I'm currently working on deploying the LightGlue model from this repository on the NXP i.MX 93 NPU using the EIQ toolkit. However, I've encountered a challenge: the toolkit supports only ONNX opset 15, which seems not to include support for the GridSample layer used in the PyTorch model.
Could you suggest any alternatives to the GridSample operation that would be compatible with ONNX opset 15, or any other workaround that might help in converting and running this model on the specified NPU?
I am not experienced in ONNX, so I can not really help here. The grid sample in python is basically an interpolation of the descriptor map, you could try to port the interpolation from PixLoc.
Hi,
I'm currently working on deploying the LightGlue model from this repository on the NXP i.MX 93 NPU using the EIQ toolkit. However, I've encountered a challenge: the toolkit supports only ONNX opset 15, which seems not to include support for the GridSample layer used in the PyTorch model.
Could you suggest any alternatives to the GridSample operation that would be compatible with ONNX opset 15, or any other workaround that might help in converting and running this model on the specified NPU?
Cheers Paul