This repository offers a collection of interactive machine learning experiments, primarily targeting developers and students interested in understanding and visualizing ML algorithms. It provides both Jupyter notebooks for training insights and browser-based demos for immediate interaction, facilitating hands-on learning of various ML techniques.

How It Works

The project showcases supervised and unsupervised learning through experiments built with TensorFlow 2 and Keras. For supervised learning, it includes Multilayer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs) for tasks like digit and sketch recognition, and image classification. Recurrent Neural Networks (RNNs) are demonstrated for sequence-based tasks like text generation. Unsupervised learning is explored with Generative Adversarial Networks (GANs) for tasks like clothes generation. Models are converted to TensorFlow.js formats for browser-based demos.

Quick Start & Requirements

• Setup: Create and activate Python virtual environments (.virtualenvs/experiments, .virtualenvs/converter). Install dependencies using pip install -r requirements.txt and pip install -r requirements.converter.txt.
• Jupyter: Launch with jupyter notebook.
• Demos: Navigate to demos directory, install Node.js dependencies with yarn install, and start the server with yarn start or yarn start-https.
• Requirements: Python > 3.7.3, Node >= 12.4.0, Yarn >= 1.13.0.

Highlighted Details

• Interactive Jupyter notebooks for model training.
• Browser-based demos using React and TensorFlow.js.
• Experiments cover MLP, CNN, RNN, and GAN architectures.
• Models are converted from Keras .h5 to TensorFlow.js formats for web deployment.

Maintenance & Community

The project is authored by @trekhleb. Links to related projects like "Homemade GPT" and "Homemade Machine Learning" are provided.

Licensing & Compatibility

The repository does not explicitly state a license in the provided README. Users should verify licensing for commercial or closed-source use.

Limitations & Caveats

The README explicitly states this is a sandbox/playground, not production-ready code. Models may not perform well, and issues like overfitting/underfitting are expected. Converting large models directly to the browser for demos is noted as inefficient for production.