Support Converting TF dialect to HLO dialect entirely? about mlir-hlo HOT 4 CLOSED

You version of Resnet seems to be using reference variable, I don't think these have even been supported in TensorFlow when using XLA. They also have been deprecated in TensorFlow for many years, any chance you could be using resource variables instead?

from mlir-hlo.

Yes, it is okay to use resource variables. Could you please give me a link? Now I just would like to demonstrate the translation from the a TF model to MHLO dialect, I thought it was ready before, but it seems not to work now. After MHLO, more work will be focused on the MLIR infrastructure.

Alternatively, do you have any suggestion of how to port to MLIR from the AI framework, such as TensorFlow and PyTorch?

from mlir-hlo.

Interfacing from PyTorch is happening in a dedicated project that you can track/participate in: https://github.com/llvm/torch-mlir

It is early for that but they have e2e full models for ResNet and BERT either working or close to working. If you ask there, they can likely give you some commands to run easily to get what you need (with the caveat that it is end of year holidays for a lot of the contributors right now).

from mlir-hlo.

@stellaraccident Many thanks for your reply, quite helpful.

from mlir-hlo.