API reference

Every public class and function, with its signature, tensor shapes and defaults, as written in the source.

The public API lives in three namespaces: everything for fitting coordinate/value pairs is exported at the top level, and image compression and radiance fields have their own.

Namespace	Contents
inro	Models, datasets, train, evaluate, chunked_inference, losses, metrics, quantization, plotting and utilities
inro.compression	encode_futon, decode_futon, encode_pil, decode_pil, evaluate_compression, arithmetic coding
inro.nerf	BlenderDataset, RadianceField, renderers, train, evaluate, render_image, render_views, camera helpers

• FUTON, bases and combiners

  ---

  FUTON, its six bases and the features they return, the combiners, the solvers and the error bounds.

• inro.models

  ---

  Every other architecture, with its layers and encodings.

• Datasets

  ---

  Coordinate datasets for images, volumes and MRI, Deep Image Prior, and super-resolution's blocks with their bicubic operator.

• Training and evaluation

  ---

  train, evaluate, chunked_inference.

• Metrics and losses

  ---

  PSNR, SSIM, MS-SSIM, LPIPS, NMSE, IoU, bits per pixel and super-resolution's metrics; rate-distortion, SDF, super-resolution and entropy losses.

• Quantization

  ---

  Uniform quantization of a model's parameters.

• Compression

  ---

  FUTON as a codec, PIL baselines, their evaluation, and arithmetic coding.

• Radiance fields

  ---

  The inro.nerf package.

• Utilities

  ---

  Module specs, parameter counts, logging, serialization and the seaborn line plot.

• Custom components

  ---

  Your own basis, combiner, head, loss, metric, dataset and model, each with its contract.

Reading the docstrings

(N, C) is a flat batch and (..., C) allows any leading dimensions. C is the number of coordinate axes, D of output channels, K_c of basis functions on axis c, and R a rank; the conventions page has the full notation.