AXERA-TECH/dinov3_vits_pretrain_lvd1689m

DINOv3 for AX650

This repository provides an AX650 deployment package for facebook/dinov3-vits16-pretrain-lvd1689m.

It includes:

• an AX650 .axmodel
• a reference .onnx model
• inference and comparison scripts
• example inputs and outputs

The .axmodel in this package was converted for AX650 with u16 quantization.

Compatible Pulsar2 version: 5.1-patch1

Repository Structure

.
├── models-ax650/   # model files
├── scripts/        # inference and comparison scripts
└── examples/       # sample images and reference outputs

Platform Support

Chip	Model	Latency
AX650	dinov3-vits16-pretrain-lvd1689m	9.841 ms

Quick Start

Install Python dependencies:

pip install -r scripts/requirements.txt
pip install matplotlib

For AXModel inference, install pyaxengine separately:

wget https://github.com/AXERA-TECH/pyaxengine/releases/download/0.1.3.rc2/axengine-0.1.3-py3-none-any.whl
pip install axengine-0.1.3-py3-none-any.whl

Run ONNX inference:

python scripts/infer_dinov3_backbone_onnx.py \
  --model models-ax650/dinov3-vits16-pretrain-lvd1689m.onnx \
  --image examples/ILSVRC2012_val_00000005.jpeg

Run AXModel inference:

python scripts/infer_dinov3_backbone_axmodel.py \
  --model models-ax650/dinov3-vits16-pretrain-lvd1689m.axmodel \
  --image examples/ILSVRC2012_val_00000005.jpeg

Compare ONNX and AXModel outputs:

python scripts/compare_onnx_axmodel_feature_embeddings.py \
  --npy-a examples/onnx_ILSVRC2012_val_00000005.npy \
  --npy-b examples/compiled__ILSVRC2012_val_00000005_output.npy \
  --output-prefix compare_result

Example comparison outputs for examples/ILSVRC2012_val_00000005.jpeg:

Input image:

Comparison plot:

Side-by-side heatmap:

Per-output heatmaps:

For detailed script usage, see:

• English: scripts/README_en.md
• 中文: scripts/README_zh.md

Model Conversion References

If you want to reproduce the conversion pipeline, start from:

• Original DINOv3 repository
• AXERA conversion repository

Notes

• This repository is intended for inference delivery and validation.
• Training and conversion source code are not included here.