Diffract¶

Diffract is a Python library for analyzing deep neural network weights and tracking how they evolve during training.

Installation¶

Diffract requires Python 3.12:

pip install diffract-core

For development, clone the repository and install with uv, which provisions Python automatically:

git clone https://github.com/Risk-AI-Research/diffract.git
cd diffract
uv sync --extra dev

Extra	Description
`torch`	PyTorch model support
`tensorflow`	TensorFlow/Keras model support
`flax`	Flax/JAX model support
`onnx`	ONNX model support
`frameworks`	TensorFlow + Flax + ONNX bundle
`viz`	Plotly visualization helpers
`pandas`	Export to pandas DataFrames
`polars`	Export to polars DataFrames
`redis`	Redis cache backend
`zarr`	Zarr storage backend (cloud-native arrays via fsspec)
`taichi`	Taichi-accelerated heavy-tailed fitting and bootstrap p-value kernels
`common`	viz + pandas + polars (recommended)
`all`	torch + viz + taichi + pandas + polars
`notebooks`	Dependencies for the example notebooks
`docs`	Build this documentation

Install extras with pip install "diffract-core[torch,common]" (quotes matter in zsh), or uv sync --extra torch --extra common in a clone.

Overview

What Diffract is, core concepts, and when to use it.

Quickstart

5-minute tour: Session → models.add → compute.apply → results.export_metrics → viz.draw.

Recipes

Focused how-tos: filtering, kernels, exports, storage backends.

API Reference

Docstring-driven reference starting from Session.