Lab 2 – MLOps (IE-7374)

Overview

This lab focused on exposing a trained machine learning model as a REST API using FastAPI and Uvicorn.
Instead of using the traditional Iris dataset, this implementation uses the Breast Cancer dataset from scikit-learn to build a binary classifier that predicts whether a tumor is benign or malignant.

The objective was to go beyond model training by learning how to serve models in real time — a key step in the modern MLOps workflow.

---

Project Structure

lab_2/
│── model/
│   └── cancer_model.pkl               # trained Decision Tree model
│── src/
│   ├── data.py                        # loads Breast Cancer dataset
│   ├── train.py                       # trains and saves model
│   ├── predict.py                     # loads model and performs predictions
│   └── main.py                        # FastAPI app exposing prediction endpoint
│── requirements.txt                   # dependencies
│── README.md                          # setup and usage guide
│── .gitignore                         # excluded files and folders

---

Workflow

Data Loading (data.py)

• Implemented a reusable function get_data() that loads the Breast Cancer dataset as Pandas DataFrames and returns (X, y) along with feature names.
• Makes it easy to switch datasets later without modifying training or inference code.

Model Training (train.py)

• Trained a DecisionTreeClassifier using 4 key features:
  ["mean radius", "mean texture", "mean smoothness", "mean compactness"]
• Saved the trained model as cancer_model.pkl using joblib.
• Designed so training can be rerun independently before deployment.

Prediction Logic (predict.py)

• Loads the model and performs predictions via a function predict_data().
• Keeps inference logic separate from web serving code for cleaner maintainability.

API Development (main.py)

• Built a FastAPI application with two endpoints:
  • GET / → health check
  • POST /predict → accepts JSON input and returns prediction
• Defined a Pydantic model for input validation (CancerData).
• Added human-readable outputs ("Benign" / "Malignant") for clarity.

Model Serving

Run the API with: bash uvicorn main:app --reload

What I Changed / Improved

• Replaced the default Iris model with a Breast Cancer classifier while keeping the FastAPI architecture identical.
• Simplified API inputs to 4 representative numeric features for cleaner and faster testing.
• Modularized dataset handling with data.py to improve reusability and separation of concerns.
• Added clear logging and success messages during model training for better traceability.
• Created a refined .gitignore and requirements.txt to ensure a clean and professional GitHub setup.

Results

Local Testing

• Model trained successfully and saved as:
• ../model/cancer_model.pkl

• Example input:

{ "mean_radius": 14.5, "mean_texture": 19.0, "mean_smoothness": 0.10, "mean_compactness": 0.14 }

• Example output:

{ "predicted_class": 1, "label": "Benign" }

Swagger UI

•	FastAPI auto-generated interactive API documentation available at:

http://127.0.0.1:8000/docs #### Successfully tested the /predict endpoint with multiple inputs

GitHub Integration

•	Repository created and pushed: fastapi
•	.gitignore excludes environment folders, model files, and cache.
•	README.md structured to reflect professional MLOps documentation standards.