Hello!
We would like to deploy MiVOLO, a face-and-body age/gender estimation model, on an Axelera Edge 130p (Metis), which we are planning to use.
Our setup:
- Architecture: MiVOLO / VOLO, a transformer with attention.
- Model: ONNX, opset 17; compatibility adaptations applied, trained weights retained.
- Input: face_body, float32 [1,6,384,384], NCHW.
- Output: gender_age, float32 [1,3].
- Compiler: axelera-devkit 1.9.0, Python compiler API.
- Environment: Google Colab, Linux, Python 3.13.16, PyTorch 2.9.1.
- No hardware testing yet.
We have not obtained a usable compiled model. In one diagnostic run, quantization and export completed, but numerical validation against ONNX Runtime failed (max_abs ≈ 0.279, relative L2 ≈ 1.445), so we did not proceed to full ELF compilation.
Settings: both_symmetric + per_tensor_min_max. We used 9 synthetic calibration inputs and 3 separate evaluation inputs. We understand that representative calibration data is needed.
Is this architecture supported? Could you recommend the correct compilation workflow, an example configuration and calibration requirements for Edge 130p?
Alternatively, is there a supported age/gender model in your model zoo?
We can provide the ONNX model and logs through an agreed channel if needed.
Thank you!
