// the find
labmlai/annotated_deep_learning_paper_implementations
🧑🏫 60+ Implementations/tutorials of deep learning papers with side-by-side notes 📝; including transformers (original, xl, switch, feedback, vit, ...), optimizers (adam, adabelief, sophia, ...), gans(cyclegan, stylegan2, ...), 🎮 reinforcement learning (ppo, dqn), capsnet, distillation, ... 🧠
60+ annotated PyTorch implementations of ML papers, rendered as side-by-side code-and-notes pages on nn.labml.ai. It's a learning resource first, a code library second — the whole point is reading the implementations alongside the paper explanations, not importing them into your project. Aimed at developers who want to understand how attention, diffusion, GANs, and optimizers actually work rather than just calling into HuggingFace.
The breadth is genuinely impressive: transformers from vanilla through Flash Attention and RoPE, diffusion from DDPM through Stable Diffusion, GANs through StyleGAN2, plus optimizers, normalization layers, and RL — most of the modern ML canon is here. The literate-programming format works well; code and explanation live next to each other rather than in a separate doc that drifts. The implementations are clean enough to actually read, not the kind of research code that's been hacked into submission. LoRA and GPT-NeoX coverage means it's not stuck in 2021.
Last push was January 2026 and the README still says 'adding new implementations almost weekly' — that pace has clearly slowed, and some newer architectures (Mamba, Moe with expert routing, anything post-2024) are missing. The labml dependency for experiment tracking creates friction if you just want to run a single implementation without pulling in their ecosystem. No GPU memory estimates or hardware requirements anywhere, so the Stable Diffusion and NeoX sections will silently fail on modest hardware. The Japanese translation is a nice touch but it's clearly auto-generated and occasionally confusing.