C.W.K.
Stream
Quiz · 8 questions

🧮 Autograd — Automatic Differentiation

The thing that makes neural networks trainable. Dynamic graphs, backward passes, and the gotchas you'll hit.

Level 0Tensor Curious
0 XP0/62 lessons0/13 achievements
0/120 XP to next level120 XP to go0% complete

Quiz

01Why is reverse-mode AD the right choice for training neural networks?
02Why does every training loop have optimizer.zero_grad()?
03Which is the modern, recommended context manager for pure inference?
04After backward(), why is b.grad None for an intermediate tensor b?
05What is the modern API for mixed-precision training?
06When using bfloat16 autocast, do you typically need a GradScaler?
Hint
What problem does GradScaler solve, and does bf16 have it?
07Where in the training loop should gradient clipping go?
08What does torch.func.vmap let you do most cleanly?
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.