Skip to content

File Structure (LSTM Traffic Predictor)

Overall Project Structure

lstm-predictor-service/
├── app/
│   ├── main.py
│   ├── images/
│   │   ├── logo.png
│   │   └── favicon.ico
│   ├── models/
│   │   ├── lstm_model.py
│   │   └── lstm_train.py
│   ├── trained_models/
│   │   ├── lstm_model.keras
│   │   ├── lstm_model_tf/
│   │   └── scaler.pkl
│   └── utils/
│       ├── stationarity.py
│       ├── feature_engineering.py
│       └── metrics.py
├── edge-analysis.py
├── trip-analysis.py
├── Dockerfile
├── requirements.txt
├── .env.example
├── .gitignore
├── README.md
└── docker-compose.override.yml

File Descriptions

app/main.py

FastAPI application for inference: - Loads trained LSTM model from app/trained_models/lstm_model.keras - Initializes MinMaxScaler from app/trained_models/scaler.pkl - Implements 4 endpoints: /health, /model-info, /predict, /favicon.ico - Pydantic models for request/response validation - Comprehensive error handling with HTTP exceptions - Custom Swagger UI with favicon support

app/models/lstm_train.py

Complete training pipeline: - Parses SUMO edgeData.xml - Identifies top 5 most congested edges by average density - Normalizes data using MinMaxScaler - Creates sequences: 3 timesteps input → 1 timestep forecast target - Builds LSTM model: 64 units → Dense(32) → Dense(5 outputs) - Trains for 50 epochs, batch size 2 - Evaluates on 20% test set (temporal split, no shuffle) - Saves in both Keras .keras and TensorFlow SavedModel formats - Output: Test Loss (MSE) 0.0698, Test MAE 0.2084

app/models/lstm_model.py

Data preparation and exploration script: - Parses edgeData.xml from SUMO results - Identifies top 5 congested edges - Normalizes data with MinMaxScaler - Outputs dataset shape and sample records for inspection

app/trained_models/lstm_model.keras

Trained LSTM model (production format): - Keras native format (TensorFlow 2.11+) - Input: (batch, 3, 5) — 3 timesteps × 5 edges - Output: (batch, 5) — 5 edge density predictions - MAE: 0.2084 (normalized scale) - MSE: 0.0698 (normalized scale)

app/trained_models/lstm_model_tf/

Alternative TensorFlow SavedModel format: - For production deployment with TensorFlow Serving - Compatible with cloud platforms - Includes model signature and metadata

app/trained_models/scaler.pkl

MinMaxScaler pickle file: - Fitted on training data - Scales density values to [0, 1] range - Used for both training and inference normalization - Loaded at startup in main.py

app/utils/

Utility modules: - stationarity.py: Stationarity checking with scipy - feature_engineering.py: Feature engineering functions (placeholder) - metrics.py: Model evaluation metrics (placeholder)

app/images/

Static assets: - logo.png: Project logo (used in docs) - favicon.ico: Browser tab icon

edge-analysis.py

SUMO data analysis script: - Parses edgeData.xml from SUMO simulation - Computes density, occupancy, waiting time statistics - Groups by edge and identifies top 10 most congested - Useful for data exploration and validation

trip-analysis.py

Trip statistics script: - Parses tripinfo.xml from SUMO simulation - Analyzes trip duration, departure delays, waiting times - Computes aggregate statistics across all vehicle trips - Useful for understanding overall traffic performance

Dockerfile

Production container build: - Base: python:3.9-slim - Installs dependencies from requirements.txt - Runs: uvicorn app.main:app --host 0.0.0.0 --port 8000

requirements.txt

Python dependencies: - FastAPI, Uvicorn (web framework) - TensorFlow, Keras (model training/inference) - Pandas, Numpy, scikit-learn (data processing) - Colorama (colored terminal output)

README.md

Service documentation: - Quick start guide - Data flow diagram - LSTM concepts explanation - Available endpoints

.env.example

Environment variable template: - API_PORT, MODEL_PATH, etc.

Status

✅ Fully Operational

Working: - [x] Data loading and preprocessing from SUMO - [x] Model training on historical patterns - [x] Inference service (3 endpoints) - [x] Swagger UI documentation - [x] Health monitoring and logging

Future: - [ ] 15-minute ahead forecasting - [ ] MAE < 10% on raw traffic values - [ ] Handle missing sensor data (KNN imputation) - [ ] Docker containerization - [ ] Integration with RL Inference Service