Resource Import Failed & ONNX Model
What layers are allowed in snap ml? I see this compatibility list but it is inaccurate for successful imports.
I've been getting all sorts of layer import failures like "Gather" and "Transpose" which I thought are pretty standard for an onnx-able model.
As a follow up, is there a snap ml compatibility list for onnx-able python methods? Something like the onnx op set but for snap ml so that I could know before making my model to avoid matrix indexing (translates to the "Gather" layer that can't be imported to snap ml).
Good question! I had some failures in snap ml too :/
Hi Alec,
Thanks for raising this issue.
1. We are sorry that the compatibility list only included client side inference time supported ops, we will update the list to include supported ops at import-time as well.
2. When exporting Pytorch/TensorFlow models to ONNX, sometimes a single op may be translated to complex patterns that composed by standard ops like "Gather" and "Transpose" layers, and this pattern may be inconsistent across different ONNX opset versions. We are trying our best to support all patterns but our converter may still fail with upstream changes on Pytorch/TensorFlow -> ONNX side.
We are more than willing to help you to get your model imported successfully and it will be great if you could share your model with us
Best,
Eric