DDeepanshiindpanshi.hashnode.dev·6d ago · 6 min readLet's Talk SSLThis week I did a deep dive on image recognition based self supervised learning algorithms. The idea was to do a rigorous trial of three different extremely popular architectures: MAE, DINO and IJEPA 00
DDeepanshiindpanshi.hashnode.dev·Aug 1 · 5 min readTransformers from scratch: things that silently brokeOver the last three weeks I built a decoder-only GPT, a Vision Transformer, RoPE, and GQA, all from raw PyTorch, no copy-pasting a reference implementation. The models train and the loss curves look r00
