Tensors, autograd and the training loop — the foundations every AI engineer stands on. We rebuilt nn.Linear from scratch.
What we covered
Most PyTorch tutorials start with model.fit and never look underneath. We went the other way: raw tensors first, then the autograd engine, then the training loop — so every abstraction above feels earned, not magical.
- Tensors as the lingua franca — shapes, strides, broadcasting, and why
(batch, features)runs the world. - Autograd by hand — a tiny scalar engine that builds the computation graph and backpropagates, before touching
torch.autograd. - The training loop, deconstructed — forward, loss, backward, step, zero-grad. Every line explained, nothing hidden behind a trainer class.
What we built
By the end of the morning, every laptop in the room had a from-scratch linear layer training on real data — then we swapped in nn.Linear and watched the same loop run on the real thing. Same math, same gradients, zero mystery.
Takeaways
- If you can implement
backward()for a scalar graph on paper, PyTorch stops being a black box forever. - Shape errors are the compiler telling you your mental model is wrong — read them, don’t fear them.
- The training loop is five lines. Everything else is convenience.
Join the next one
We run sessions like this every Saturday in Hyderabad — free, hands-on, and open to everyone. RSVP on Meetup and bring a laptop with Python 3.10+.
