ankit-jha/kfp

ML Ops Code Structure

ml

├── README.md

├── bert_pipeline.py Pipeline that runs all the kubeflow components

├── components Different kubeflow components as per the functionality

    ├── deploy        Components for deploying to serving endpoint

    ├── evaluate      Components for capturing metadata  

    ├── preprocess    Data Preprocessing component

    ├── storage       Component for moving files from/to GCS 

    └── train         Training component code. e.g BERT training 

├── loadtest Load testing using Jmeter

├── data Sample representative data for testing

├── notebooks Exploration Notebooks

├── pipelines Eventually all pipelines will be moved here

├── python_package Houses python package for training models

└── requirements.txt

Installing programs

• Create a virtual environment

python3 -m venv <virtualenv-name>

• We should update pip, Python's dependency manager:

pip install --upgrade pip

• Install the packages from requirements.txt

pip3 install -r requirements.txt

NOTE : If an error message of Installation Error - Failed building wheel for tokenizers then please install rust compiler and re-run the installation command above.

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
source $HOME/.cargo/env

How to run the pipeline

Vertex AI

You'll need to authenticate against a service account that has access to Vertex Ai and Vertex AI Service Agent

gcloud auth <service-account>  --key-file=<key-file>

You'll need to set the below variables to run the pipeline.

export PROJECT_ID="mlops"
export REGION="us-central1"
export BUCKET_NAME="ankit-test-pipeline"
export BERT_TRAINING_DATA_BUCKET="ankit-test-data"
export BERT_TRAINING_DATA="reviews.csv"
export MODEL_NAME="finetuned-bert-classifier"

python bert_pipeline.py

Manual steps that still need automation

1. Python Packaging Python Package needs to be uploaded to the bucket before training. For automation we need a pipeline component to copy the code from git to gcs bucket

    cd python_package;
    tar cvf bert.tar bert; gzip -f bert.tar; gsutil cp bert.tar.gz gs://${BUCKET_NAME}/bert.tar.gz

2. Model archiver needs certain files to be uploaded to the bucket

   gsutil cp ./torch/predictor/custom_handler.py gs://${BUCKET_NAME}/${MODEL_NAME}/serve/predictor/
   gsutil cp ./torch/predictor/index_to_name.json gs://${BUCKET_NAME}/${MODEL_NAME}/serve/predictor/
   gsutil cp ./torch/predictor/Dockerfile.serve gs://${BUCKET_NAME}/${MODEL_NAME}/serve/

Kubeflow

Install Kubeflow