The Reproducible Python Library

Keep track of your results.

Ever produced a result for a paper, only to realize a few months later that you could not reproduce it? That you had no idea which version of the code, and which parameter values were used back then?

The reproducible library, developped by the Cognitive Neuro-Robotics Unit at the Okinawa Institute of Science and Technology (OIST), aims to provide an easy way to gather and save important information about the context in which a result was computed. This includes details about the OS, the Python version, the time, the git commit, the command-line arguments, hashes of input and output files, and any user provided data.

Other Python libraries doing just that exists such as Recipy and Sumatra. And they are good. Do try them. They each have their own design philosophy, which proved to be difficult to interface with some of the workflows of the Cognitive Neuro-Robotics Unit lab at OIST.

With reproducible the goal was to have a small non-intrusive library allowing precise control over the data collected and how to output it. In particular, the goal was to have the tracking info sitting next to—or better, directly embedded in—the result files. That makes sending results to collaborators or packaging them for publication straightforward.

The reproducible library is licensed under the LGPL version 3, to allow you to use it along-side code that use other licenses.

The library is in beta; expect some changes. Python 3.5 or later is officially supported, but for the time being, the code runs on 2.7 as well.

Install

pip install reproducible

Instant Tutorial

Say this is your code, which is fully committed using git:

import random
import pickle

def walk(n):
    """A simple random walk generator"""
    steps = [0]
    for i in range(n):
        steps.append(steps[-1] + random.choice([-1, 1]))
    return steps

if __name__ == '__main__':
    random.seed(1)
    results = walk(10)
    with open('results.pickle', 'wb') as f:
        pickle.dump(results, f)

To add reproducible tracking:

import random
import pickle

import reproducible

def walk(n):
    """A simple random walk generator"""
    steps = [0]
    for i in range(n):
        steps.append(steps[-1] + random.choice([-1, 1]))
    return steps

if __name__ == '__main__':
    # create a reproducible.Context instance, that will hold all the
    # tracked data.
    context = reproducible.Context()

    # recording git repository state
    # here we are okay with running our code with uncommitted changes, but
    # we record a diff of the changes in the tracked data.
    context.add_repo(path='.', allow_dirty=True, diff=True)

    # recording parameters; this is not necessarily needed, as the code state
    # is recorded, but it is convenient.
    seed = 1
    random.seed(seed)
    context.add_data('seed', seed)

    # add_data return the provided value (here 10), so you can do this:
    n = reproducible.add_data('n', 10)
    results = walk(n)

    with open('results.pickle', 'wb') as f:
        pickle.dump(results, f)
    # recording the SHA1 hash of the output file
    context.add_file('results.pickle', category='output')

    # you can examine the tracked data and add or remove from it at any moment
    # with `context.data`: it is a simple dictionary. For instance, the
    # cpu info is quite detailed. Let's remove it to keep the yaml output short.
    context.data.pop('cpuinfo')

    # exporting the provenance data to disk
    context.export_yaml('results_prov.yaml')

This is the resulting yaml file output containing the tracking data:

argv: [example_after.py]
data: {n: 10, seed: 1}
files:
  output:
    results.pickle: {mtime: 1531381834.0666547, sha256: 395d8846640c012e3e5c642e7737173a1a74120275b37fa2ded13a211df3264e}
packages: [gitdb2==2.0.3, GitPython==2.1.10, pip==10.0.1, py-cpuinfo==4.0.0, PyYAML==4.2b4,
  reproducible==0.1.2, setuptools==39.0.1, smmap2==2.0.3]
platform: Darwin-17.6.0-x86_64-i386-64bit
python:
  branch: ''
  compiler: Clang 9.1.0 (clang-902.0.39.2)
  implementation: CPython
  revision: ''
  version: ['3', '7', '0']
repositories:
  .: {diff: null, dirty: false, hash: 88c1de4ba5fb5cb2564b60245f26d3226ecb20c9, version: git
      version 2.18.0}
timestamp: ['2018-07-12T07:50:34.033829Z']

See also the The API Reference.

Roadmap

• Retrieve GPU information.
• More configurability.
• Optionally capture input, output (sys.stderr, sys.stdout).
• Easy disabling/reenabling of reproducible
• optional SHA256 in the filename of external files

Changelog

version 0.4.1, 20200803

• CPU info collection disabled by default for module-level Context instance.
• new add_cpu_info() function.
• add missing module-level functions that were only accesible from instances before.
• fix leaked-semaphore issue.

version 0.4.0, 20200802

• new functions sha256(), untrack_file(), find_editable_repos(), add_editable_repos().
• fix tests.

version 0.3.0, 20190703 This version introduces API and logic-breaking changes.

• add_file() overwrites by default now, and category is now an optional argument.
• context.data() becomes context.data.
• Context(repo_path='.', allow_dirty=False, allow_untracked=False, diff=True, cpuinfo=True) becomes Context(cpuinfo=True, pip_packages=True): add_repo() needs to be called explicitly now, and pip_packages queries can be made optional.
• reset() does not accept any arguments anymore; remembers __init__() argument values instead.
• fixed missing reproducible.add_pip_packages().

version 0.2.4, 20170809

• hotfix for Python 2.7---because I am stupid.

version 0.2.3, 20170809

• add json(), yaml() and requirements() function to access the result of export functions programmatically.
• YAML output is now generated using yaml.safe_dump rather than yaml.dump. Leads to safer and simpler output.

version 0.2.2, 20170717

• fix for deprecated save_yaml(), save_json() functions.

version 0.2.1, 20170717

• add readme, license to pypi package.

version 0.2.0, 20170717

• renamed save_json() and save_yaml() as export_json() and export_yaml(). The old name remain for now with a deprecation warning.
• Context instances for more flexible, non-module level, behavior, much like the Random instances of the standard random module.
• reproducible.function_args() function to retrieve arguments from inside a function.
• reproducible.reset() function for clearing tracked data.
• reproducible.export_requirements() to create requirements files from the retrieved list of installed packages.
• Fix import of the freeze command from the pip package.
• Updated readme: yaml output of the example, roadmap, changlog.

version 0.1.2, 20170611

• Various bug fixes.
• The save_json() and save_yaml() functions now return the SHA256 hash of the file they produce.

version 0.1.1, 20170608

• reproducible.data() function to access and modify the collected data.
• more unit tests

version 0.1.0, 20170607

• first version: add_repo(), add_file(), add_data(), add_random_state(), git_info(), git_dirty(), save_json(), save_yaml() functions.