[Example] YOLOv8-ONNXRuntime-Rust example (#6583)

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>
1 year ago · fdcf0dd4fd
parent 3c277347e4
commit fdcf0dd4fd
8 changed files with 1888 additions and 0 deletions
--- a/examples/YOLOv8-ONNXRuntime-Rust/Cargo.toml
+++ b/examples/YOLOv8-ONNXRuntime-Rust/Cargo.toml
@ -0,0 +1,21 @@
+[package]
+name = "yolov8-rs"
+version = "0.1.0"
+edition = "2021"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+clap = { version = "4.2.4", features = ["derive"] }
+image = { version = "0.24.7", default-features = false, features = ["jpeg", "png", "webp-encoder"] }
+imageproc = { version = "0.23.0", default-features = false }
+ndarray = { version = "0.15.6" }
+ort =  {version = "1.16.3", default-features = false, features = ["load-dynamic", "copy-dylibs", "half"]}
+rusttype = { version = "0.9", default-features = false }
+anyhow = { version = "1.0.75"}
+regex = { version = "1.5.4" }
+rand = { version ="0.8.5" }
+chrono = { version = "0.4.30" }
+half = { version = "2.3.1" }
+dirs = { version = "5.0.1" }
+ureq = { version = "2.9.1" }
--- a/examples/YOLOv8-ONNXRuntime-Rust/README.md
+++ b/examples/YOLOv8-ONNXRuntime-Rust/README.md
@ -0,0 +1,222 @@
+# YOLOv8-ONNXRuntime-Rust for All the Key YOLO Tasks
+
+This repository provides a Rust demo for performing YOLOv8 tasks like `Classification`, `Segmentation`, `Detection` and `Pose Detection` using ONNXRuntime.
+
+## Features
+
+- Support `Classification`, `Segmentation`, `Detection`, `Pose(Keypoints)-Detection` tasks.
+- Support `FP16` & `FP32` ONNX models.
+- Support `CPU`, `CUDA` and `TensorRT` execution provider to accelerate computation.
+- Support dynamic input shapes(`batch`, `width`, `height`).
+
+## Installation
+
+### 1. Install Rust
+
+Please follow the Rust official installation. (https://www.rust-lang.org/tools/install)
+
+### 2. Install ONNXRuntime
+
+This repository use `ort` crate, which is ONNXRuntime wrapper for Rust. (https://docs.rs/ort/latest/ort/)
+
+You can follow the instruction with `ort` doc or simply do this:
+
+- step1: Download ONNXRuntime(https://github.com/microsoft/onnxruntime/releases)
+- setp2: Set environment variable `PATH` for linking.
+
+On ubuntu, You can do like this:
+
+```
+vim ~/.bashrc
+
+# Add the path of ONNXRUntime lib
+export LD_LIBRARY_PATH=/home/qweasd/Documents/onnxruntime-linux-x64-gpu-1.16.3/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
+
+source ~/.bashrc
+```
+
+### 3. \[Optional\] Install CUDA & CuDNN & TensorRT
+
+- CUDA execution provider requires CUDA v11.6+.
+- TensorRT execution provider requires CUDA v11.4+ and TensorRT v8.4+.
+
+## Get Started
+
+### 1. Export the YOLOv8 ONNX Models
+
+```bash
+pip install -U ultralytics
+
+# export onnx model with dynamic shapes
+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
+
+
+# export onnx model with constant shapes
+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
+```
+
+### 2. Run Inference
+
+It will perform inference with the ONNX model on the source image.
+
+```
+cargo run --release -- --model <MODEL> --source <SOURCE>
+```
+
+Set `--cuda` to use CUDA execution provider to speed up inference.
+
+```
+cargo run --release -- --cuda --model <MODEL> --source <SOURCE>
+```
+
+Set `--trt` to use TensorRT execution provider, and you can set `--fp16` at the same time to use TensorRT FP16 engine.
+
+```
+cargo run --release -- --trt --fp16 --model <MODEL> --source <SOURCE>
+```
+
+Set `--device_id` to select which device to run. When you have only one GPU, and you set `device_id` to 1 will not cause program panic, the `ort` would automatically fall back to `CPU` EP.
+
+```
+cargo run --release -- --cuda --device_id 0 --model <MODEL> --source <SOURCE>
+```
+
+Set `--batch` to do multi-batch-size inference.
+
+If you're using `--trt`, you can also set `--batch-min` and `--batch-max` to explicitly specify min/max/opt batch for dynamic batch input.(https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#explicit-shape-range-for-dynamic-shape-input).(Note that the ONNX model should exported with dynamic shapes)
+
+```
+cargo run --release -- --cuda --batch 2 --model <MODEL> --source <SOURCE>
+```
+
+Set `--height` and `--width` to do dynamic image size inference. (Note that the ONNX model should exported with dynamic shapes)
+
+```
+cargo run --release -- --cuda --width 480 --height 640 --model <MODEL> --source <SOURCE>
+```
+
+Set `--profile` to check time consumed in each stage.(Note that the model usually needs to take 1~3 times dry run to warmup. Make sure to run enough times to evaluate the result.)
+
+```
+cargo run --release -- --trt --fp16 --profile --model <MODEL> --source <SOURCE>
+```
+
+Results: (yolov8m.onnx, batch=1, 3 times, trt, fp16, RTX 3060Ti)
+
+```
+==> 0
+[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
+[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
+[Model Preprocess]: 12.475332ms
+[ORT H2D]: 246.127µs
+[ORT Inference]: 5.048432ms
+[ORT D2H]: 187.117µs
+[Model Inference]: 5.493119ms
+[Model Postprocess]: 1.040906ms
+```
+
+And also:
+
+`--conf`: confidence threshold \[default: 0.3\]
+
+`--iou`: iou threshold in NMS \[default: 0.45\]
+
+`--kconf`: confidence threshold of keypoint \[default: 0.55\]
+
+`--plot`: plot inference result with random RGB color and save
+
+you can check out all CLI arguments by:
+
+```
+git clone https://github.com/ultralytics/ultralytics
+cd ultralytics/examples/YOLOv8-ONNXRuntime-Rust
+cargo run --release -- --help
+```
+
+## Examples
+
+### Classification
+
+Running dynamic shape ONNX model on `CPU` with image size `--height 224 --width 224`.
+Saving plotted image in `runs` directory.
+
+```
+cargo run --release -- --model ../assets/weights/yolov8m-cls-dyn.onnx --source ../assets/images/dog.jpg --height 224 --width 224 --plot --profile
+```
+
+You will see result like:
+
+```
+Summary:
+> Task: Classify (Ultralytics 8.0.217)
+> 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)]),
+        Bboxes: None,
+        Keypoints: None,
+        Masks: None,
+    },
+]
+
+```
+
+![2023-11-25-22-02-02-156623351](https://github.com/jamjamjon/ultralytics/assets/51357717/ef75c2ae-c5ab-44cc-9d9e-e60b51e39662)
+
+### Object Detection
+
+Using `CUDA` EP and 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
+```
+
+![det](https://github.com/jamjamjon/ultralytics/assets/51357717/5d89a19d-0c96-4a59-875c-defab6887a2c)
+
+### Pose Detection
+
+using `TensorRT` EP
+
+```
+cargo run --release -- --trt --model ../assets/weights/yolov8m-pose.onnx --source ../assets/images/bus.jpg --plot
+```
+
+![2023-11-25-22-31-45-127054025](https://github.com/jamjamjon/ultralytics/assets/51357717/157b5ba7-bfcf-47cf-bee7-68b62e0de1c4)
+
+### Instance Segmentation
+
+using `TensorRT` EP and FP16 model `--fp16`
+
+```
+cargo run --release --  --trt --fp16 --model ../assets/weights/yolov8m-seg.onnx --source ../assets/images/0172.jpg --plot
+```
+
+![seg](https://github.com/jamjamjon/ultralytics/assets/51357717/cf046f4f-9533-478a-adc7-4de22443a641)
--- a/examples/YOLOv8-ONNXRuntime-Rust/src/cli.rs
+++ b/examples/YOLOv8-ONNXRuntime-Rust/src/cli.rs
@ -0,0 +1,87 @@
+use clap::Parser;
+
+use crate::YOLOTask;
+
+#[derive(Parser, Clone)]
+#[command(author, version, about, long_about = None)]
+pub struct Args {
+    /// ONNX model path
+    #[arg(long, required = true)]
+    pub model: String,
+
+    /// input path
+    #[arg(long, required = true)]
+    pub source: String,
+
+    /// device id
+    #[arg(long, default_value_t = 0)]
+    pub device_id: u32,
+
+    /// using TensorRT EP
+    #[arg(long)]
+    pub trt: bool,
+
+    /// using CUDA EP
+    #[arg(long)]
+    pub cuda: bool,
+
+    /// input batch size
+    #[arg(long, default_value_t = 1)]
+    pub batch: u32,
+
+    /// trt input min_batch size
+    #[arg(long, default_value_t = 1)]
+    pub batch_min: u32,
+
+    /// trt input max_batch size
+    #[arg(long, default_value_t = 32)]
+    pub batch_max: u32,
+
+    /// using TensorRT --fp16
+    #[arg(long)]
+    pub fp16: bool,
+
+    /// specify YOLO task
+    #[arg(long, value_enum)]
+    pub task: Option<YOLOTask>,
+
+    /// num_classes
+    #[arg(long)]
+    pub nc: Option<u32>,
+
+    /// num_keypoints
+    #[arg(long)]
+    pub nk: Option<u32>,
+
+    /// num_masks
+    #[arg(long)]
+    pub nm: Option<u32>,
+
+    /// input image width
+    #[arg(long)]
+    pub width: Option<u32>,
+
+    /// input image height
+    #[arg(long)]
+    pub height: Option<u32>,
+
+    /// confidence threshold
+    #[arg(long, required = false, default_value_t = 0.3)]
+    pub conf: f32,
+
+    /// iou threshold in NMS
+    #[arg(long, required = false, default_value_t = 0.45)]
+    pub iou: f32,
+
+    /// confidence threshold of keypoint
+    #[arg(long, required = false, default_value_t = 0.55)]
+    pub kconf: f32,
+
+    /// plot inference result and save
+    #[arg(long)]
+    pub plot: bool,
+
+    /// check time consumed in each stage
+    #[arg(long)]
+    pub profile: bool,
+}
--- a/examples/YOLOv8-ONNXRuntime-Rust/src/lib.rs
+++ b/examples/YOLOv8-ONNXRuntime-Rust/src/lib.rs
@ -0,0 +1,119 @@
+#![allow(clippy::type_complexity)]
+
+use std::io::{Read, Write};
+
+pub mod cli;
+pub mod model;
+pub mod ort_backend;
+pub mod yolo_result;
+pub use crate::cli::Args;
+pub use crate::model::YOLOv8;
+pub use crate::ort_backend::{Batch, OrtBackend, OrtConfig, OrtEP, YOLOTask};
+pub use crate::yolo_result::{Bbox, Embedding, Point2, YOLOResult};
+
+pub fn non_max_suppression(
+    xs: &mut Vec<(Bbox, Option<Vec<Point2>>, Option<Vec<f32>>)>,
+    iou_threshold: f32,
+) {
+    xs.sort_by(|b1, b2| b2.0.confidence().partial_cmp(&b1.0.confidence()).unwrap());
+
+    let mut current_index = 0;
+    for index in 0..xs.len() {
+        let mut drop = false;
+        for prev_index in 0..current_index {
+            let iou = xs[prev_index].0.iou(&xs[index].0);
+            if iou > iou_threshold {
+                drop = true;
+                break;
+            }
+        }
+        if !drop {
+            xs.swap(current_index, index);
+            current_index += 1;
+        }
+    }
+    xs.truncate(current_index);
+}
+
+pub fn gen_time_string(delimiter: &str) -> String {
+    let offset = chrono::FixedOffset::east_opt(8 * 60 * 60).unwrap(); // Beijing
+    let t_now = chrono::Utc::now().with_timezone(&offset);
+    let fmt = format!(
+        "%Y{}%m{}%d{}%H{}%M{}%S{}%f",
+        delimiter, delimiter, delimiter, delimiter, delimiter, delimiter
+    );
+    t_now.format(&fmt).to_string()
+}
+
+pub const SKELETON: [(usize, usize); 16] = [
+    (0, 1),
+    (0, 2),
+    (1, 3),
+    (2, 4),
+    (5, 6),
+    (5, 11),
+    (6, 12),
+    (11, 12),
+    (5, 7),
+    (6, 8),
+    (7, 9),
+    (8, 10),
+    (11, 13),
+    (12, 14),
+    (13, 15),
+    (14, 16),
+];
+
+pub fn check_font(font: &str) -> rusttype::Font<'static> {
+    // check then load font
+
+    // ultralytics font path
+    let font_path_config = match dirs::config_dir() {
+        Some(mut d) => {
+            d.push("Ultralytics");
+            d.push(font);
+            d
+        }
+        None => panic!("Unsupported operating system. Now support Linux, MacOS, Windows."),
+    };
+
+    // current font path
+    let font_path_current = std::path::PathBuf::from(font);
+
+    // check font
+    let font_path = if font_path_config.exists() {
+        font_path_config
+    } else if font_path_current.exists() {
+        font_path_current
+    } else {
+        println!("Downloading font...");
+        let source_url = "https://ultralytics.com/assets/Arial.ttf";
+        let resp = ureq::get(source_url)
+            .timeout(std::time::Duration::from_secs(500))
+            .call()
+            .unwrap_or_else(|err| panic!("> Failed to download font: {source_url}: {err:?}"));
+
+        // read to buffer
+        let mut buffer = vec![];
+        let total_size = resp
+            .header("Content-Length")
+            .and_then(|s| s.parse::<u64>().ok())
+            .unwrap();
+        let _reader = resp
+            .into_reader()
+            .take(total_size)
+            .read_to_end(&mut buffer)
+            .unwrap();
+
+        // save
+        let _path = std::fs::File::create(font).unwrap();
+        let mut writer = std::io::BufWriter::new(_path);
+        writer.write_all(&buffer).unwrap();
+        println!("Font saved at: {:?}", font_path_current.display());
+        font_path_current
+    };
+
+    // load font
+    let buffer = std::fs::read(font_path).unwrap();
+    rusttype::Font::try_from_vec(buffer).unwrap()
+}
--- a/examples/YOLOv8-ONNXRuntime-Rust/src/main.rs
+++ b/examples/YOLOv8-ONNXRuntime-Rust/src/main.rs
@ -0,0 +1,28 @@
+use clap::Parser;
+
+use yolov8_rs::{Args, YOLOv8};
+
+fn main() -> Result<(), Box<dyn std::error::Error>> {
+    let args = Args::parse();
+
+    // 1. load image
+    let x = image::io::Reader::open(&args.source)?
+        .with_guessed_format()?
+        .decode()?;
+
+    // 2. model support dynamic batch inference, so input should be a Vec
+    let xs = vec![x];
+
+    // You can test `--batch 2` with this
+    // let xs = vec![x.clone(), x];
+
+    // 3. build yolov8 model
+    let mut model = YOLOv8::new(args)?;
+    model.summary(); // model info
+
+    // 4. run
+    let ys = model.run(&xs)?;
+    println!("{:?}", ys);
+
+    Ok(())
+}
--- a/examples/YOLOv8-ONNXRuntime-Rust/src/model.rs
+++ b/examples/YOLOv8-ONNXRuntime-Rust/src/model.rs
@ -0,0 +1,642 @@
+#![allow(clippy::type_complexity)]
+
+use anyhow::Result;
+use image::{DynamicImage, GenericImageView, ImageBuffer};
+use ndarray::{s, Array, Axis, IxDyn};
+use rand::{thread_rng, Rng};
+use std::path::PathBuf;
+
+use crate::{
+    check_font, gen_time_string, non_max_suppression, Args, Batch, Bbox, Embedding, OrtBackend,
+    OrtConfig, OrtEP, Point2, YOLOResult, YOLOTask, SKELETON,
+};
+
+pub struct YOLOv8 {
+    // YOLOv8 model for all yolo-tasks
+    engine: OrtBackend,
+    nc: u32,
+    nk: u32,
+    nm: u32,
+    height: u32,
+    width: u32,
+    batch: u32,
+    task: YOLOTask,
+    conf: f32,
+    kconf: f32,
+    iou: f32,
+    names: Vec<String>,
+    color_palette: Vec<(u8, u8, u8)>,
+    profile: bool,
+    plot: bool,
+}
+
+impl YOLOv8 {
+    pub fn new(config: Args) -> Result<Self> {
+        // execution provider
+        let ep = if config.trt {
+            OrtEP::Trt(config.device_id)
+        } else if config.cuda {
+            OrtEP::Cuda(config.device_id)
+        } else {
+            OrtEP::Cpu
+        };
+
+        // batch
+        let batch = Batch {
+            opt: config.batch,
+            min: config.batch_min,
+            max: config.batch_max,
+        };
+
+        // build ort engine
+        let ort_args = OrtConfig {
+            ep,
+            batch,
+            f: config.model,
+            task: config.task,
+            trt_fp16: config.fp16,
+            image_size: (config.height, config.width),
+        };
+        let engine = OrtBackend::build(ort_args)?;
+
+        //  get batch, height, width, tasks, nc, nk, nm
+        let (batch, height, width, task) = (
+            engine.batch(),
+            engine.height(),
+            engine.width(),
+            engine.task(),
+        );
+        let nc = engine.nc().or(config.nc).unwrap_or_else(|| {
+            panic!("Failed to get num_classes, make it explicit with `--nc`");
+        });
+        let (nk, nm) = match task {
+            YOLOTask::Pose => {
+                let nk = engine.nk().or(config.nk).unwrap_or_else(|| {
+                    panic!("Failed to get num_keypoints, make it explicit with `--nk`");
+                });
+                (nk, 0)
+            }
+            YOLOTask::Segment => {
+                let nm = engine.nm().or(config.nm).unwrap_or_else(|| {
+                    panic!("Failed to get num_masks, make it explicit with `--nm`");
+                });
+                (0, nm)
+            }
+            _ => (0, 0),
+        };
+
+        // class names
+        let names = engine.names().unwrap_or(vec!["Unknown".to_string()]);
+
+        // color palette
+        let mut rng = thread_rng();
+        let color_palette: Vec<_> = names
+            .iter()
+            .map(|_| {
+                (
+                    rng.gen_range(0..=255),
+                    rng.gen_range(0..=255),
+                    rng.gen_range(0..=255),
+                )
+            })
+            .collect();
+
+        Ok(Self {
+            engine,
+            names,
+            conf: config.conf,
+            kconf: config.kconf,
+            iou: config.iou,
+            color_palette,
+            profile: config.profile,
+            plot: config.plot,
+            nc,
+            nk,
+            nm,
+            height,
+            width,
+            batch,
+            task,
+        })
+    }
+
+    pub fn scale_wh(&self, w0: f32, h0: f32, w1: f32, h1: f32) -> (f32, f32, f32) {
+        let r = (w1 / w0).min(h1 / h0);
+        (r, (w0 * r).round(), (h0 * r).round())
+    }
+
+    pub fn preprocess(&mut self, xs: &Vec<DynamicImage>) -> Result<Array<f32, IxDyn>> {
+        let mut ys =
+            Array::ones((xs.len(), 3, self.height() as usize, self.width() as usize)).into_dyn();
+        ys.fill(144.0 / 255.0);
+        for (idx, x) in xs.iter().enumerate() {
+            let img = match self.task() {
+                YOLOTask::Classify => x.resize_exact(
+                    self.width(),
+                    self.height(),
+                    image::imageops::FilterType::Triangle,
+                ),
+                _ => {
+                    let (w0, h0) = x.dimensions();
+                    let w0 = w0 as f32;
+                    let h0 = h0 as f32;
+                    let (_, w_new, h_new) =
+                        self.scale_wh(w0, h0, self.width() as f32, self.height() as f32); // f32 round
+                    x.resize_exact(
+                        w_new as u32,
+                        h_new as u32,
+                        if let YOLOTask::Segment = self.task() {
+                            image::imageops::FilterType::CatmullRom
+                        } else {
+                            image::imageops::FilterType::Triangle
+                        },
+                    )
+                }
+            };
+
+            for (x, y, rgb) in img.pixels() {
+                let x = x as usize;
+                let y = y as usize;
+                let [r, g, b, _] = rgb.0;
+                ys[[idx, 0, y, x]] = (r as f32) / 255.0;
+                ys[[idx, 1, y, x]] = (g as f32) / 255.0;
+                ys[[idx, 2, y, x]] = (b as f32) / 255.0;
+            }
+        }
+
+        Ok(ys)
+    }
+
+    pub fn run(&mut self, xs: &Vec<DynamicImage>) -> Result<Vec<YOLOResult>> {
+        // pre-process
+        let t_pre = std::time::Instant::now();
+        let xs_ = self.preprocess(xs)?;
+        if self.profile {
+            println!("[Model Preprocess]: {:?}", t_pre.elapsed());
+        }
+
+        // run
+        let t_run = std::time::Instant::now();
+        let ys = self.engine.run(xs_, self.profile)?;
+        if self.profile {
+            println!("[Model Inference]: {:?}", t_run.elapsed());
+        }
+
+        // post-process
+        let t_post = std::time::Instant::now();
+        let ys = self.postprocess(ys, xs)?;
+        if self.profile {
+            println!("[Model Postprocess]: {:?}", t_post.elapsed());
+        }
+
+        // plot and save
+        if self.plot {
+            self.plot_and_save(&ys, xs, Some(&SKELETON));
+        }
+        Ok(ys)
+    }
+
+    pub fn postprocess(
+        &self,
+        xs: Vec<Array<f32, IxDyn>>,
+        xs0: &[DynamicImage],
+    ) -> Result<Vec<YOLOResult>> {
+        if let YOLOTask::Classify = self.task() {
+            let mut ys = Vec::new();
+            let preds = &xs[0];
+            for batch in preds.axis_iter(Axis(0)) {
+                ys.push(YOLOResult::new(
+                    Some(Embedding::new(batch.into_owned())),
+                    None,
+                    None,
+                    None,
+                ));
+            }
+            Ok(ys)
+        } else {
+            const CXYWH_OFFSET: usize = 4; // cxcywh
+            const KPT_STEP: usize = 3; // xyconf
+            let preds = &xs[0];
+            let protos = {
+                if xs.len() > 1 {
+                    Some(&xs[1])
+                } else {
+                    None
+                }
+            };
+            let mut ys = Vec::new();
+            for (idx, anchor) in preds.axis_iter(Axis(0)).enumerate() {
+                // [bs, 4 + nc + nm, anchors]
+                // input image
+                let width_original = xs0[idx].width() as f32;
+                let height_original = xs0[idx].height() as f32;
+                let ratio = (self.width() as f32 / width_original)
+                    .min(self.height() as f32 / height_original);
+
+                // save each result
+                let mut data: Vec<(Bbox, Option<Vec<Point2>>, Option<Vec<f32>>)> = Vec::new();
+                for pred in anchor.axis_iter(Axis(1)) {
+                    // split preds for different tasks
+                    let bbox = pred.slice(s![0..CXYWH_OFFSET]);
+                    let clss = pred.slice(s![CXYWH_OFFSET..CXYWH_OFFSET + self.nc() as usize]);
+                    let kpts = {
+                        if let YOLOTask::Pose = self.task() {
+                            Some(pred.slice(s![pred.len() - KPT_STEP * self.nk() as usize..]))
+                        } else {
+                            None
+                        }
+                    };
+                    let coefs = {
+                        if let YOLOTask::Segment = self.task() {
+                            Some(pred.slice(s![pred.len() - self.nm() as usize..]).to_vec())
+                        } else {
+                            None
+                        }
+                    };
+
+                    // confidence and id
+                    let (id, &confidence) = clss
+                        .into_iter()
+                        .enumerate()
+                        .reduce(|max, x| if x.1 > max.1 { x } else { max })
+                        .unwrap(); // definitely will not panic!
+
+                    // confidence filter
+                    if confidence < self.conf {
+                        continue;
+                    }
+
+                    // bbox re-scale
+                    let cx = bbox[0] / ratio;
+                    let cy = bbox[1] / ratio;
+                    let w = bbox[2] / ratio;
+                    let h = bbox[3] / ratio;
+                    let x = cx - w / 2.;
+                    let y = cy - h / 2.;
+                    let y_bbox = Bbox::new(
+                        x.max(0.0f32).min(width_original),
+                        y.max(0.0f32).min(height_original),
+                        w,
+                        h,
+                        id,
+                        confidence,
+                    );
+
+                    // kpts
+                    let y_kpts = {
+                        if let Some(kpts) = kpts {
+                            let mut kpts_ = Vec::new();
+                            // rescale
+                            for i in 0..self.nk() as usize {
+                                let kx = kpts[KPT_STEP * i] / ratio;
+                                let ky = kpts[KPT_STEP * i + 1] / ratio;
+                                let kconf = kpts[KPT_STEP * i + 2];
+                                if kconf < self.kconf {
+                                    kpts_.push(Point2::default());
+                                } else {
+                                    kpts_.push(Point2::new_with_conf(
+                                        kx.max(0.0f32).min(width_original),
+                                        ky.max(0.0f32).min(height_original),
+                                        kconf,
+                                    ));
+                                }
+                            }
+                            Some(kpts_)
+                        } else {
+                            None
+                        }
+                    };
+
+                    // data merged
+                    data.push((y_bbox, y_kpts, coefs));
+                }
+
+                // nms
+                non_max_suppression(&mut data, self.iou);
+
+                // decode
+                let mut y_bboxes: Vec<Bbox> = Vec::new();
+                let mut y_kpts: Vec<Vec<Point2>> = Vec::new();
+                let mut y_masks: Vec<Vec<u8>> = Vec::new();
+                for elem in data.into_iter() {
+                    if let Some(kpts) = elem.1 {
+                        y_kpts.push(kpts)
+                    }
+
+                    // decode masks
+                    if let Some(coefs) = elem.2 {
+                        let proto = protos.unwrap().slice(s![idx, .., .., ..]);
+                        let (nm, nh, nw) = proto.dim();
+
+                        // coefs * proto -> mask
+                        let coefs = Array::from_shape_vec((1, nm), coefs)?; // (n, nm)
+                        let proto = proto.to_owned().into_shape((nm, nh * nw))?; // (nm, nh*nw)
+                        let mask = coefs.dot(&proto).into_shape((nh, nw, 1))?; // (nh, nw, n)
+
+                        // build image from ndarray
+                        let mask_im: ImageBuffer<image::Luma<_>, Vec<f32>> =
+                            match ImageBuffer::from_raw(nw as u32, nh as u32, mask.into_raw_vec()) {
+                                Some(image) => image,
+                                None => panic!("can not create image from ndarray"),
+                            };
+                        let mut mask_im = image::DynamicImage::from(mask_im); // -> dyn
+
+                        // rescale masks
+                        let (_, w_mask, h_mask) =
+                            self.scale_wh(width_original, height_original, nw as f32, nh as f32);
+                        let mask_cropped = mask_im.crop(0, 0, w_mask as u32, h_mask as u32);
+                        let mask_original = mask_cropped.resize_exact(
+                            // resize_to_fill
+                            width_original as u32,
+                            height_original as u32,
+                            match self.task() {
+                                YOLOTask::Segment => image::imageops::FilterType::CatmullRom,
+                                _ => image::imageops::FilterType::Triangle,
+                            },
+                        );
+
+                        // crop-mask with bbox
+                        let mut mask_original_cropped = mask_original.into_luma8();
+                        for y in 0..height_original as usize {
+                            for x in 0..width_original as usize {
+                                if x < elem.0.xmin() as usize
+                                    || x > elem.0.xmax() as usize
+                                    || y < elem.0.ymin() as usize
+                                    || y > elem.0.ymax() as usize
+                                {
+                                    mask_original_cropped.put_pixel(
+                                        x as u32,
+                                        y as u32,
+                                        image::Luma([0u8]),
+                                    );
+                                }
+                            }
+                        }
+                        y_masks.push(mask_original_cropped.into_raw());
+                    }
+                    y_bboxes.push(elem.0);
+                }
+
+                // save each result
+                let y = YOLOResult {
+                    probs: None,
+                    bboxes: if !y_bboxes.is_empty() {
+                        Some(y_bboxes)
+                    } else {
+                        None
+                    },
+                    keypoints: if !y_kpts.is_empty() {
+                        Some(y_kpts)
+                    } else {
+                        None
+                    },
+                    masks: if !y_masks.is_empty() {
+                        Some(y_masks)
+                    } else {
+                        None
+                    },
+                };
+                ys.push(y);
+            }
+
+            Ok(ys)
+        }
+    }
+
+    pub fn plot_and_save(
+        &self,
+        ys: &[YOLOResult],
+        xs0: &[DynamicImage],
+        skeletons: Option<&[(usize, usize)]>,
+    ) {
+        // check font then load
+        let font = check_font("Arial.ttf");
+        for (_idb, (img0, y)) in xs0.iter().zip(ys.iter()).enumerate() {
+            let mut img = img0.to_rgb8();
+
+            // draw for classifier
+            if let Some(probs) = y.probs() {
+                for (i, k) in probs.topk(5).iter().enumerate() {
+                    let legend = format!("{} {:.2}%", self.names[k.0], k.1);
+                    let scale = 32;
+                    let legend_size = img.width().max(img.height()) / scale;
+                    let x = img.width() / 20;
+                    let y = img.height() / 20 + i as u32 * legend_size;
+                    imageproc::drawing::draw_text_mut(
+                        &mut img,
+                        image::Rgb([0, 255, 0]),
+                        x as i32,
+                        y as i32,
+                        rusttype::Scale::uniform(legend_size as f32 - 1.),
+                        &font,
+                        &legend,
+                    );
+                }
+            }
+
+            // draw bboxes & keypoints
+            if let Some(bboxes) = y.bboxes() {
+                for (_idx, bbox) in bboxes.iter().enumerate() {
+                    // rect
+                    imageproc::drawing::draw_hollow_rect_mut(
+                        &mut img,
+                        imageproc::rect::Rect::at(bbox.xmin() as i32, bbox.ymin() as i32)
+                            .of_size(bbox.width() as u32, bbox.height() as u32),
+                        image::Rgb(self.color_palette[bbox.id()].into()),
+                    );
+
+                    // text
+                    let legend = format!("{} {:.2}%", self.names[bbox.id()], bbox.confidence());
+                    let scale = 40;
+                    let legend_size = img.width().max(img.height()) / scale;
+                    imageproc::drawing::draw_text_mut(
+                        &mut img,
+                        image::Rgb(self.color_palette[bbox.id()].into()),
+                        bbox.xmin() as i32,
+                        (bbox.ymin() - legend_size as f32) as i32,
+                        rusttype::Scale::uniform(legend_size as f32 - 1.),
+                        &font,
+                        &legend,
+                    );
+                }
+            }
+
+            // draw kpts
+            if let Some(keypoints) = y.keypoints() {
+                for kpts in keypoints.iter() {
+                    for kpt in kpts.iter() {
+                        // filter
+                        if kpt.confidence() < self.kconf {
+                            continue;
+                        }
+
+                        // draw point
+                        imageproc::drawing::draw_filled_circle_mut(
+                            &mut img,
+                            (kpt.x() as i32, kpt.y() as i32),
+                            2,
+                            image::Rgb([0, 255, 0]),
+                        );
+                    }
+
+                    // draw skeleton if has
+                    if let Some(skeletons) = skeletons {
+                        for &(idx1, idx2) in skeletons.iter() {
+                            let kpt1 = &kpts[idx1];
+                            let kpt2 = &kpts[idx2];
+                            if kpt1.confidence() < self.kconf || kpt2.confidence() < self.kconf {
+                                continue;
+                            }
+                            imageproc::drawing::draw_line_segment_mut(
+                                &mut img,
+                                (kpt1.x(), kpt1.y()),
+                                (kpt2.x(), kpt2.y()),
+                                image::Rgb([233, 14, 57]),
+                            );
+                        }
+                    }
+                }
+            }
+
+            // draw mask
+            if let Some(masks) = y.masks() {
+                for (mask, _bbox) in masks.iter().zip(y.bboxes().unwrap().iter()) {
+                    let mask_nd: ImageBuffer<image::Luma<_>, Vec<u8>> =
+                        match ImageBuffer::from_vec(img.width(), img.height(), mask.to_vec()) {
+                            Some(image) => image,
+                            None => panic!("can not crate image from ndarray"),
+                        };
+
+                    for _x in 0..img.width() {
+                        for _y in 0..img.height() {
+                            let mask_p = imageproc::drawing::Canvas::get_pixel(&mask_nd, _x, _y);
+                            if mask_p.0[0] > 0 {
+                                let mut img_p = imageproc::drawing::Canvas::get_pixel(&img, _x, _y);
+                                // img_p.0[2] = self.color_palette[bbox.id()].2 / 2;
+                                // img_p.0[1] = self.color_palette[bbox.id()].1 / 2;
+                                // img_p.0[0] = self.color_palette[bbox.id()].0 / 2;
+                                img_p.0[2] /= 2;
+                                img_p.0[1] = 255 - (255 - img_p.0[2]) / 2;
+                                img_p.0[0] /= 2;
+                                imageproc::drawing::Canvas::draw_pixel(&mut img, _x, _y, img_p)
+                            }
+                        }
+                    }
+                }
+            }
+
+            // mkdir and save
+            let mut runs = PathBuf::from("runs");
+            if !runs.exists() {
+                std::fs::create_dir_all(&runs).unwrap();
+            }
+            runs.push(gen_time_string("-"));
+            let saveout = format!("{}.jpg", runs.to_str().unwrap());
+            let _ = img.save(saveout);
+        }
+    }
+
+    pub fn summary(&self) {
+        println!(
+            "\nSummary:\n\
+            > Task: {:?}{}\n\
+            > EP: {:?} {}\n\
+            > Dtype: {:?}\n\
+            > Batch: {} ({}), Height: {} ({}), Width: {} ({})\n\
+            > nc: {} nk: {}, nm: {}, conf: {}, kconf: {}, iou: {}\n\
+            ",
+            self.task(),
+            match self.engine.author().zip(self.engine.version()) {
+                Some((author, ver)) => format!(" ({} {})", author, ver),
+                None => String::from(""),
+            },
+            self.engine.ep(),
+            if let OrtEP::Cpu = self.engine.ep() {
+                ""
+            } else {
+                "(May still fall back to CPU)"
+            },
+            self.engine.dtype(),
+            self.batch(),
+            if self.engine.is_batch_dynamic() {
+                "Dynamic"
+            } else {
+                "Const"
+            },
+            self.height(),
+            if self.engine.is_height_dynamic() {
+                "Dynamic"
+            } else {
+                "Const"
+            },
+            self.width(),
+            if self.engine.is_width_dynamic() {
+                "Dynamic"
+            } else {
+                "Const"
+            },
+            self.nc(),
+            self.nk(),
+            self.nm(),
+            self.conf,
+            self.kconf,
+            self.iou,
+        );
+    }
+
+    pub fn engine(&self) -> &OrtBackend {
+        &self.engine
+    }
+
+    pub fn conf(&self) -> f32 {
+        self.conf
+    }
+
+    pub fn set_conf(&mut self, val: f32) {
+        self.conf = val;
+    }
+
+    pub fn conf_mut(&mut self) -> &mut f32 {
+        &mut self.conf
+    }
+
+    pub fn kconf(&self) -> f32 {
+        self.kconf
+    }
+
+    pub fn iou(&self) -> f32 {
+        self.iou
+    }
+
+    pub fn task(&self) -> &YOLOTask {
+        &self.task
+    }
+
+    pub fn batch(&self) -> u32 {
+        self.batch
+    }
+
+    pub fn width(&self) -> u32 {
+        self.width
+    }
+
+    pub fn height(&self) -> u32 {
+        self.height
+    }
+
+    pub fn nc(&self) -> u32 {
+        self.nc
+    }
+
+    pub fn nk(&self) -> u32 {
+        self.nk
+    }
+
+    pub fn nm(&self) -> u32 {
+        self.nm
+    }
+
+    pub fn names(&self) -> &Vec<String> {
+        &self.names
+    }
+}
--- a/examples/YOLOv8-ONNXRuntime-Rust/src/ort_backend.rs
+++ b/examples/YOLOv8-ONNXRuntime-Rust/src/ort_backend.rs
@ -0,0 +1,534 @@
+use anyhow::Result;
+use clap::ValueEnum;
+use half::f16;
+use ndarray::{Array, CowArray, IxDyn};
+use ort::execution_providers::{CUDAExecutionProviderOptions, TensorRTExecutionProviderOptions};
+use ort::tensor::TensorElementDataType;
+use ort::{Environment, ExecutionProvider, Session, SessionBuilder, Value};
+use regex::Regex;
+
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, ValueEnum)]
+pub enum YOLOTask {
+    // YOLO tasks
+    Classify,
+    Detect,
+    Pose,
+    Segment,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord)]
+pub enum OrtEP {
+    // ONNXRuntime execution provider
+    Cpu,
+    Cuda(u32),
+    Trt(u32),
+}
+
+#[derive(Debug)]
+pub struct Batch {
+    pub opt: u32,
+    pub min: u32,
+    pub max: u32,
+}
+
+impl Default for Batch {
+    fn default() -> Self {
+        Self {
+            opt: 1,
+            min: 1,
+            max: 1,
+        }
+    }
+}
+
+#[derive(Debug, Default)]
+pub struct OrtInputs {
+    // ONNX model inputs attrs
+    pub shapes: Vec<Vec<i32>>,
+    pub dtypes: Vec<TensorElementDataType>,
+    pub names: Vec<String>,
+    pub sizes: Vec<Vec<u32>>,
+}
+
+impl OrtInputs {
+    pub fn new(session: &Session) -> Self {
+        let mut shapes = Vec::new();
+        let mut dtypes = Vec::new();
+        let mut names = Vec::new();
+        for i in session.inputs.iter() {
+            let shape: Vec<i32> = i
+                .dimensions()
+                .map(|x| if let Some(x) = x { x as i32 } else { -1i32 })
+                .collect();
+            shapes.push(shape);
+            dtypes.push(i.input_type);
+            names.push(i.name.clone());
+        }
+        Self {
+            shapes,
+            dtypes,
+            names,
+            ..Default::default()
+        }
+    }
+}
+
+#[derive(Debug)]
+pub struct OrtConfig {
+    // ORT config
+    pub f: String,
+    pub task: Option<YOLOTask>,
+    pub ep: OrtEP,
+    pub trt_fp16: bool,
+    pub batch: Batch,
+    pub image_size: (Option<u32>, Option<u32>),
+}
+
+#[derive(Debug)]
+pub struct OrtBackend {
+    // ORT engine
+    session: Session,
+    task: YOLOTask,
+    ep: OrtEP,
+    batch: Batch,
+    inputs: OrtInputs,
+}
+
+impl OrtBackend {
+    pub fn build(args: OrtConfig) -> Result<Self> {
+        // build env & session
+        let env = Environment::builder()
+            .with_name("YOLOv8")
+            .with_log_level(ort::LoggingLevel::Verbose)
+            .build()?
+            .into_arc();
+        let session = SessionBuilder::new(&env)?.with_model_from_file(&args.f)?;
+
+        // get inputs
+        let mut inputs = OrtInputs::new(&session);
+
+        // batch size
+        let mut batch = args.batch;
+        let batch = if inputs.shapes[0][0] == -1 {
+            batch
+        } else {
+            assert_eq!(
+                inputs.shapes[0][0] as u32, batch.opt,
+                "Expected batch size: {}, got {}. Try using `--batch {}`.",
+                inputs.shapes[0][0] as u32, batch.opt, inputs.shapes[0][0] as u32
+            );
+            batch.opt = inputs.shapes[0][0] as u32;
+            batch
+        };
+
+        // input size: height and width
+        let height = if inputs.shapes[0][2] == -1 {
+            match args.image_size.0 {
+                Some(height) => height,
+                None => panic!("Failed to get model height. Make it explicit with `--height`"),
+            }
+        } else {
+            inputs.shapes[0][2] as u32
+        };
+        let width = if inputs.shapes[0][3] == -1 {
+            match args.image_size.1 {
+                Some(width) => width,
+                None => panic!("Failed to get model width. Make it explicit with `--width`"),
+            }
+        } else {
+            inputs.shapes[0][3] as u32
+        };
+        inputs.sizes.push(vec![height, width]);
+
+        // build provider
+        let (ep, provider) = match args.ep {
+            OrtEP::Cuda(device_id) => Self::set_ep_cuda(device_id),
+            OrtEP::Trt(device_id) => Self::set_ep_trt(device_id, args.trt_fp16, &batch, &inputs),
+            _ => (OrtEP::Cpu, ExecutionProvider::CPU(Default::default())),
+        };
+
+        // build session again with the new provider
+        let session = SessionBuilder::new(&env)?
+            // .with_optimization_level(ort::GraphOptimizationLevel::Level3)?
+            .with_execution_providers([provider])?
+            .with_model_from_file(args.f)?;
+
+        // task: using given one or guessing
+        let task = match args.task {
+            Some(task) => task,
+            None => match session.metadata() {
+                Err(_) => panic!("No metadata found. Try making it explicit by `--task`"),
+                Ok(metadata) => match metadata.custom("task") {
+                    Err(_) => panic!("Can not get custom value. Try making it explicit by `--task`"),
+                    Ok(value) => match value {
+                        None => panic!("No correspoing value of `task` found in metadata. Make it explicit by `--task`"),
+                        Some(task) => match task.as_str() {
+                            "classify" => YOLOTask::Classify,
+                            "detect" => YOLOTask::Detect,
+                            "pose" => YOLOTask::Pose,
+                            "segment" => YOLOTask::Segment,
+                            x => todo!("{:?} is not supported for now!", x),
+                        },
+                    },
+                },
+            },
+        };
+
+        Ok(Self {
+            session,
+            task,
+            ep,
+            batch,
+            inputs,
+        })
+    }
+
+    pub fn fetch_inputs_from_session(
+        session: &Session,
+    ) -> (Vec<Vec<i32>>, Vec<TensorElementDataType>, Vec<String>) {
+        // get inputs attrs from ONNX model
+        let mut shapes = Vec::new();
+        let mut dtypes = Vec::new();
+        let mut names = Vec::new();
+        for i in session.inputs.iter() {
+            let shape: Vec<i32> = i
+                .dimensions()
+                .map(|x| if let Some(x) = x { x as i32 } else { -1i32 })
+                .collect();
+            shapes.push(shape);
+            dtypes.push(i.input_type);
+            names.push(i.name.clone());
+        }
+        (shapes, dtypes, names)
+    }
+
+    pub fn set_ep_cuda(device_id: u32) -> (OrtEP, ExecutionProvider) {
+        // set CUDA
+        if ExecutionProvider::CUDA(Default::default()).is_available() {
+            (
+                OrtEP::Cuda(device_id),
+                ExecutionProvider::CUDA(CUDAExecutionProviderOptions {
+                    device_id,
+                    ..Default::default()
+                }),
+            )
+        } else {
+            println!("> CUDA is not available! Using CPU.");
+            (OrtEP::Cpu, ExecutionProvider::CPU(Default::default()))
+        }
+    }
+
+    pub fn set_ep_trt(
+        device_id: u32,
+        fp16: bool,
+        batch: &Batch,
+        inputs: &OrtInputs,
+    ) -> (OrtEP, ExecutionProvider) {
+        // set TensorRT
+        if ExecutionProvider::TensorRT(Default::default()).is_available() {
+            let (height, width) = (inputs.sizes[0][0], inputs.sizes[0][1]);
+
+            // dtype match checking
+            if inputs.dtypes[0] == TensorElementDataType::Float16 && !fp16 {
+                panic!(
+                    "Dtype mismatch! Expected: Float32, got: {:?}. You should use `--fp16`",
+                    inputs.dtypes[0]
+                );
+            }
+
+            // dynamic shape: input_tensor_1:dim_1xdim_2x...,input_tensor_2:dim_3xdim_4x...,...
+            let mut opt_string = String::new();
+            let mut min_string = String::new();
+            let mut max_string = String::new();
+            for name in inputs.names.iter() {
+                let s_opt = format!("{}:{}x3x{}x{},", name, batch.opt, height, width);
+                let s_min = format!("{}:{}x3x{}x{},", name, batch.min, height, width);
+                let s_max = format!("{}:{}x3x{}x{},", name, batch.max, height, width);
+                opt_string.push_str(s_opt.as_str());
+                min_string.push_str(s_min.as_str());
+                max_string.push_str(s_max.as_str());
+            }
+            let _ = opt_string.pop();
+            let _ = min_string.pop();
+            let _ = max_string.pop();
+            (
+                OrtEP::Trt(device_id),
+                ExecutionProvider::TensorRT(TensorRTExecutionProviderOptions {
+                    device_id,
+                    fp16_enable: fp16,
+                    timing_cache_enable: true,
+                    profile_min_shapes: min_string,
+                    profile_max_shapes: max_string,
+                    profile_opt_shapes: opt_string,
+                    ..Default::default()
+                }),
+            )
+        } else {
+            println!("> TensorRT is not available! Try using CUDA...");
+            Self::set_ep_cuda(device_id)
+        }
+    }
+
+    pub fn fetch_from_metadata(&self, key: &str) -> Option<String> {
+        // fetch value from onnx model file by key
+        match self.session.metadata() {
+            Err(_) => None,
+            Ok(metadata) => match metadata.custom(key) {
+                Err(_) => None,
+                Ok(value) => value,
+            },
+        }
+    }
+
+    pub fn run(&self, xs: Array<f32, IxDyn>, profile: bool) -> Result<Vec<Array<f32, IxDyn>>> {
+        // ORT inference
+        match self.dtype() {
+            TensorElementDataType::Float16 => self.run_fp16(xs, profile),
+            TensorElementDataType::Float32 => self.run_fp32(xs, profile),
+            _ => todo!(),
+        }
+    }
+
+    pub fn run_fp16(&self, xs: Array<f32, IxDyn>, profile: bool) -> Result<Vec<Array<f32, IxDyn>>> {
+        // f32->f16
+        let t = std::time::Instant::now();
+        let xs = xs.mapv(f16::from_f32);
+        if profile {
+            println!("[ORT f32->f16]: {:?}", t.elapsed());
+        }
+
+        // h2d
+        let t = std::time::Instant::now();
+        let xs = CowArray::from(xs);
+        let xs = vec![Value::from_array(self.session.allocator(), &xs)?];
+        if profile {
+            println!("[ORT H2D]: {:?}", t.elapsed());
+        }
+
+        // run
+        let t = std::time::Instant::now();
+        let ys = self.session.run(xs)?;
+        if profile {
+            println!("[ORT Inference]: {:?}", t.elapsed());
+        }
+
+        // d2h
+        Ok(ys
+            .iter()
+            .map(|x| {
+                // d2h
+                let t = std::time::Instant::now();
+                let x = x.try_extract::<_>().unwrap().view().clone().into_owned();
+                if profile {
+                    println!("[ORT D2H]: {:?}", t.elapsed());
+                }
+
+                // f16->f32
+                let t_ = std::time::Instant::now();
+                let x = x.mapv(f16::to_f32);
+                if profile {
+                    println!("[ORT f16->f32]: {:?}", t_.elapsed());
+                }
+                x
+            })
+            .collect::<Vec<Array<_, _>>>())
+    }
+
+    pub fn run_fp32(&self, xs: Array<f32, IxDyn>, profile: bool) -> Result<Vec<Array<f32, IxDyn>>> {
+        // h2d
+        let t = std::time::Instant::now();
+        let xs = CowArray::from(xs);
+        let xs = vec![Value::from_array(self.session.allocator(), &xs)?];
+        if profile {
+            println!("[ORT H2D]: {:?}", t.elapsed());
+        }
+
+        // run
+        let t = std::time::Instant::now();
+        let ys = self.session.run(xs)?;
+        if profile {
+            println!("[ORT Inference]: {:?}", t.elapsed());
+        }
+
+        // d2h
+        Ok(ys
+            .iter()
+            .map(|x| {
+                let t = std::time::Instant::now();
+                let x = x.try_extract::<_>().unwrap().view().clone().into_owned();
+                if profile {
+                    println!("[ORT D2H]: {:?}", t.elapsed());
+                }
+                x
+            })
+            .collect::<Vec<Array<_, _>>>())
+    }
+
+    pub fn output_shapes(&self) -> Vec<Vec<i32>> {
+        let mut shapes = Vec::new();
+        for o in &self.session.outputs {
+            let shape: Vec<_> = o
+                .dimensions()
+                .map(|x| if let Some(x) = x { x as i32 } else { -1i32 })
+                .collect();
+            shapes.push(shape);
+        }
+        shapes
+    }
+
+    pub fn output_dtypes(&self) -> Vec<TensorElementDataType> {
+        let mut dtypes = Vec::new();
+        self.session
+            .outputs
+            .iter()
+            .for_each(|x| dtypes.push(x.output_type));
+        dtypes
+    }
+
+    pub fn input_shapes(&self) -> &Vec<Vec<i32>> {
+        &self.inputs.shapes
+    }
+
+    pub fn input_names(&self) -> &Vec<String> {
+        &self.inputs.names
+    }
+
+    pub fn input_dtypes(&self) -> &Vec<TensorElementDataType> {
+        &self.inputs.dtypes
+    }
+
+    pub fn dtype(&self) -> TensorElementDataType {
+        self.input_dtypes()[0]
+    }
+
+    pub fn height(&self) -> u32 {
+        self.inputs.sizes[0][0]
+    }
+
+    pub fn width(&self) -> u32 {
+        self.inputs.sizes[0][1]
+    }
+
+    pub fn is_height_dynamic(&self) -> bool {
+        self.input_shapes()[0][2] == -1
+    }
+
+    pub fn is_width_dynamic(&self) -> bool {
+        self.input_shapes()[0][3] == -1
+    }
+
+    pub fn batch(&self) -> u32 {
+        self.batch.opt
+    }
+
+    pub fn is_batch_dynamic(&self) -> bool {
+        self.input_shapes()[0][0] == -1
+    }
+
+    pub fn ep(&self) -> &OrtEP {
+        &self.ep
+    }
+
+    pub fn task(&self) -> YOLOTask {
+        self.task.clone()
+    }
+
+    pub fn names(&self) -> Option<Vec<String>> {
+        // class names, metadata parsing
+        // String format: `{0: 'person', 1: 'bicycle', 2: 'sports ball', ..., 27: "yellow_lady's_slipper"}`
+        match self.fetch_from_metadata("names") {
+            Some(names) => {
+                let re = Regex::new(r#"(['"])([-()\w '"]+)(['"])"#).unwrap();
+                let mut names_ = vec![];
+                for (_, [_, name, _]) in re.captures_iter(&names).map(|x| x.extract()) {
+                    names_.push(name.to_string());
+                }
+                Some(names_)
+            }
+            None => None,
+        }
+    }
+
+    pub fn nk(&self) -> Option<u32> {
+        // num_keypoints, metadata parsing: String `nk` in onnx model: `[17, 3]`
+        match self.fetch_from_metadata("kpt_shape") {
+            None => None,
+            Some(kpt_string) => {
+                let re = Regex::new(r"([0-9]+), ([0-9]+)").unwrap();
+                let caps = re.captures(&kpt_string).unwrap();
+                Some(caps.get(1).unwrap().as_str().parse::<u32>().unwrap())
+            }
+        }
+    }
+
+    pub fn nc(&self) -> Option<u32> {
+        // num_classes
+        match self.names() {
+            // by names
+            Some(names) => Some(names.len() as u32),
+            None => match self.task() {
+                // by task calculation
+                YOLOTask::Classify => Some(self.output_shapes()[0][1] as u32),
+                YOLOTask::Detect => {
+                    if self.output_shapes()[0][1] == -1 {
+                        None
+                    } else {
+                        // cxywhclss
+                        Some(self.output_shapes()[0][1] as u32 - 4)
+                    }
+                }
+                YOLOTask::Pose => {
+                    match self.nk() {
+                        None => None,
+                        Some(nk) => {
+                            if self.output_shapes()[0][1] == -1 {
+                                None
+                            } else {
+                                // cxywhclss3*kpt
+                                Some(self.output_shapes()[0][1] as u32 - 4 - 3 * nk)
+                            }
+                        }
+                    }
+                }
+                YOLOTask::Segment => {
+                    if self.output_shapes()[0][1] == -1 {
+                        None
+                    } else {
+                        // cxywhclssnm
+                        Some((self.output_shapes()[0][1] - self.output_shapes()[1][1]) as u32 - 4)
+                    }
+                }
+            },
+        }
+    }
+
+    pub fn nm(&self) -> Option<u32> {
+        // num_masks
+        match self.task() {
+            YOLOTask::Segment => Some(self.output_shapes()[1][1] as u32),
+            _ => None,
+        }
+    }
+
+    pub fn na(&self) -> Option<u32> {
+        // num_anchors
+        match self.task() {
+            YOLOTask::Segment | YOLOTask::Detect | YOLOTask::Pose => {
+                if self.output_shapes()[0][2] == -1 {
+                    None
+                } else {
+                    Some(self.output_shapes()[0][2] as u32)
+                }
+            }
+            _ => None,
+        }
+    }
+
+    pub fn author(&self) -> Option<String> {
+        self.fetch_from_metadata("author")
+    }
+
+    pub fn version(&self) -> Option<String> {
+        self.fetch_from_metadata("version")
+    }
+}
--- a/examples/YOLOv8-ONNXRuntime-Rust/src/yolo_result.rs
+++ b/examples/YOLOv8-ONNXRuntime-Rust/src/yolo_result.rs
@ -0,0 +1,235 @@
+use ndarray::{Array, Axis, IxDyn};
+
+#[derive(Clone, PartialEq, Default)]
+pub struct YOLOResult {
+    // YOLO tasks results of an image
+    pub probs: Option<Embedding>,
+    pub bboxes: Option<Vec<Bbox>>,
+    pub keypoints: Option<Vec<Vec<Point2>>>,
+    pub masks: Option<Vec<Vec<u8>>>,
+}
+
+impl std::fmt::Debug for YOLOResult {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("YOLOResult")
+            .field(
+                "Probs(top5)",
+                &format_args!("{:?}", self.probs().map(|probs| probs.topk(5))),
+            )
+            .field("Bboxes", &self.bboxes)
+            .field("Keypoints", &self.keypoints)
+            .field(
+                "Masks",
+                &format_args!("{:?}", self.masks().map(|masks| masks.len())),
+            )
+            .finish()
+    }
+}
+
+impl YOLOResult {
+    pub fn new(
+        probs: Option<Embedding>,
+        bboxes: Option<Vec<Bbox>>,
+        keypoints: Option<Vec<Vec<Point2>>>,
+        masks: Option<Vec<Vec<u8>>>,
+    ) -> Self {
+        Self {
+            probs,
+            bboxes,
+            keypoints,
+            masks,
+        }
+    }
+
+    pub fn probs(&self) -> Option<&Embedding> {
+        self.probs.as_ref()
+    }
+
+    pub fn keypoints(&self) -> Option<&Vec<Vec<Point2>>> {
+        self.keypoints.as_ref()
+    }
+
+    pub fn masks(&self) -> Option<&Vec<Vec<u8>>> {
+        self.masks.as_ref()
+    }
+
+    pub fn bboxes(&self) -> Option<&Vec<Bbox>> {
+        self.bboxes.as_ref()
+    }
+
+    pub fn bboxes_mut(&mut self) -> Option<&mut Vec<Bbox>> {
+        self.bboxes.as_mut()
+    }
+}
+
+#[derive(Debug, PartialEq, Clone, Default)]
+pub struct Point2 {
+    // A point2d with x, y, conf
+    x: f32,
+    y: f32,
+    confidence: f32,
+}
+
+impl Point2 {
+    pub fn new_with_conf(x: f32, y: f32, confidence: f32) -> Self {
+        Self { x, y, confidence }
+    }
+
+    pub fn new(x: f32, y: f32) -> Self {
+        Self {
+            x,
+            y,
+            ..Default::default()
+        }
+    }
+
+    pub fn x(&self) -> f32 {
+        self.x
+    }
+
+    pub fn y(&self) -> f32 {
+        self.y
+    }
+
+    pub fn confidence(&self) -> f32 {
+        self.confidence
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Default)]
+pub struct Embedding {
+    // An float32 n-dims tensor
+    data: Array<f32, IxDyn>,
+}
+
+impl Embedding {
+    pub fn new(data: Array<f32, IxDyn>) -> Self {
+        Self { data }
+    }
+
+    pub fn data(&self) -> &Array<f32, IxDyn> {
+        &self.data
+    }
+
+    pub fn topk(&self, k: usize) -> Vec<(usize, f32)> {
+        let mut probs = self
+            .data
+            .iter()
+            .enumerate()
+            .map(|(a, b)| (a, *b))
+            .collect::<Vec<_>>();
+        probs.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
+        let mut topk = Vec::new();
+        for &(id, confidence) in probs.iter().take(k) {
+            topk.push((id, confidence));
+        }
+        topk
+    }
+
+    pub fn norm(&self) -> Array<f32, IxDyn> {
+        let std_ = self.data.mapv(|x| x * x).sum_axis(Axis(0)).mapv(f32::sqrt);
+        self.data.clone() / std_
+    }
+
+    pub fn top1(&self) -> (usize, f32) {
+        self.topk(1)[0]
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Default)]
+pub struct Bbox {
+    // a bounding box around an object
+    xmin: f32,
+    ymin: f32,
+    width: f32,
+    height: f32,
+    id: usize,
+    confidence: f32,
+}
+
+impl Bbox {
+    pub fn new_from_xywh(xmin: f32, ymin: f32, width: f32, height: f32) -> Self {
+        Self {
+            xmin,
+            ymin,
+            width,
+            height,
+            ..Default::default()
+        }
+    }
+
+    pub fn new(xmin: f32, ymin: f32, width: f32, height: f32, id: usize, confidence: f32) -> Self {
+        Self {
+            xmin,
+            ymin,
+            width,
+            height,
+            id,
+            confidence,
+        }
+    }
+
+    pub fn width(&self) -> f32 {
+        self.width
+    }
+
+    pub fn height(&self) -> f32 {
+        self.height
+    }
+
+    pub fn xmin(&self) -> f32 {
+        self.xmin
+    }
+
+    pub fn ymin(&self) -> f32 {
+        self.ymin
+    }
+
+    pub fn xmax(&self) -> f32 {
+        self.xmin + self.width
+    }
+
+    pub fn ymax(&self) -> f32 {
+        self.ymin + self.height
+    }
+
+    pub fn tl(&self) -> Point2 {
+        Point2::new(self.xmin, self.ymin)
+    }
+
+    pub fn br(&self) -> Point2 {
+        Point2::new(self.xmax(), self.ymax())
+    }
+
+    pub fn cxcy(&self) -> Point2 {
+        Point2::new(self.xmin + self.width / 2., self.ymin + self.height / 2.)
+    }
+
+    pub fn id(&self) -> usize {
+        self.id
+    }
+
+    pub fn confidence(&self) -> f32 {
+        self.confidence
+    }
+
+    pub fn area(&self) -> f32 {
+        self.width * self.height
+    }
+
+    pub fn intersection_area(&self, another: &Bbox) -> f32 {
+        let l = self.xmin.max(another.xmin);
+        let r = (self.xmin + self.width).min(another.xmin + another.width);
+        let t = self.ymin.max(another.ymin);
+        let b = (self.ymin + self.height).min(another.ymin + another.height);
+        (r - l + 1.).max(0.) * (b - t + 1.).max(0.)
+    }
+
+    pub fn union(&self, another: &Bbox) -> f32 {
+        self.area() + another.area() - self.intersection_area(another)
+    }
+
+    pub fn iou(&self, another: &Bbox) -> f32 {
+        self.intersection_area(another) / self.union(another)
+    }
+}