Skip to content

Notebook Gallery

Every notebook below runs in your browser — click Open in Colab, no install required. They are ordinary, executed Jupyter notebooks in the repo, so you can also run them locally (jupyter lab notebooks/ or examples/).

New to Tangent? Start with the Tutorial, then see the readable gradients that are the whole point, then a case study in your domain.

Start here

  • Tangent Tutorial Open In Colab
    End-to-end tour: grad, the NumPy/JAX/PyTorch/TensorFlow backends, control flow, containers, and higher-order derivatives.

  • Tangent 2 — reproducible writeup Open In Colab
    The short paper on this fork, written so its code runs — a one-page tour of everything below, with every result reproducible in the notebook or via python paper/reproduce.py.

See the readable gradients — the niche

Tangent differentiates by source-to-source transformation: the gradient is Python you can read, debug, and edit. Tracing autodiff (JAX, PyTorch, TensorFlow) gives an opaque graph or tape; these notebooks show what having the actual adjoint source buys you.

  • Gallery of Gradients Open In Colab
    Eight functions, each shown next to its generated gradient code — from a polynomial to loops and control flow.

  • Gradient Surgery Open In Colab
    Read and edit the backward pass with insert_grad_of — scale, clip, log, and guard gradients mid-flow. The demo no tracing AD can copy.

Case studies — differentiate a messy NumPy simulator

Gnarly, loop-and-branch scientific and quantitative code, differentiated as written (no framework rewrite), with the gradient validated against finite differences.

  • LIBOR Market Model greeks (finance) Open In Colab
    A Monte-Carlo forward-rate simulator with state-dependent drift; exact deltas and vegas in one reverse pass instead of bump-and-revalue.

  • Projectile with drag (physics / ODE) Open In Colab
    A nonlinear ODE integrator (quadratic drag, no closed form); sensitivities to the launch parameters, then optimize the launch by gradient descent.

  • SIR epidemic calibration (epidemiology) Open In Colab
    Differentiate through a data-dependent lockdown branch to calibrate the transmission and recovery rates from a case curve.

  • Building energy optimization (applied) Open In Colab
    A thermal building simulation differentiated for gradient-based control / design optimization.

Run locally

pip install tangent-ad[all]
git clone https://github.com/pedronahum/tangent.git
jupyter lab tangent/notebooks tangent/examples

Prefer the terminal? python -m tangent doctor checks your install (and warns if the dead 2017 tangent package is shadowing tangent-ad).