ONNX custom OP uses deprecated functions

[nrakltx]

Ask a Question

Question

I am attempting to implement a custom op using CUDA kernels and started looking into existing guides and how-to's available. The simplest and easiest I could find exists on this repo, here. Notwithstanding this is the main E2E example for exporting custom ops in ORT, it seems to be using deprecated functions; Inside this file, functions such as GetTensorData are deprecated as per this, and such are some other functions from this tutorial.

Further information

• Relevant Area (e.g. model usage, best practices, shape_inference, version_converter, training, test): model inference, custom ops.

• Is this issue related to a specific model?
  Model name (e.g. mnist): - Model opset (e.g. 7): - Please upload the model and provide the link if possible.

Notes

Any additional information, code snippets.

[yuslepukhin]

Would you like to contribute and fix those? It would also be a good learning experience, as well as a good testing ground before the next release. The deprecation of the functions has not been released yet and you encounter these problems only if you build from source code.

Cc: @jcwchen