pyexperiment

Pyexperiment facilitates the development of small, reproducible experiments with minimal boilerplate code. Consider the following example, implementing a simple program that stores numbers and computes their sum:

from pyexperiment import experiment, state, conf, log

conf['pyexperiment.save_state'] = True
conf['pyexperiment.load_state'] = True
conf['message'] = "The stored numbers are: "

def store(number):
    """Store a number"""
    if 'numbers' not in state:
        log.debug("Initialize state['numbers'] to empty list")
        state['numbers'] = []

    log.debug("Store number: %s", number)
    state['numbers'].append(float(number))

 def show():
     """Show the stored numbers and compute their sum"""
     if not 'numbers' in state:
         print('No numbers stored yet')
         return

     print(conf['message'] + str(state['numbers']))
     with log.timed("sum"):
             total = sum(state['numbers'])
     print("The total is: " + str(total))

if __name__ == '__main__':
    experiment.main(default=show,
                    commands=[store, show])

Pyexperiment makes it easy to experiment with this code from the command line:

$ ./numbers store 42
$ ./numbers store 3.14
$ ./numbers
The stored numbers are: [42.0, 3.14]
The total is: 45.14
$ ./numbers -o message "Numbers: "
Numbers: [42.0, 3.14]
The total is: 45.14
$ ./numbers -v
[DEBUG   ] [0.069s] Start: './numbers -v'
[DEBUG   ] [0.069s] Time: '2015-08-14 14:23:00.027378'
[INFO    ] [0.075s] Loading state from file 'experiment_state.h5f'
The stored numbers are: [42.0, 3.14]
[DEBUG   ] [0.076s] sum took 0.000025s
The total is: 45.14
[DEBUG   ] [0.078s] No need to save unchanged state
[DEBUG   ] [0.078s] End: './numbers -v show'
[DEBUG   ] [0.078s] Time: '2015-08-14 14:23:00.037124'
[DEBUG   ] [0.078s] Took: 0.010s
$ ./numbers -h
usage: numbers [-h] [-c CONFIG] [-o key value] [-i]
               [--verbosity {DEBUG,INFO,WARNING,ERROR,CRITICAL}] [-v]
               [-j PROCESSES]
               [{store,show,help,test,show_tests,show_config,save_config,show_state,show_commands}]
               [argument [argument ...]]

Thanks for using numbers.

positional arguments:
  {store,show,help,test,show_tests,show_config,save_config,show_state,show_commands}
                        choose a command to run, running show by default
  argument              argument to the command

optional arguments:
  -h, --help            show this help message and exit
  -c CONFIG, --config CONFIG
                        specify a configuration file
  -o key value, --option key value
                        override a configuration option
  -i, --interactive     drop to interactive prompt after COMMAND
  --verbosity {DEBUG,INFO,WARNING,ERROR,CRITICAL}
                        choose the console logger's verbosity
  -v                    shortcut for --verbosity DEBUG
  -j PROCESSES, --processes PROCESSES
                        set number of parallel processes used

available commands:

  store:                Store a number
  show (default):       Show the stored numbers and compute their sum
  help:                 Shows help for a specified command
  test:                 Run tests for the experiment
  show_tests:           Show available tests for the experiment
  show_config:          Print the configuration
  save_config:          Save a configuration file to a filename
  show_state:           Shows the contents of the state loaded by the configuration or from the file specified as an argument
  show_commands:        Print the available commands

Motivation

There is no shortage of great Python libraries for command line interfaces, logging, configuration file management, persistent state, parallelism, or plotting. When writing small scripts for quick experiments though, it's often too much effort to configure these components, and one ends up rewriting the same setup code over and over again.

Pyexperiment fixes this by providing a simple way to jump start short experiments. Importing pyexperiment will give you:

• A basic command line interface that allows calling arbitrary functions (and passing arguments) from the command prompt, providing help text derived from the functions' docstrings and zsh/bash autocompletion (based on the standard library's argparse and argcomplete).
• A simple configuration management with an easy way to provide default values (based on the excellent configobj library).
• A thread-safe logger with configurable logging levels, timing utilities with statistics, and rotating log files (based on the standard library's logging module).
• Persistent state with platform independent, configurable, (optionally rotating) state files that are compatible with many other programs (based on h5py).
• Parallel execution of replicates.
• A sensible setup for plotting (based on matplotlib, and optionally seaborn), with configurable defaults and asynchronous plotting.
• Many other bits and pieces that might come in handy...

As a design principle, pyexperiment's components come ready to use without any further configuration. Inevitably then, the choices made in this setup are opinionated and may or may not fit your personal taste. Feel free to start a discussion on the issues page.

For more documentation, see the automatically generated pages here. For more usage examples, check the examples folder.

Installation

The easiest way to install pyexperiment is from pypi, just call pip install --user pyexperiment (alternatively, use pip install pyexperiment in a virtualenv, or prepend sudo for system wide installation).

The pyexperiment package has a few external dependencies (as you can see in the requirements.txt):

• six
• configobj
• numpy
• h5py
• matplotlib
• lockfile
• toolz
• IPython (optional, allows using IPython with the --interactive command)
• argcomplete (optional, adds activate_autocompletion command)
• seaborn (optional, adds more plotting options)

If you install (the h5py dependency) from pypi, you may need to install libhdf5 first, e.g., by running sudo apt-get install libhdf5-dev. You may also find that you need to install cython first, e.g., by running either sudo apt-get install Cython or pip install Cython.

Reproducible experiments

To keep your experiments reproducible and avoid dependency problems, it is a good idea to automate the setup of your development environment, e.g., using a Vagrant box, or - in many cases even better - a Docker image. To get started with pyexperiment using Vagrant or Docker, you can use the Vagrantfile and setup script here, or the Dockerfile and setup scripts here.

License

The pyexperiment package is licensed under an MIT licence (see the LICENSE).