel(preTrainedModelPath, skipNcNotEqualLayers: true);
predictor.Train(trainDataPath, valDataPath, outputPath: outputPath, batchSize: batchSize, epochs: epochs, useMosaic: true);

//ImagePredict image
predictor.LoadModel(Path.Combine(outputPath, "best.bin"));
List<Predictor.PredictResult> predictResult = predictor.ImagePredict(predictImage, predictThreshold, iouThreshold);

Use yolov5n pre-trained model to detect.

Segment

You can use it with the code below:

MagickImage predictImage = new MagickImage(predictImagePath);

// Create segmenter
Segmenter segmenter = new Segmenter(sortCount, yoloType: yoloType, deviceType: deviceType, yoloSize: yoloSize, dtype: dtype);
segmenter.LoadModel(preTrainedModelPath, skipNcNotEqualLayers: true);

// Train model
segmenter.Train(trainDataPath, valDataPath, outputPath: outputPath, batchSize: batchSize, epochs: epochs, useMosaic: false);
segmenter.LoadModel(Path.Combine(outputPath, "best.bin"));

// ImagePredict image
var (predictResult, resultImage) = segmenter.ImagePredict(predictImage, predictThreshold, iouThreshold);

Use yolov8n-seg pre-trained model to detect.