SegFormer: Optimized for AMD ROCm
SegFormer is a semantic segmentation model that assigns a class label to every pixel across 19 Cityscapes categories. 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 SegFormer found here. This repository contains configurations and scripts optimized for AMD® ROCm™ platforms. You can use the segformer AMD scripts to reproduce results or export with custom configurations. More details on model performance can be found here.
Task Overview
Task: Semantic segmentation
Dataset: Cityscapes val (500 images, 19 classes)
Output metrics: mIoU (mean Intersection over Union)
Model: SegFormer-B5 only at 1024×1024 — no
MODEL_SIZEvariants.
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 (FP32/FP16/BF16/INT8), GPU (MIGraphX — FP32/FP16/BF16/INT8), and NPU (VitisAI — FP32/FP16/BF16/INT8).
- No code changes required versus the upstream SegFormer implementation — only environment/runtime configuration differs.
- NPU INT8 requires a separate Quark + Cityscapes calibration quantization step before evaluation.
| 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) | First run pays a 30+ minute graph-compilation cost |
| ONNX Runtime | FP32 / FP16 / BF16 / INT8 | VitisAI Execution Provider | AMD Ryzen AI NPU | INT8 requires Quark + Cityscapes calibration quantization step |
Getting Started
For setup instructions, evaluation scripts, and custom configuration options, see the segformer on GitHub.
Model Details
Model Type: Semantic segmentation (Transformer-based)
Base Model: SegFormer-B5, 1024×1024 input resolution
Model Stats:
- Input (
input):(1, 3, 1024, 1024)float32 - Output (
logits):(1, 19, 1024, 1024)float32 - Precision tested: FP32, FP16, BF16, INT8 (CPU/GPU/NPU)
Accuracy Pipeline
Higher mIoU means predicted pixel labels agree more closely with ground truth across all 19 Cityscapes classes — 100% is perfect overlap, 0% is no agreement. Values above ~80% on Cityscapes val are considered strong for this task. Paper mIoU for SegFormer-B5 is 82.4%.
Metrics Explained
| Metric | Description |
|---|---|
| mIoU | Primary segmentation metric — mean Intersection-over-Union averaged across all 19 Cityscapes classes. Higher means better boundary alignment and class assignment across the full val set. |
Accuracy Results
Full Dataset Evaluation (Cityscapes val) — SegFormer-B5:
| Device | Precision | mIoU |
|---|---|---|
| CPU | FP32 | 82.25% |
| GPU | FP32 | 82.25% |
| GPU | FP16 | 82.20% |
| GPU | BF16 | 82.26% |
| GPU | INT8 | 82.25% |
| NPU | FP32 | 82.25% |
| NPU | FP16 | 82.25% |
| NPU | BF16 | 82.23% |
| NPU | INT8 | 1.34% |
Dig Deeper
Want to explore the full evaluation scripts, config options, and other AMD-optimized model examples?
📂 View the full project on GitHub
The GitHub repository includes:
- Setup and prerequisites for ROCm environments
- Cityscapes val dataset staging and mIoU evaluation pipeline via mmsegmentation
- NPU INT8 quantization workflow (AMD Quark + Cityscapes calibration)
- Benchmarking and reproduction instructions across CPU, GPU, and NPU