juniorPyTorch
What is autograd in PyTorch?
Updated May 17, 2026
Short answer
Autograd is PyTorch’s automatic differentiation engine.
Deep explanation
It tracks operations on tensors and builds a computation graph dynamically, enabling gradient computation via backpropagation.
Real-world example
Used in training neural networks where gradients are required for optimization.
Common mistakes
- Forgetting requires_grad or calling backward multiple times without zeroing gradients.
Follow-up questions
- What is a computation graph?
- What does detach() do?