Skip to content

Guides

Task-oriented pages, grouped by what you are trying to do. Each one is self-contained and links to the API it uses; the tutorials are the guided path through the same material.

Build a problem

  • Bring your own problem

    nefi.from_forward(forward, y, shape, prior=...): any differentiable PyTorch function in about twenty lines; the prior DSL; what "auto" decides and how to override it.

  • Operator, field & loss zoo

    Every forward operator, representation, head and loss that composes with everything else — with a short guide to choosing among them.

  • A new forward operator

    Write the physics of your instrument, make it work at every curriculum resolution, check its gradient and pair it with an inverse-crime-safe data generator.

  • A new instance

    Package a complete, reproducible problem behind the Instance protocol and get the CLI, benchmarks, diagnostics and gallery for free.

Solve it well

  • Geometric representations

    Layered media, anomalies, star shapes, warps and spectral preconditioning; residual-aware annealing, capacity growth, held-out representation selection and spectral match reports.

  • Auto-tuning

    Detect and fix gauges, budgets that stop short of the noise floor and mis-weighted priors; an acquisition report for what no tuning can fix.

  • Edge refinement

    Crisp interfaces from a smooth reconstruction, under the same physics and data — with an acceptance test that refuses what the data do not support.

  • Robustness & adaptivity

    NaN guard with rollback, restarts, discrepancy stopping, nuisance parameters, ensembles and robust fidelities for unattended campaigns and real data.

Evaluate

  • Diagnostics

    Sensitivity, iter-0 gradients, filter-kernel rows, singular values, energy barriers and the data-fit paradox — why a method works, or fails.

  • Benchmarking

    The papers' protocol for any instance: paired seeds, 95 % confidence intervals, ablations, sweeps, runtime and memory, and the inverse-crime guard.

  • Baselines

    Grid (Tikhonov), L-BFGS, ADMM, Gaussian splats, Deep Decoder and closed-form references — and when a comparison is fair.

  • Visualization

    Reconstruction figures, measurement viewers, training dynamics, 3-D views, the interactive viewer and the multi-physics gallery.

  • Limitations and caveats

    The model-error floor, neural fields versus grids on linear problems, refused refinements, acquisition limits, seed variation and estimated CUDA numbers — each with what to do.

Scale up

  • Performance

    What is fast and what is slow, batched solving, whole-step compilation, the heat stencil, sharded benchmarks and how to profile.

  • GPU servers

    Sync, environment, run scripts, SLURM arrays, memory and time expectations, checkpoints.

  • Reproducing the papers

    The exact commands for the headline tables of NeTMY and NeFTY, their cost, and the caveats to check before comparing numbers.

  • FAQ — failure modes

    Center collapse, cross artifacts, high-frequency leakage, back-face artifacts, stalls at high contrast, the data-fit paradox — symptoms, checks and fixes.