A readable zoo of ML algorithms, from bandits to ResNet

What it does mlimpl is a curated collection of educational implementations covering statistical learning, deep learning, and reinforcement learning. Most classes follow a scikit-learn-style API with fit, predict, and score methods, but the internals are left open for reading and hacking. The repo spans multi-armed bandits and Markov decision processes on the RL side, through LSTMs, VAEs, and a full ResNet50 training loop on the deep learning side.

The interesting bit The author leaves in the battle scars. The ResNet50 write-up openly admits hitting a CPU bottleneck while sharding ImageNet1K into HDF5 files, and only reaching 68.1% validation accuracy after 64 epochs—well below the paper’s 78%. There is even a bare-bones autograd engine built without PyTorch, tracing gradients through a manual computational graph so you can see how automatic differentiation actually works.

Key highlights

• Scikit-learn-like class structure (fit, predict, score) across most algorithms.
• Broad coverage: RL basics (UCB, policy gradient, MDP), classic deep learning (GAN, CNN, VAE, LSTM), and modern self-supervised learning (Barlow Twins, VicReg, SimCLR).
• A from-scratch autograd system using computational graphs, no Torch required.
• Honest notes on practical hurdles, like grayscale edge cases in custom ImageNet dataloaders and TensorFlow 1-era CAPTCHA CNNs.

Caveats

• The codebase is a mix of PyTorch, TensorFlow 1, MATLAB, and NumPy, so consistency varies by folder.
• Some entries read more like course homework or competition write-ups than polished libraries.

Verdict Ideal for students, interview preppers, or anyone who learns by reading source code rather than import statements. Skip it if you need production-grade, maintained packages.