Caliban simplifies research workflows by enabling reproducible, isolated execution of numerical experiments, from local workstations to cloud platforms. It targets machine learning researchers and engineers, facilitating seamless transition from prototype to large-scale cloud deployments with robust job tracking and management.

How It Works

Caliban leverages Docker to create consistent, isolated execution environments for experiments. It builds Docker images based on project dependencies (specified in requirements.txt or setup.py) and executes code within these containers. This approach ensures that experiments run identically across local machines and cloud environments, eliminating "it works on my machine" issues and simplifying GPU/dependency management.

Quick Start & Requirements

• Install Caliban: pip install caliban
• Prerequisites: Docker (running), Python >= 3.6.
• Local MNIST training: git clone https://github.com/google/caliban.git && cd caliban/tutorials/basic && caliban run --nogpu mnist.py
• Documentation: Read The Docs

Highlighted Details

• Supports local execution and submission to Google Cloud AI Platform.
• Enables parameter sweeping via JSON configuration or stdin.
• Provides caliban shell and caliban notebook for interactive development within the execution environment.
• Offers caliban cluster for GKE integration and caliban status for job management.

Maintenance & Community

This is a research project, not an official Google product. Issues and contributions are managed via GitHub Issues.

Licensing & Compatibility

Licensed under the Apache License, Version 2.0. Compatible with commercial use.

Limitations & Caveats

Described as a research project with potential bugs and "sharp edges." Cloud submission currently focuses on Google Cloud AI Platform.