The examples for nvidia-resnet cannot be built using existing scripts.

[a1856315445]

Feature Area

/area backend
/area sdk

The examples for nvidia-resnet cannot be built using existing scripts.

What feature would you like to see?

Update existing nvidia-resnet or build a new version

What is the use case or pain point?

This example is based on Resnet on image classification, however, it can not longer be built becuase:

1. This example is built with the latest tensorRT which is tagged with v2.45.0, however it only can work with v2.6.0 due to code structure chaged after v2.6.0. (https://github.com/triton-inference-server/server/tree/v2.6.0)
2. This example is still using TensorFlow-19.03 containers. https://github.com/kubeflow/pipelines/blob/master/samples/contrib/nvidia-resnet/components/train/Dockerfile
3. webapp_launcher, built on top of Ubuntu16.04, has also not been updated.
   (https://github.com/kubeflow/pipelines/blob/master/samples/contrib/nvidia-resnet/components/webapp_launcher/Dockerfile)
4. Kubeflow pipeline is about to be updated to v2, while this example is still on v1.
   (https://github.com/kubeflow/pipelines/blob/master/samples/contrib/nvidia-resnet/pipeline/Dockerfile)

Due to the aforementioned issues, it's not easy for users (like myself) to encounter difficulties in building this example.

Is there a workaround currently?

Due to the aforementioned issues, it is hard to rebuild this example.

The proposed change list

1. The three items mentioned above: tensorRT SDK, TensorFlow container and webapp_launcher.
2. Replace kfp v1 with v2