YOLO26-pose: Optimized for AMD ROCm
YOLO26-pose is a real-time human pose-estimation model that detects people and localizes 17 COCO body keypoints per person in a single forward pass. This repository packages inference for pose estimation using ONNX Runtime (CPU, NPU) and PyTorch (GPU), exported and validated for AMD ROCm so it runs efficiently on AMD GPUs, CPUs, and NPUs.
This is based on the implementation of YOLO26 found
here.
This repository contains configurations and scripts optimized for
AMD® ROCm™ platforms. You can use the
yolo26_pose AMD scripts to reproduce results or export with custom configurations. More details on model performance can be found
here.
Task Overview
Task: Pose estimation (17 COCO keypoints)
Dataset: COCO-Pose val2017 (a verified human-image sample is staged under datasets/samples/)
Output metrics: Throughput (inferences/sec), latency (mean/P95/P99 ms), per-operator profiling breakdown
Model variants: Default is n (nano). Override with MODEL_SIZE=n/s/m/l/x — note the pose download list ships n/s/m/l weights only, so MODEL_SIZE=x may not resolve to a hosted weight.
GPU note: The GPU path uses the PyTorch/ROCm backend rather than MIGraphX (the ONNX/MIGraphX path hits reduce-operator errors for pose), so all GPU precision targets currently share the same FP32 PyTorch backend (INT8 is a no-op skip).
AMD ROCm Optimization
This model export has been adapted and validated for AMD Instinct™ / Radeon™ GPUs running ROCm, as well as AMD CPUs and AMD Ryzen AI NPUs. Key points:
- Validated backends: ONNX Runtime on CPU (FP32) and NPU (VitisAI — BF16), and PyTorch/ROCm on GPU (FP32).
- No code changes required versus the upstream Ultralytics YOLO26-pose implementation — only environment/runtime configuration differs.
- NPU evaluation is not fully functional yet — some layers have numerical-stability issues on NPU (to be addressed in a future release); NPU eval uses a lowered
--conf-threshold 0.05 to compensate for BF16 precision loss.
| Runtime | Precision | Backend | Hardware | Notes |
|---|
| ONNX Runtime | FP32 | CPU Execution Provider | AMD CPU | — |
| PyTorch | FP32 | ROCm (HIP) | AMD Instinct™ / Radeon™ GPU | FP32/FP16/BF16/INT8 targets all share this backend; INT8 is a no-op skip |
| ONNX Runtime | BF16 | VitisAI Execution Provider | AMD Ryzen AI NPU | Known numerical-stability issues on some layers |
Getting Started
For setup instructions, evaluation scripts, and custom configuration options, see the
yolo26_pose on GitHub.
Model Details
Model Type: Human pose estimation (17 COCO keypoints)
Base Model: YOLO26-pose (Ultralytics), size n (nano) default
Model Stats:
- Input (
images): (1, 3, 640, 640) float32
- Output (
output0): (1, 300, 57) float32 — up to 300 detections × (box[4] + score[1] + class[1] + 17 keypoints × [x, y, visibility])
- Model sizes:
n/s/m/l ship pretrained pose weights (x may not resolve to a hosted weight)
- Precision tested: FP32 (CPU, GPU); BF16 (NPU)
Accuracy Pipeline
Accuracy evaluation is not yet implemented for this model. The staged COCO-Pose sample image is used only for visual sanity checks via annotated skeleton overlays, not full COCO-Pose keypoint accuracy scoring.
Dig Deeper
Want to explore the full evaluation scripts, config options, and other AMD-optimized model examples?
The GitHub repository includes:
- Setup and prerequisites for ROCm environments
- Per-operator latency profiling scripts (including NPU AI Analyzer integration)
- Sample-image evaluation with annotated skeleton overlays
- Benchmarking and reproduction instructions across CPU, GPU, and NPU
