This project demonstrates an end-to-end MLOps pipeline using Google Cloud's Vertex AI. The pipeline automates data fetching, model training and evaluation, and model deployment on Vertex AI for serving predictions.

• Google Cloud Project: A GCP project with billing enabled.
• Vertex AI API: Ensure that the Vertex AI API and other necessary APIs are enabled in your project.
• Service Account: A service account with the following permissions:
  • Access to BigQuery datasets.
  • Permissions to create and manage Vertex AI resources.
  • Permissions to read and write to Google Cloud Storage.
• Local Environment:
  • Docker and Docker Compose installed.
  • Google Cloud SDK (gcloud CLI) installed and authenticated.
  • jq command-line JSON processor installed.
• Configuration Files:
  • config.json: Copy from config.json.sample and fill in your project-specific details.
  • credential.json: Downloaded service account key file placed in the project root.

The project focuses on automating the lifecycle of a machine learning model for text classification using Vertex AI. It includes:

• Fetching data from BigQuery.
• Training a DistilBERT-based text classification model using Hugging Face Transformers.
• Deploying the trained model to Vertex AI for real-time predictions.
• Building a custom Docker image for model serving.
• Orchestrating the entire pipeline using Kubeflow Pipelines (KFP).

• Dataset: AG News dataset, a collection of news articles labeled into four categories.
• Data Preparation:
  • The prepare_dataset.py script loads and preprocesses the data.
  • Data is uploaded to BigQuery with an added timestamp field to simulate recent data entries.
• Data Fetching:
  • The pipeline fetches data from BigQuery based on a timestamp filter (last two days). We do not need to use BigQuery for this dataset, and it is just for demonstration purposes towards a real-world scenario.

• Architecture: Based on distilbert-base-uncased from Hugging Face Transformers.
• Purpose: Fine-tuned for text classification to categorize news articles.
• Classes: Classifies texts into four categories:
  • World
  • Sports
  • Business
  • Sci/Tech
• Training Details:
  • Tokenization using AutoTokenizer.
  • Training with Trainer API.
  • Evaluation metrics: accuracy, precision, recall, F1 score.
• Artifacts: The trained model, tokenizer, and label encoder are saved for deployment.

cd serving
CONFIG=$(cat ../config.json | jq '.model_tag, .region, .repository_name, .image_name' | sed 's/"//g' | tr '\n' ',')
gcloud builds submit \
    --config cloudbuild.yaml \
    --substitutions _MODEL_VERSION=$(echo $CONFIG | cut -d, -f1),_LOCATION=$(echo $CONFIG | cut -d, -f2),_REPOSITORY_NAME=$(echo $CONFIG | cut -d, -f3),_IMAGE_NAME=$(echo $CONFIG | cut -d, -f4)

For the first time, you need to authenticate with your Google Cloud service account:

gcloud auth activate-service-account <service-account> --key-file=./credential.json
gcloud auth configure-docker <location>-docker.pkg.dev

docker compose build

Here, I will use IMDb dataset for training. The script will register only part of it to Google BigQuery.

docker compose run mlops-v1 python prepare_dataset.py

docker compose run mlops-v1

In my case, the pipeline completed in 1 hour.

I have prepared a sample request script sample-request.sh to test the deployed model.

chmod +x sample-request.sh
./sample-request.sh

The inference output will look like:

{
  "predictions": [
    {
      "probabilities": {
        "Sci/Tech": 0.0032518664374947548,
        "World": 0.98935961723327637,
        "Sports": 0.0034750890918076038,
        "Business": 0.0039134090766310692
      },
      "predicted_class": "World"
    },
    {
      "probabilities": {
        "World": 0.013002258725464341,
        "Sports": 0.97295308113098145,
        "Business": 0.0041450257413089284,
        "Sci/Tech": 0.00989964697510004
      },
      "predicted_class": "Sports"
    },
    {
      "probabilities": {
        "World": 0.0058785984292626381,
        "Sports": 0.0053764986805617809,
        "Sci/Tech": 0.029164601117372509,
        "Business": 0.95958030223846436
      },
      "predicted_class": "Business"
    },
    {
      "predicted_class": "Sci/Tech",
      "probabilities": {
        "Business": 0.018011067062616348,
        "Sports": 0.002777427202090621,
        "Sci/Tech": 0.97608381509780884,
        "World": 0.003127649892121553
      }
    }
  ],
  "deployedModelId": "xxx",
  "model": "projects/xxx/locations/xxx/models/text-classification-model",
  "modelDisplayName": "text-classification-model",
  "modelVersionId": "1"
}