Experimenter

Use git tags to log experiments and the exact code that was used to run those experiments. The goal is to make every experiment completely reproducible. To install use

pip install experimenter

All contributions are welcome. Experimenter was inspired by Ali's and Charles' blog posts.

Use Cases

• It is the case that you might make some changes in your machine learning code, not necessarily worth committing, and spawn a process to run an experiment that uses these changes. However, by the time the experiment has finished you don't know what changes you were testing (presumably because you made more changes in the meanwhile).
• You need a distributed way of collecting all the experiments (parameters and results), making sure that they have been run on the exact same version of the code.

Usage

Create an ExperimentLogger object, passing the parameters of the experiment. When the experiment is finished, call the record_results() method of that object, ie:

with ExperimentLogger(name="NameOfExperiment", parameters=parameters_dict) as experiment_logger:
    ...
    experiment_logger.record_results(dict_of_results)

Behind the scenes, a git tag will be created (committing any changes you may have in the working tree, into a different branch). The tag will have a name of the form exp_NameOfExperiment_timestamp and in the message it will have a JSON representation of the parameters and the results (when/if recorded). The working state of the current branch will seemingly remain unaffected. Note that this is not thread-safe.

If no result is recorded (ie record_results is not called) within the with then the experiment will be deleted upon exit from that block. This is useful when stopping experiments or when experiments fail before finishing.

There is a command-line tool that helps with retrieving tests. In the git folder, run the command

> experimenter -c SHAofCodeState

to retrieve all the experiments that happened with the give code version. If -c is not provided then all experiments will be shown. If -s is provided only experiments that have results will be shown. The strict command is off by default. Use --help for more information.

TODO

• A command line tool for starting/stopping experiments.
• Auto-push tags method in ExperimentLogger