ultralytics/examples/YOLOv8-ONNXRuntime-Rust

YOLOv8-ONNXRuntime-Rust for All Key YOLO Tasks

This repository provides a Rust demonstration for performing Ultralytics YOLOv8 tasks like Classification, Segmentation, Detection, Pose Estimation, and Oriented Bounding Box (OBB) detection using the ONNXRuntime.

✨ Recently Updated

• Added YOLOv8-OBB demo.
• Updated ONNXRuntime dependency to 1.19.x.

Newly updated YOLOv8 example code is located in this repository.

🚀 Features

• Supports Classification, Segmentation, Detection, Pose(Keypoints)-Detection, and OBB tasks.
• Supports FP16 & FP32 ONNX models.
• Supports CPU, CUDA, and TensorRT execution providers to accelerate computation.
• Supports dynamic input shapes (batch, width, height).

🛠️ Installation

1. Install Rust

Please follow the official Rust installation guide: https://www.rust-lang.org/tools/install.

2. ONNXRuntime Linking

• For detailed setup instructions, refer to the ORT documentation.

• For Linux or macOS Users:

  • Download the ONNX Runtime package from the Releases page.
  • Set up the library path by exporting the ORT_DYLIB_PATH environment variable:

    export ORT_DYLIB_PATH=/path/to/onnxruntime/lib/libonnxruntime.so.1.19.0 # Adjust version/path as needed
    

3. [Optional] Install CUDA & CuDNN & TensorRT

• The CUDA execution provider requires CUDA v11.6+.
• The TensorRT execution provider requires CUDA v11.4+ and TensorRT v8.4+. You may also need cuDNN.

▶️ Get Started

1. Export the Ultralytics YOLOv8 ONNX Models

First, install the Ultralytics package:

pip install -U ultralytics

Then, export the desired Ultralytics YOLOv8 models to the ONNX format. See the Export documentation for more details.

# Export ONNX model with dynamic shapes (recommended for flexibility)
yolo export model=yolov8m.pt format=onnx simplify dynamic
yolo export model=yolov8m-cls.pt format=onnx simplify dynamic
yolo export model=yolov8m-pose.pt format=onnx simplify dynamic
yolo export model=yolov8m-seg.pt format=onnx simplify dynamic
# yolo export model=yolov8m-obb.pt format=onnx simplify dynamic # Add OBB export if needed

# Export ONNX model with constant shapes (if dynamic shapes are not required)
# yolo export model=yolov8m.pt format=onnx simplify
# yolo export model=yolov8m-cls.pt format=onnx simplify
# yolo export model=yolov8m-pose.pt format=onnx simplify
# yolo export model=yolov8m-seg.pt format=onnx simplify
# yolo export model=yolov8m-obb.pt format=onnx simplify

2. Run Inference

This command will perform inference using the specified ONNX model on the source image using the CPU.

cargo run --release -- --model MODEL_PATH.onnx --source SOURCE_IMAGE.jpg

Using GPU Acceleration

Set --cuda to use the CUDA execution provider for faster inference on NVIDIA GPUs.

cargo run --release -- --cuda --model MODEL_PATH.onnx --source SOURCE_IMAGE.jpg

Set --trt to use the TensorRT execution provider. You can also set --fp16 simultaneously to leverage the TensorRT FP16 engine for potentially even greater speed, especially on compatible hardware.

cargo run --release -- --trt --fp16 --model MODEL_PATH.onnx --source SOURCE_IMAGE.jpg

Specifying Device and Batch Size

Set --device_id to select a specific GPU device. If the specified device ID is invalid (e.g., setting device_id 1 when only one GPU exists), ort will automatically fall back to the CPU execution provider without causing a panic.

cargo run --release -- --cuda --device_id 0 --model MODEL_PATH.onnx --source SOURCE_IMAGE.jpg

Set --batch to perform inference with a specific batch size.

cargo run --release -- --cuda --batch 2 --model MODEL_PATH.onnx --source SOURCE_IMAGE.jpg

If you're using --trt with a model exported with dynamic batch dimensions, you can explicitly specify the minimum, optimal, and maximum batch sizes for TensorRT optimization using --batch-min, --batch, and --batch-max. Refer to the TensorRT Execution Provider documentation for details.

Dynamic Image Size

Set --height and --width to perform inference with dynamic image sizes. Note: The ONNX model must have been exported with dynamic input shapes (dynamic=True).

cargo run --release -- --cuda --width 480 --height 640 --model MODEL_PATH_dynamic.onnx --source SOURCE_IMAGE.jpg

Profiling Performance

Set --profile to measure the time consumed in each stage of the inference pipeline (preprocessing, H2D transfer, inference, D2H transfer, postprocessing). Note: Models often require a few "warm-up" runs (1-3 iterations) before reaching optimal performance. Ensure you run the command enough times to get a stable performance evaluation.

cargo run --release -- --trt --fp16 --profile --model MODEL_PATH.onnx --source SOURCE_IMAGE.jpg

Example Profile Output (yolov8m.onnx, batch=1, 3 runs, trt, fp16, RTX 3060Ti):

==> 0 # Warm-up run
[Model Preprocess]: 12.75788ms
[ORT H2D]: 237.118µs
[ORT Inference]: 507.895469ms
[ORT D2H]: 191.655µs
[Model Inference]: 508.34589ms
[Model Postprocess]: 1.061122ms
==> 1 # Stable run
[Model Preprocess]: 13.658655ms
[ORT H2D]: 209.975µs
[ORT Inference]: 5.12372ms
[ORT D2H]: 182.389µs
[Model Inference]: 5.530022ms
[Model Postprocess]: 1.04851ms
==> 2 # Stable run
[Model Preprocess]: 12.475332ms
[ORT H2D]: 246.127µs
[ORT Inference]: 5.048432ms
[ORT D2H]: 187.117µs
[Model Inference]: 5.493119ms
[Model Postprocess]: 1.040906ms

Other Options

• --conf: Confidence threshold for detections [default: 0.3].
• --iou: IoU (Intersection over Union) threshold for Non-Maximum Suppression (NMS) [default: 0.45].
• --kconf: Confidence threshold for keypoints (in Pose Estimation) [default: 0.55].
• --plot: Plot the inference results with random RGB colors and save the output image to the runs directory.

You can view all available command-line arguments by running:

# Clone the repository if you haven't already
# git clone https://github.com/ultralytics/ultralytics
# cd ultralytics/examples/YOLOv8-ONNXRuntime-Rust

cargo run --release -- --help

🖼️ Examples

Classification

Running a dynamic shape ONNX classification model on the CPU with a specific image size (--height 224 --width 224). The plotted result image will be saved in the runs directory.

cargo run --release -- --model ../assets/weights/yolov8m-cls-dyn.onnx --source ../assets/images/dog.jpg --height 224 --width 224 --plot --profile

Example output:

Summary:
> Task: Classify (Ultralytics 8.0.217) # Version might differ
> EP: Cpu
> Dtype: Float32
> Batch: 1 (Dynamic), Height: 224 (Dynamic), Width: 224 (Dynamic)
> nc: 1000 nk: 0, nm: 0, conf: 0.3, kconf: 0.55, iou: 0.45

[Model Preprocess]: 16.363477ms
[ORT H2D]: 50.722µs
[ORT Inference]: 16.295808ms
[ORT D2H]: 8.37µs
[Model Inference]: 16.367046ms
[Model Postprocess]: 3.527µs
[
    YOLOResult {
        Probs(top5): Some([(208, 0.6950566), (209, 0.13823675), (178, 0.04849795), (215, 0.019029364), (212, 0.016506357)]), # Class IDs and confidences
        Bboxes: None,
        Keypoints: None,
        Masks: None,
    },
]

Object Detection

Using the CUDA execution provider and a dynamic image size (--height 640 --width 480).

cargo run --release -- --cuda --model ../assets/weights/yolov8m-dynamic.onnx --source ../assets/images/bus.jpg --plot --height 640 --width 480

Pose Detection

Using the TensorRT execution provider.

cargo run --release -- --trt --model ../assets/weights/yolov8m-pose.onnx --source ../assets/images/bus.jpg --plot

Instance Segmentation

Using the TensorRT execution provider with an FP16 model (--fp16).

cargo run --release -- --trt --fp16 --model ../assets/weights/yolov8m-seg.onnx --source ../assets/images/0172.jpg --plot

🤝 Contributing

Contributions are welcome! If you find any issues or have suggestions for improvement, please feel free to open an issue or submit a pull request to the main Ultralytics repository.