Octopi 🐙🐙🐙¶

Built for cryo-ET

3D U-Net models designed specifically for the challenges of tomographic data — missing wedge, low SNR, and anisotropic resolution.

Autonomous architecture search

Bayesian optimization via Optuna finds the best model architecture for your data — no manual hyperparameter tuning required.

Storage-agnostic data layer

Built on CoPick — seamlessly reads tomograms and writes picks from local disk, S3, or any remote store.

HPC-ready

Submit training and architecture search jobs directly to SLURM clusters. Multi-GPU inference included out of the box.

Import Data

Set up a CoPick project and import your tomograms and initial picks.

Import data

Pick Particles

Generate initial particle picks using the interactive GUI or existing pick files.

Pick particles

Train Models

Train a 3D U-Net or run autonomous architecture search with model-explore.

Train

Inference

Segment tomograms, localize particles, and evaluate against ground truth.

Run inference

API Overview

Introduction to driving octopi programmatically from Python.

Read more

Quick Start

End-to-end particle picking in a Jupyter notebook.

Get started

Training Guide

Customize training loops, loss functions, and data augmentation via the API.

Customize

Adding New Models

Register new MONAI architectures in the octopi model registry.

Extend