exprun 0.0.1

exprun: Automatized python pipelines to run reproducible experiments.

ExpRun

exprun allows building simple python pipelines for reproducible numerical experiments.

Quick start

With exprun, an experiment is run using an instance of the Runner class, based on a yaml configuration file, with a specified results directory, save directory and number of repeats.

from exprun import Experiment, Runner

class MyExperiment(Experiment):
    # Define the experiment here (see below)
    ...

config_path = './config.yml'
result_dir = './results/'
save_dir = './saves/'

runner = Runner()
runner.run(MyExperiment, config_path, result_dir, repeats=10)
runner.plot(MyExperiment, config_path, result_dir, save_dir)

The above code runs 10 times MyExperiment, each result is saved at result_dir as a pickle file. The results found in this directory that match the current configuration are recovered, plotted, and saved at save_dir as a pickle file.

In the MyExperiment class, four functions need to be specified by the user:

• setup(self) -> None: Set up the experiment from the information in the configuration file.
• run(self) -> dict: Perform one run of the experiment and return the results as a dict.
• cleanup(self) -> None: Clean up the experimental data.
• plot(self, results: list) -> dict: Process the results obtained from a list of all the results found in the results directory that match the configuration file. Returns the plot data that must be saved.

For instance, a simple experiment that computes the sum of two numbers within a range specified in the configuration file could be defined as follows. First, let's write the config.yml file.

min: 1
max: 10

Next, let's fill the MyExperiment class functions.

class MyExperiment(Experiment):

    def setup(self) -> None:
        from random import randrange
        # You can access the config data from self.config
        self.a = randrange(self.config["min"], self.config["max"])
        self.b = randrange(self.config["min"], self.config["max"])
    
    def run(self) -> dict:
        # You can access the data of the setup function
        return {"sum": self.a + self.b}

    def cleanup(self) -> None:
        # Nothing to clean up here
        pass

    def plot(self, results: list) -> dict:
        # Print the mean of the sums among all the results
        # found matching the current configuration file
        sums = [result["sum"] for result in results]
        print("Mean of sums:", sum(sums) / len(sums))
        return {"sums": sums}

And as simple as that, you have a reproducible experiment that can be run multiple times with different configurations. More advanced examples can be found in the examples directory.

Contribute

exprun is still in its early stages of development. Feel free to contribute by reporting any bugs on the issue page or by opening a pull request. Any feedback or contribution is welcome.

License

exprun is distributed under the MIT license.