MLOps API: Sentiment Analysis with DistilBERT by AMAH DanielMLOps API: Sentiment Analysis with DistilBERT by AMAH Daniel

MLOps API: Sentiment Analysis with DistilBERT

AMAH Daniel

Completed work

Automation Engineer

Data Engineer

ML Engineer

FastAPI

GitHub

Hugging Face

Artificial Intelligence

Like this project

Completed work

Posted Apr 6, 2025

Exploring a simple text classification on emotions with DISTILBert and enabling MLOPs with colab, MLFlow, wandb, fastapi, Docker and Github Actions - mlops-sen…

Likes

Views

Timeline

Apr 4, 2025 - Apr 6, 2025

Emotion Classification API

A production-ready machine learning service that classifies text into emotions using a fine-tuned DistilBERT model with full CI/CD pipeline integration.

Architecture Overview

This MLOps solution consists of:

Fine-tuned DistilBERT model hosted on Hugging Face Hub

FastAPI service for low-latency model serving

Automated CI/CD pipeline with GitHub Actions

Containerized deployment on Render

Getting Started

Prerequisites

Python 3.11+

Docker (for local container testing)

Hugging Face account (for model hosting)

Local Development

Clone the repository:

git clone https://github.com/AkanimohOD19A/mlops-sentiment-distilbert.git
cd emotion-classifier

Install dependencies:

pip install -r requirements.txt

Run the API locally:

# Using the Hugging Face model
MODEL_ID="AfroLogicInsect/emotionClassifier" uvicorn app.main:app --reload

# OR using a local model
MODEL_PATH="./model_export" uvicorn app.main:app --reload

Access the interactive API documentation at http://localhost:8000/docs

Running Tests

pytest --cov=app tests/

CI/CD Pipeline

The automated GitHub Actions workflow:

Test: Runs unit tests and code quality checks

Build: Packages the application into a Docker container

Deploy: Pushes to Docker Hub and triggers deployment on Render

Model Information

The emotion classification model is fine-tuned on [dataset description] and detects six emotions: anger, fear, joy, love, sadness, and surprise.

Training metrics and experiment tracking are available in the training notebook.

The CI/CD pipeline automatically deploys the application to Render:

Demo URL: https://mlops-sentiment-distilbert.onrender.com/

Making Predictions

cURL Example

emotion-classifier/
├── .github/workflows/      # GitHub Actions workflows
├── app/                    # FastAPI application
│   ├── main.py             # Main API endpoints
│   └── ml/                 # ML code
│       └── predictor.py    # Model prediction class
├── tests/                  # Unit tests
├── Dockerfile              # Docker configuration
├── README.md               # This file
└── requirements.txt        # Python dependencies

Making Prediction Calls

Now having deployed this project, you can just make API Calls to the deployed render service E.g:

curl -X POST "https://mlops-sentiment-distilbert.onrender.com/predict" \
  -H "Content-Type: application/json" \
  -d '{"text": "I am feeling very happy today!"}'

Python Example

import requests
import json

url = "https://mlops-sentiment-distilbert.onrender.com/predict"
data = {"text": "I am feeling very happy today!"}
headers = {"Content-Type": "application/json"}

response = requests.post(url, data=json.dumps(data), headers=headers)
print(response.json())

Sample Response

{
  "emotion": "joy",
  "confidence": 0.9988011121749878,
  "all_emotions": {
    "anger": 0.00021597424347419292,
    "fear": 0.00011065993021475151,
    "joy": 0.9988011121749878,
    "love": 0.00037375965621322393,
    "sadness": 0.00031321862479671836,
    "surprise": 0.00018527933571022004
  }
}

Other NOTES

-> SOFTMAX Calc. [0.1, 0.2, 0.7] The softmax calculation takes exponentials of each value and then normalizes them to sum to 1. When the differences between values are small, the probabilities are more evenly distributed.

emotion-classifier/
├── .github/workflows/      # CI/CD automation
├── app/                    # FastAPI application
│   ├── main.py             # API endpoints
│   └── ml/                 # ML components
│       └── predictor.py    # Model prediction logic
├── tests/                  # Test suite
├── Dockerfile              # Container configuration
├── README.md               # Project documentation
└── requirements.txt        # Dependencies

Technical Details

Model Architecture: DistilBERT (faster and lighter version of BERT)

Inference Optimization: Batched inference, model quantization

Monitoring: Basic metrics via Render dashboard

Security: API secret key authentication (optional)

Flowchart

Training to Deployment

flowchart LR
    subgraph Training
        A[Raw Text Data] -->|Preprocessing| B[Processed Dataset]
        B -->|Fine-tuning| C[Base DistilBERT]
        C -->|Training| D[Fine-tuned Model]
        D -->|Evaluation| E[Validated Model]
    end
    
    subgraph Deployment
        E -->|Model Upload| F[Hugging Face Hub]
        F -->|Model Registry| G[Model Serving API]
        G -->|Container| H[Docker Image]
        H -->|Hosting| I[Render Service]
    end
    
    style Training fill:#e6f7ff,stroke:#1890ff
    style Deployment fill:#f6ffed,stroke:#52c41a

CI/CD Workflow

flowchart TD
    A1[GitHub Push] -->|Trigger| B1[GitHub Actions]
    
    subgraph CI
        B1 -->|Install Requirements| C1[Setup Environment]
        C1 -->|Run Tests| D1[Test Suite]
        D1 -->|Code Quality| E1[Flake8 Checks]
    end
    
    subgraph CD
        E1 -->|If Tests Pass| F1[Build Docker Image]
        F1 -->|Docker Login| G1[Push to Docker Hub]
        G1 -->|Deployment Trigger| H1[Deploy to Render]
        H1 -->|Health Check| I1[Verify Deployment]
    end
    
    style CI fill:#f9f0ff,stroke:#722ed1
    style CD fill:#fff2e8,stroke:#fa541c

Model Serving

flowchart LR
    A2[Client] -->|HTTP Request| B2[Load Balancer]
    B2 -->|Route Request| C2[FastAPI Service]
    
    subgraph API
        C2 -->|Validation| D2[Input Processing]
        D2 -->|Model Inference| E2[DistilBERT Model]
        E2 -->|Post-Processing| F2[Format Response]
    end
    
    F2 -->|HTTP Response| A2
    
    style API fill:#e6fffb,stroke:#13c2c2

Monitoring (wandb/mlflow)*

flowchart TD
    A3[Inference Requests] -->|Logging| B3[Request Logs]
    A3 -->|Metrics Collection| C3[Performance Metrics]
    A3 -->|Error Tracking| D3[Error Logs]
    
    subgraph Monitoring
        B3 -->|Log Analysis| E3[Request Patterns]
        C3 -->|Dashboards| F3[Performance Visualization]
        D3 -->|Alerts| G3[Error Notifications]
    end
    
    subgraph Observability
        E3 --> H3[Model Drift Detection]
        F3 --> I3[SLA Monitoring]
        G3 --> J3[Incident Response]
    end
    
    style Monitoring fill:#fff1f0,stroke:#f5222d
    style Observability fill:#fcffe6,stroke:#a0d911