Style Synth - Fashion Recommendation System

Team: Courtney Green, Li-Wen Hu, Satomi Ito, Nandini Kodali, Sophia Rutman

A fashion recommendation system that helps users style their own clothing or build new outfits from an uploaded image. The system combines computer vision (ResNet-50 embeddings) with efficient similarity search and recommendation algorithms.

Architecture

Key Features

Outfit Generation

1. Upload clothing items to your wardrobe
2. Select occasion (Casual, Formal, Business, Party) and season
3. Click "Generate" to receive AI-recommended outfit combinations
4. Customize outfits by adding or removing items
5. Save your favorite outfits for later

Wardrobe Management

• Upload images with automatic category classification via ResNet-50
• Edit item details including brand, color, occasion, and season tags
• Delete items you no longer want in your wardrobe

Saved Outfits

• View all saved outfits with filtering by occasion and season
• Edit saved outfits by adding/removing items
• Rate and review outfits for future reference

Running with Docker

To run the entire application stack locally:

docker-compose up --build

This starts the FastAPI backend, Streamlit frontend, PostgreSQL database, and Prometheus/Grafana monitoring.

Project Structure

Computer Vision Files (ComputerVisionFiles/)

• ResNet.ipynb - Colab notebook used to preprocess Fashion MNIST, extract embeddings using ResNet-50, and generate visualization outputs
• fashion_mnist_classes.txt - Text file listing the 10 Fashion MNIST class names
• fashion_mnist_labels.npy - Array of label indices for each embedded image (stored with Git LFS)
• fashion_mnist_resnet50_embeddings.npy - Array of 2048-dimensional ResNet-50 embeddings for the entire Fashion MNIST training set (stored with Git LFS)

Recommendation Engine (RecommendationFiles/)

• recommendation_engine.py - Main recommendation engine module with PCA dimensionality reduction and FAISS similarity search
• example_usage.py - Example usage script demonstrating all features
• benchmark.py - Comprehensive benchmarking script
• test_integration.py - Integration tests
• quick_test.py - Quick verification script
• run_tests.sh - Automated test runner
• requirements.txt - Python dependencies
• README.md - Detailed documentation

Quick Start

1. Install Dependencies

cd RecommendationFiles
pip install -r requirements.txt

cd backend
pip install -r requirements.txt

cd frontend
pip install -r requirements.txt

2. Ensure Data Files Are Available

# From project root
git lfs install
git lfs pull

3. Run Quick Test

cd RecommendationFiles
python quick_test.py

Getting Started

For detailed instructions, see the Recommendation Engine README.

Git LFS

The .npy files are large and are therefore tracked using Git LFS instead of regular Git.

After cloning the repository, run:

git lfs install
git lfs pull

Project Components

Computer Vision Component

The CV team provides ResNet-50 embeddings (2048-dimensional) extracted from Fashion MNIST images. These embeddings are stored in the ComputerVisionFiles/ directory.

Recommendation Engine Component

The recommendation engine provides:

• Dimensionality Reduction: PCA-based reduction from 2048D to configurable dimensions (default: 128D)
• FAISS Integration: High-performance similarity search
• Flexible Filtering: Filter by class, exclude items, combine filters
• Performance Optimized: ~10,000+ queries/second, ~30MB memory footprint

See RecommendationFiles/README.md for complete documentation.

Backend

The backend uses FastAPI to connect S3 and RDS services from AWS to our site. In order to run and test the API calls locally, run

uvicorn backend.app.main:app --reload

Then, navigate to http://127.0.0.1:8000 on your machine to access specific calls.

The backend is deployed on the cloud using Render, located at https://dsan6700.onrender.com.

Frontend

The frontend is written using Streamlit, where the API calls are connected to the various functionalities. To run the frontend locally, run

streamlit run frontend/app.py

The frontend is deployed on the cloud using Streamlit, located at https://stylesynth.streamlit.app.