Compresso

Compresso is an interactive application for dataset compression based on the concept of Minimal Finite Covering (MFC). It provides an intuitive, visual tool for exploring how datasets can be compressed while preserving their essential structure and information.

Features

• Minimal Finite Covering–based compression for efficient data representation
• Interactive visualization showing how compressed samples can represent the original dataset
• Multiple dataset support, including built-in options:
  • MNIST, CIFAR-10, CIFAR-100, and SVHN
  • plus the ability for users to upload custom datasets, strictly following the required formats
• Integrated frontend + backend architecture — runs locally via Docker, similar to a self-contained Jupyter-style environment
• Browser-based interactive interface, making experimentation simple and visual

Tech Stack

• Backend: Python 3.11· FastAPI · PyTorch
• Frontend: C# · ASP.NET Core · Blazor Server (.NET 9)
  • Includes custom JavaScript interop for client-side rendering and user interaction
  • Interactive visualizations built with Plotly.js
• Communication: HTTP (JSON) · Server-Sent Events (SSE) for real-time progress updates
• Containerization: Docker · Docker Compose
• Visualization: Plotly.js · NetworkX (graph layout generation and interactive web UI)
• Data Formats: JSON · .pt (trained models and compressed datasets in storage)

Project Structure:

Compresso/
├── Backend/ (FastAPI backend service)
│   ├── data/
│   ├── requirements.txt
│   ├── xxx.py
│   └── Dockerfile
├── CompressoApp/ (ASP.NET Core frontend)
│   ├── Components/
│   ├── Models/
│   ├── Services/
│   ├── wwwroot/
│   ├── xxx.cs
│   └── Dockerfile
├── docker-compose.yml (Orchestrates backend and frontend)
└── README.md

Prerequisites:

• Docker installed (https://docs.docker.com/get-docker/)
• Docker Compose installed (usually included with Docker)

Quick Start:

1. Clone the repository

git clone https://github.com/BinaryHexedecimal/Compresso.git

cd Compresso

2. Prepare the Gurobi license (Optional)

2.1 Linear Optimization Setup

This project includes three linear optimization solvers: CBC, SCIP, and Gurobi.
Among them, Gurobi generally offers the best performance and is the recommended option. However, it requires a valid license (free for personal or academic use).

2.2 Choosing an Optimizer

If you prefer not to use Gurobi for any reason, you can skip its setup — the compression page will still provide access to the other two optimizers (CBC and SCIP).

If you wish to use Gurobi, you can obtain a free Web License Service (WLS) license at:
https://www.gurobi.com

Note:

• A local Gurobi license (the standard .lic file) typically will not work inside Docker. You must use a WLS (Web License Service) license for containerized environments.
• If you already have a WLS license for another container, you can reuse it here.

2.3 One-Time Setup

Once you have your WLS credentials, create a file named .env in the project’s root directory (Compresso/) and add the following environment variables:

GRB_WLSACCESSID=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx
GRB_WLSSECRET=xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
GRB_LICENSEID=1234567

You only need to do this once — the same credentials can be reused for future runs, unless the license expires or is revoked.

3. Build and run the application

docker compose up --build

Note:

• After the initialization, you can start existing containers simply by running:docker compose up
• However, if you have made changes to the code, dependencies, or Docker configuration, you should rebuild the images using docker compose up --build

4. Open Compresso in your browser

http://localhost:3000

5. Stop the application

docker compose down

or use Ctrl + C in the terminal.

How It Works:

The backend automatically downloads and preprocesses the four built-in datasets on startup.

All permanent and temporary data is stored under data/, which is accessible to the user.
The data/ directory inside the container mirrors your local folder Compresso/Backend/data/, allowing users to view and manage their data directly.

The frontend communicates with the backend API to visualize compressed datasets, perform training, and interactively explore MFC-based compression results.

Common Issues and Solutions:

Problem: Slow first startup

Cause: Installing packages and Preloading built-in datasets

Solution: Wait until “webmcs-frontend | info: Microsoft.Hosting.Lifetime[0]” appears in the logs.

By default, the backend is allowed up to 40 minutes to initialize before it stops automatically. You can adjust this duration in the docker-compose.yml under the healthcheck section.

More issues may be added as they arise.

Note

• Modify docker-compose.yml to change ports if needed
• For visualization purposes, we use 15% of each of the four built-in training datasets in compression. This configuration is designed to ensure smooth operation on a typical personal computer. The parameter can be adjusted by modifying BUILT_IN_DATASET_PERCENT in Compresso/Backend/globals.py as needed.

License:

This project is released under the MIT License.
You are free to use, modify, and distribute it with attribution.

Author:

Ying Pei
pwb749@alumni.ku.dk