This directory contains API implementations for image classification using a fine-tuned model on the Quickdraw dataset. The models are hosted on Hugging Face and configured via environment variables specified in a .env file located at the root of the project. A performance review indicates that you can run these file on practically all devices. It only takes few MB on RAM or VRAM.

The endpoints/ directory includes three Python scripts, each designed to run the image classification API on different hardware setups: CPU, GPU, and MPS (Apple's Metal Performance Shaders). The API uses FastAPI for serving predictions via HTTP requests.

• FastAPI
• Uvicorn
• PyTorch
• Transformers library from Hugging Face
• A .env file containing:
  • MODEL_CKPT: Model checkpoint on Hugging Face
  • HOST: Host address (usually localhost or 127.0.0.1)
  • PORT_APP: Local web view port
  • PORT_ENDPOINT: Port number for the API server
  • SUPABASE_URL: Supabase API private URL
  • SUPABASE_KEY: Supabase private API key

1. Environment Setup:

   • Install required packages:
     pip install -r requirements.txt

2. Configuration:

   • Create a .env file in the root directory with the following content:

     MODEL_CKPT=<model-checkpoint>
     HOST=localhost
     PORT_APP=5500
     PORT_ENDPOINT=8000
     SUPABASE_URL=...
     SUPABASE_KEY=...
     

     However, to get the SUPABASE_URL and SUPABASE_KEY parameters and access to the Supabase API. Your have to ask permissions to the owner of the repository.

3. Running the API:

   • Navigate to the directory containing the desired script based on your hardware. Be careful, you must have a CUDA compatible GPU to run predict_gpu.py.
   • Execute the script using Uvicorn:

     python endpoints/predict_cpu.py # For CPU
     python endpoints/predict_gpu.py # For GPU
     python endpoints/predict_mps.py # For MPS
     

   • Access the API at http://localhost:8000/docs to interact with the Swagger UI and test the endpoints.

• GET /: Returns a welcome message and a link to the API documentation.
• GET /add_score: Add a score to the score database. Your must provide user, score, mean_time, modeand difficulty.
• GET /device: Returns the type of device being used for inference: cpuor gpu.
• GET /labels: Return the label set used by the model for predictions.
• GET /model: Returns the model checkpoint.
• GET /scores: Returns the top 3 scores from the database with parameters modeand difficulty.
• POST /predict_with_path: Accepts an image path and returns the classification results.
• POST /predict_with_array: Accepts an image as a nested list of integers and returns the classification results.
• POST /predict_with_array: Accepts an image file and returns the classification results.

• Ensure that the model checkpoint specified in .env matches the device compatibility (CPU/GPU/MPS).
• Modify the host and port settings as required by your deployment environment.