It's non-trivial, certain models like GRU has fundamentally different implementation, so they are not cross-compatible. TF also has no incentive to merge any PR back into its code-base for compatibility, and dragged its feet for months. I have spent a long time investigating into this and hit brick walls after brick walls.
Comments
Might be a stupid question, but what's the advantage of using PyTorch Mobile, compared to converting to ONNX-> TF -> TF lite (or something similar)
It's non-trivial, certain models like GRU has fundamentally different implementation, so they are not cross-compatible. TF also has no incentive to merge any PR back into its code-base for compatibility, and dragged its feet for months. I have spent a long time investigating into this and hit brick walls after brick walls.