This project contains simple jupyter notebook lab exercises demonstrating ONNX model conversion. This includes:

• Exporting a simple model from Pytorch to ONNX format.
• Converting a tensorflow model to ONNX, using tensorflow-onnx.

These examples require you install needed packages. This is not done as part of the notebook; the steps here will guide you through building and running a docker container to complete the lab exercises.

This lab is intended to be run on an x86/x64 environment to demonstrate converting a trained model to ONNX prior to deploying on IBM zSystems and LinuxONE.

1. Clone the lab github repository git clone https://github.com/IBM/ai-on-z-samples.git

2. Navigate to the subdirectoy.

3. Run docker build using the provided dockerfile docker build .

   • Using a python base image, this will create an environment with both recent Pytorch and TensorFlow releases, model conversion libraries, as well as the lab jupyter notebooks.

4. Create and run a docker container using the image created on the prior step. As part of this step, you should map the jupyter notebook port, 8888, to a port on your local system. An example follows:

   • docker run -it --rm -p 8571:8888 <image id>
   • This states the image in interactive mode, tells docker to delete the container upon exit, and publishes container port 8888 to host port 8571.

5. From a web browser, connect to the jupyter URL provided on the prior step. Note, you must change port 8888 to port 8571.

6. Run through the lab exercises:

   • tfonnx_conversion.ipynb
   • torch_export_onnx.ipynb

7. Download the .ONNX models and inspect them using Netron

There are numerous additional examples and guidance available, not only for Pytorch and TensorFlow, but for other frameworks as well.

This includes:

• Building, exporting or converting a model here
• tensorflow-onnx examples
• pytorch guidance and examples

On IBM Z and LinuxONE, you can run these models using ONNX-MLIR. For z/OS users, we recommend you try Watson Machine Learning for z/OS Trial edition, available here.