---
library_name: onnx
license: mit
tags:
  - foundation
  - amd
  - rocm
  - image-segmentation
pipeline_tag: image-segmentation
---

![](https://huggingface.co/AMD-PAVS-AI/hrnetv2/resolve/main/hrnetv2.png)

# HRNetv2-W48: Optimized for AMD ROCm

HRNetv2-W48 is a semantic segmentation model that assigns a class label to every pixel while maintaining high-resolution feature representations throughout the network. This repository packages inference for semantic segmentation using **ONNX Runtime**, exported and validated for **AMD ROCm** so it runs efficiently on AMD GPUs, CPUs, and NPUs.

This is based on the implementation of HRNetv2 found [here](https://github.com/HRNet/HRNet-Semantic-Segmentation).
This repository contains configurations and scripts optimized for **AMD® ROCm™** platforms. You can use the [hrnetv2 AMD scripts](https://github.com/AMD-PAVS/physical_ai_sdk/blob/main/models/hrnetv2) to reproduce results or export with custom configurations. More details on model performance can be found [here](#accuracy-pipeline).

---

## Task Overview

**Task:** Semantic segmentation

**Dataset:** Cityscapes val (500 images) · ADE20k val · LIP val (variant-specific)

**Output metrics:** mIoU (mean Intersection over Union), pixel accuracy, mean class accuracy

> **Model variants:** Default is **Cityscapes** at 1024×2048 (paper mIoU 80.9%). Override with `HRNET_VARIANT` or `HRNET_MODEL` (e.g. `export HRNET_MODEL=ade20k`) or run `make set-variant VARIANT=MODEL_LIP` for a persistent override. Run `make list-variants` for aliases.

> **NPU note:** NPU float precisions (FP32/FP16/BF16) use per-dtype `config/vitisai_config_*.json`; NPU INT8 requires `make quantize-npu-int8` before benchmark/eval.

---

## 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** across CPU, GPU (MIGraphX execution provider), and NPU (VitisAI execution provider).
- No code changes required versus the upstream HRNet implementation — only environment/runtime configuration differs.
- CPU fallback path supported for environments without a ROCm-capable GPU.

| Runtime | Precision | Backend | Hardware | Notes |
|---|---|---|---|---|
| ONNX Runtime | FP32 / FP16 / BF16 / INT8 | CPU Execution Provider | AMD CPU | — |
| ONNX Runtime | FP32 / FP16 / BF16 / INT8 | MIGraphX Execution Provider | AMD Instinct™ / Radeon™ GPU (ROCm) | — |
| ONNX Runtime | FP32 / FP16 / BF16 / INT8 | VitisAI Execution Provider | AMD Ryzen AI NPU | INT8 requires Quark calibration via `make quantize-npu-int8` |

---

## Getting Started

For setup instructions, evaluation scripts, and custom configuration options, see the [hrnetv2 on GitHub](https://github.com/AMD-PAVS/physical_ai_sdk/blob/main/models/hrnetv2).

---

## Model Details

**Model Type:** Semantic segmentation, HRNetv2-W48

**Base Model:** `HRNet/HRNet-Semantic-Segmentation` (HRNetv2-W48)

**Model Stats:**
- Model variant: Cityscapes, 1024×2048 input (default) — ADE20k (520×520) and LIP (473×473) also supported
- Precision tested: FP32, FP16, BF16, INT8

---

## Accuracy Pipeline

Higher mIoU means predicted pixel labels agree more closely with ground truth — 100% is perfect overlap, 0% is no agreement.

### Metrics Explained

| Metric | Description |
|--------|-------------|
| mIoU | Primary segmentation metric — mean Intersection-over-Union averaged across all classes. Higher means better boundary alignment and class assignment across the validation set. |
| Pixel accuracy | Fraction of correctly labeled pixels — rewards overall coverage but can hide poor performance on rare classes. |
| Mean class accuracy | Average per-class accuracy — exposes imbalance when large classes dominate pixel accuracy. |

### Accuracy Results

**Full Dataset Evaluation (Cityscapes val)** — filled from `evaluation_results/`; run `make metrics` to refresh:

<!-- accuracy-table-start -->
| Device | Backend | Precision | Variant | Accuracy (%) |
|--------|---------|-----------|---------|--------------|
| CPU | ONNX Runtime | FP32 | Cityscapes | 40.54 |
| CPU | ONNX Runtime | FP16 | Cityscapes | 40.75 |
| CPU | ONNX Runtime | BF16 | Cityscapes | 40.82 |
| CPU | ONNX Runtime | INT8 | Cityscapes | 40.99 |
| GPU | ONNX Runtime | FP32 | Cityscapes | 40.86 |
| GPU | ONNX Runtime | FP16 | Cityscapes | 40.63 |
| GPU | ONNX Runtime | BF16 | Cityscapes | 40.67 |
| GPU | ONNX Runtime | INT8 | Cityscapes | 40.82 |
| NPU | ONNX Runtime | FP32 | Cityscapes | 40.49 |
| NPU | ONNX Runtime | FP16 | Cityscapes | 41.10 |
| NPU | ONNX Runtime | BF16 | Cityscapes | 40.92 |
| NPU | ONNX Runtime | INT8 | Cityscapes | 0.00 |
<!-- accuracy-table-end -->

---

## Dig Deeper

Want to explore the full evaluation scripts, config options, and other AMD-optimized model examples?

📂 **[View the full project on GitHub](https://github.com/AMD-PAVS/physical_ai_sdk/blob/main/models/hrnetv2)**

The GitHub repository includes:
- Benchmark, profile, and evaluation Makefile targets for CPU, GPU, and NPU
- Dataset staging for Cityscapes, ADE20k, and LIP variants
- ONNX export scripts and the patched upstream HRNet evaluation tooling
- Additional HRNetv2-W48 variants (Cityscapes, ADE20k, LIP)