I'm trying to deploy MitoNet v1 (a mitochondria segmentation model based on PanopticDeepLab) on the Axelera Metis AIPU using the Voyager SDK v1.5, and I'm running into issues exporting the model to ONNX.
MitoNet is available via the empanada-napari library and its weights are stored as a TorchScript archive (.pth). When loaded, the model is a `torch.jit.RecursiveScriptModule` that takes a (1, 1, H, W) float32 tensor as input and returns a dict with keys `sem_logits`, `ctr_hmp`, and `offsets`.
The problem is that the model uses `aten::__interpolate` with an `int[]` size argument, which blocks all ONNX export paths I've tried:
1. `torch.onnx.export` with opset 11 and 17 → `INTERNAL ASSERT FAILED: We don't have an op for aten::__interpolate`
2. `TS2EPConverter(model, args).convert()` → same `aten::__interpolate` error
3. `axelera.compiler.quantize()` directly on the TorchScript model → `ValueError: Failed to get input shape from calibration dataset`
Environment:
- Voyager SDK v1.5
- PyTorch 2.10.0+cu128 (in axelera-env)
- Python 3.10
- Metis PCIe board, 4GiB, firmware 1.5.0
My questions:
1. Is there a recommended way to export TorchScript models that use `aten::__interpolate` with a list size argument to ONNX for the Axelera compiler?
2. Can the Axelera compiler API handle TorchScript models directly without going through ONNX, and if so what is the correct input format for the calibration dataset?
3. Is there any workaround or patch for the interpolation op issue?
Any help would be greatly appreciated. Happy to share more details or the model config if needed.
Thanks!
