This phase you stop reading about neural nets and run them: a playground net you can break, backprop you rebuild by hand, a CNN whose every pixel you can poke, a GAN duel you can sabotage — then PyTorch, a free GPU, and your first deployed vision model. Covers the whole deep-learning core and the CV backbone of the car-damage capstone.
Train a CNN on a meaningful image dataset (plant disease, X-rays, or your own photos) twice — once from scratch, once by fine-tuning a pretrained backbone — and compare honestly: accuracy, data needed, GPU minutes. Then ship it: a Gradio app on Hugging Face Spaces that anyone can open on a phone. Alongside, finish the from-scratch set — BYO-5 (mini-PyTorch) and BYO-6 (CNN) — so the framework never gets to be magic. You're ready for Phase 5 when you can build, train, debug and deploy a PyTorch model on a free GPU, and explain backprop, conv, pooling and LSTMs from memory.
Full guide: phase-4-deep-learning.md ↗ · capstone tie-in: car-damage cost estimator (India).