Fine-Tuning BERT for Sentiment Analysis and Deployment

Overview

This repo simulates how an ML model moves to production in an industry setting. The goal is to build, deploy, monitor, and retrain a sentiment analysis model using Kubernetes (minikube) and FastAPI.

Note

Models and dataset files are not included in this repository due to size constraints. You can find the dataset at: https://www.kaggle.com/datasets/kazanova/sentiment140

Tasks & Implementation

Running a Basic Sentiment Analysis

• Dataset: Publicly available labeled dataset containing tweets with sentiment labels. https://www.kaggle.com/datasets/kazanova/sentiment140
• Notebook: bert-finetuning-sentiment-analysis-training.ipynb
  • Performed initial dataset analysis and basic cleaning.
  • Used a Hugging Face BERT model for fine-tuning on ~60% of the dataset.
  • Saved trained model artifacts for future use.
  • Notebook was run on Kaggle, and model files were recovered from persisted files.
• Location: dataset/

Wrapping the Model as an API

• Folder: inference/
• Framework: FastAPI
• Functionality:
  • API accepts a tweet as input and returns sentiment prediction along with confidence scores. It run at endpoint '/predict' as post request and return predicted class and confidence score.
• Deployment:
  • Model is included as part of inference/app.
  • Built and deployed in local minikube.
  • Tested using Postman (POST API call).
• Files:
  • app\:
    • models
    • main.py
  • Dockerfile
  • inference-deployment.yaml

Asynchronous Prediction for Large Batches

• Folder: batch_runner/
• Functionality:
  • Periodic job in Kubernetes (minikube) runs daily to predict sentiment on a new batch of data.
  • Uses half of the remaining dataset to simulate incoming data.
  • Saves results to a mounted volume in minikube.
• Deployment:
  • Requires minikube-volume to be mounted.
  • Contains both a Kubernetes Job (for testing) and a CronJob (set to run daily).
• Files:
  • app/
    • models
    • batch_runner.py
  • Dockerfile
  • batch_job_manual.yaml (for testing)
  • batch_job.yaml (for daily execution)

Data Drift & Model Retraining

Data Drift Monitoring

• check_data_distribution() Function added to batch_runner.py to check class distribution in predicted data.
• Sends Slack alerts in case of a major class imbalance.
• Note: Slack webhook URL is hardcoded; ideally, use Kubernetes secrets.

Model Retraining

• Folder: retrain_runner/
• Retrains the model periodically using additional data.
• Saves new model artifacts.
• Requires minikube-volume for accessing fine-tuning data.
• Tested with limited data.
• Deployment:
  • Job (for testing)
  • CronJob (set to run once a month)
• Files:
  • Dockerfile
  • retrain_job.yaml
  • retrain_cronjob.yaml

Further steps

• Ensure minikube-volume is mounted properly to access batch and retraining data, or use other methods for data access.
• Implement a more robust data drift monitoring system.
• Consider using Kubernetes secrets for sensitive information (e.g., Slack webhook).

---