Kubeflow pipelines built on top of Tensorflow TFX library
General info
This repository contains machine learning pipelines based on Tensorflow TFX library. Every pipeline is designed to be published on a Kubernetes/Kubeflow cluster on premise.
Each folder contains needed code and data for the Kubeflow Pipeline, plus a README that includes:
pipeline general information
specific data handling about pipeline on premise
interactive notebooks instructions
build and launch procedure
Further pipelines are welcome via pull request.
Pipelines:
iris - Complete pipeline for a simple (Keras) model on IRIS dataset.
cifar-10 - Complete pipeline for a CNN model on CIFAR-10 dataset [NEEDS UPDATE].
inat-2019 - Complete pipeline for a MobilenetV2 model on iNaturalist 2019 dataset [NEEDS UPDATE].
TFX Custom image
Pipelines are actually using custom TFX images containing NVIDIA drivers for GPU usage from tfx-nvidia-gpu
Prerequisites
Here some prerequisites needed to deploy this repo.
Platform versions
Kubeflow version >=1.0
Tensorflow >=2.1.0
Tensorflow TFX ==0.21.1
Kubernetes cluster
A PersistentVolumeClaim called tfx-pvc is needed so the cluster should have one ready before dropping the pipelines.
Here an example of a 100Gb claim with a local-path storageClass onboard.