Backprop through loops, ifs, and plain NumPy

What it does Autograd wraps NumPy to compute exact derivatives of native Python functions. It supports both reverse-mode (backpropagation) and forward-mode differentiation, and the two can be composed arbitrarily. The main goal is gradient-based optimization, and it will happily differentiate through loops, conditionals, recursion, and closures.

The interesting bit Rather than restricting you to a static tensor API, Autograd differentiates native Python as written. Dynamic control flow is treated as a first-class citizen—fluid simulations and neural Turing machines are included as examples.

Key highlights

• Supports higher-order derivatives out of the box: you can differentiate the same function repeatedly.
• Reverse-mode and forward-mode differentiation can be mixed and matched arbitrarily.
• Handles a large subset of Python itself, including loops, if-statements, recursion, and closures.
• End-to-end examples cover convnets, LSTMs, Gaussian processes, and backpropagating through a fluid simulation.

Verdict Autograd suits researchers and tinkerers who want gradients for dynamic, NumPy-centric Python without adopting a heavyweight framework. If you are looking for a modern, production-grade deep-learning stack, you may find it too lightweight.