About ten hours a week. Start with Karpathy, or with fast.ai lessons 1–3 if you need an early win (see below).
| Week | Focus | Do this | Done when |
|---|---|---|---|
| 14 | Autograd | Karpathy micrograd / Zero to Hero 1–2; rebuild Value from an empty file | Backward pass on paper + code |
| 15 | PyTorch basics | Official PyTorch tutorial; write a training loop and explain it | Loop runs; you can diagnose a planted mode or gradient bug |
| 16 | Debug training | Overfit 10 samples; learning-rate sweep; loss curves | Can name 3 failure modes |
| 17 | MLP / CNN | Dive into Deep Learning or CS231n notes; one small classifier | Baseline comparison and loss curves explained |
| 18 | Transfer learning | fast.ai lessons 1–4 or fastbook ch 1–4; train the head, then compare unfreezing the backbone | Demo on your images or domain |
| 19 | Attention intuition | Karpathy build-GPT video; type along | Can draw the attention block |
| 20 | Tokeniser | Karpathy tokeniser video; byte-level on small corpus | Encode/decode round-trip works |
| 21 | Your model + efficiency | Pick capstone; profile eager training; save and resume a full checkpoint; try one supported precision/compile setting | Config reproduces the run and the bottleneck is measured |
| 22 | Ablation + deploy | Report quality, memory and runtime; include failed runs; finish demo | Untouched test, checkpoint replay and honest ablation in README |