https://github.com/persteenolsen/fastapi-jwt-auth-dl-two

Python FastAPI ML Inference Service with ONNX Runtime and PyTorch-Trained Model for House Price Prediction (v5)
https://github.com/persteenolsen/fastapi-jwt-auth-dl-two

deep-learning fastapi jwt onnx python pytorch

Last synced: 9 days ago
JSON representation

Python FastAPI ML Inference Service with ONNX Runtime and PyTorch-Trained Model for House Price Prediction (v5)

• Host: GitHub
• URL: https://github.com/persteenolsen/fastapi-jwt-auth-dl-two
• Owner: persteenolsen
• Created: 2026-04-27T13:08:30.000Z (about 2 months ago)
• Default Branch: main
• Last Pushed: 2026-04-27T13:27:04.000Z (about 2 months ago)
• Last Synced: 2026-04-27T15:25:51.121Z (about 2 months ago)
• Topics: deep-learning, fastapi, jwt, onnx, python, pytorch
• Language: Python
• Homepage: https://fastapi-jwt-auth-dl-two.vercel.app/docs
• Size: 33.2 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# 🏠 v5 - House Price Prediction API (FastAPI + PyTorch + ONNX)

Last updated:

- 03-05-2026

A production-style machine learning backend system that predicts house prices using a neural network trained in PyTorch and deployed using ONNX Runtime with FastAPI.

This project demonstrates a complete end-to-end ML engineering workflow from data generation to secure API deployment, including model tuning for stable and realistic regression behavior.

---

## 🚀 Features

- End-to-end ML pipeline (data → training → inference)

- Regression model for house price prediction

- Synthetic dataset generation with engineered features

- Feature scaling and target normalization

- PyTorch neural network training

- ONNX model export for fast inference

- FastAPI REST API for serving predictions

- JWT authentication (OAuth2 password flow)

- Environment-based configuration (.env)

- Production-ready dependency separation

- Vercel deployment compatible structure

- Tuned model for smooth and realistic price behavior

---

## 🧱 Tech Stack

- Python 3.12

- PyTorch

- NumPy

- ONNX Runtime

- FastAPI

- Uvicorn

- python-jose (JWT authentication)

- python-dotenv

---

## 📁 Project Structure

```

.

├── main.py                 # FastAPI inference API

├── train.py               # Model training + ONNX export

├── data.py                # Synthetic dataset generator

├── model.onnx            # Exported ONNX model

├── preprocessing.json     # Normalization parameters

├── requirements/

│   ├── train.txt

│   ├── dev.txt

├── requirements.txt      # Production (Vercel)

└── .env                  # Environment variables (not committed)

```

---

## ⚙️ Installation

git clone https://github.com/your-repo/house-price-api.git

cd house-price-api

python -m venv .venv

source .venv/bin/activate  # Windows: .venv\Scripts\activate

pip install -r requirements/dev.txt

---

## 🔐 Environment Variables (.env)

JWT_SECRET=your_secret_key  

JWT_ALGORITHM=HS256  

ACCESS_TOKEN_EXPIRE_MINUTES=60  

ADMIN_USERNAME=admin  

ADMIN_PASSWORD=password  

---

## 🏋️ Training the Model

Generate dataset and train model:

python data.py  

python train.py  

Outputs:

- model.onnx

- preprocessing.json

---

## 🔧 Model Tuning

During development, the model was tuned to improve stability and realism of predictions.

Key tuning changes:

- Reduced hidden layer size (8 → 6 neurons)

- Lowered learning rate (0.01 → 0.002–0.003)

- Increased training epochs (200 → 350)

- Added weight decay (L2 regularization)

- Introduced early stopping for training stability

- Improved numerical stability in normalization

Result:

- Smooth, monotonic price curves

- Stable age depreciation behavior

- More consistent size scaling

- Reduced high-range prediction jumps

- Better generalization without overfitting

---

## 🚀 Run API Locally

uvicorn main:app --reload  

Swagger UI:  

http://127.0.0.1:8000/docs  

---

## 🔐 Authentication

Get Token:

POST /token  

username=admin  

password=password  

Use Token:

Authorization: Bearer   

---

## 📡 Prediction Endpoint

POST /predict  

Example request with base values like in test.py:

{

  "size": 100,

  "rooms": 3,

  "age": 10,

  "distance": 5,

  "income_area": 50

}

Response:

{

  "predicted_price": 371053.23,

  "user": {

    "sub": "testuser",

    "exp": 1777793908

  }

}

Predictions are generated by a trained neural network and reflect learned relationships from the synthetic dataset.

---

## 🧪 Training Pipeline

- Generate synthetic housing dataset

- Apply feature engineering:

  - size

  - rooms

  - age

  - distance penalty

  - income influence

- Normalize inputs and target values

- Train neural network in PyTorch using tuned hyperparameters

- Apply regularization for smoother behavior

- Export trained model to ONNX format

- Save preprocessing metadata for inference

---

## 🚀 Deployment (Vercel)

Production environment uses:

requirements.txt  

- ONNX Runtime for inference

- No PyTorch included in production

- FastAPI serverless-compatible design

---

## 🔁 System Comparison (XOR vs House Price Model v5)

XOR Project (Previous Version)

- Binary classification problem

- 2 input features

- Simple neural network (logic learning)

- Fixed dataset (truth table)

- Minimal preprocessing

- Educational focus only

House Price Project (Current Version)

- Regression problem (continuous output)

- 5 engineered input features

- Realistic synthetic dataset with noise

- Full preprocessing pipeline (X + Y normalization)

- Tuned neural network for stable behavior

- ONNX-based production inference

- JWT-secured API

- Production-style architecture

Key Upgrade:

ML concept demonstration → ML system engineering with tuning and deployment awareness

---

## 🧠 Final Summary

This project demonstrates a complete machine learning backend system covering the full lifecycle of an ML application:

- Data generation and feature engineering

- Model training using PyTorch

- Careful hyperparameter tuning for stable regression behavior

- Input and output normalization for stable learning

- Regularization and early stopping for generalization control

- Model export to ONNX for efficient inference

- FastAPI-based REST API for serving predictions

- JWT authentication with environment-based configuration

- Production-style separation between training and inference environments

It represents a transition from basic ML experimentation to real-world ML system design and tuning practices, similar to production machine learning services used in industry.

---

## 👨‍💻 Author

Built as a learning project to understand production-style machine learning system design, deployment, inference architecture, and model tuning in practice.