memento-ml 1.1.0

A Python library for running computationally expensive experiments

MEMENTO

MEMENTO is a Python library for running computationally expensive experiments.

Running complex sets of machine learning experiments is challenging and time-consuming due to the lack of a unified framework. This leaves researchers forced to spend time implementing necessary features such as parallelization, caching, and checkpointing themselves instead of focussing on their project. To simplify the process, we introduce MEMENTO, a Python package that is designed to aid researchers and data scientists in the efficient management and execution of computationally intensive experiments. MEMENTO has the capacity to streamline any experimental pipeline by providing a straightforward configuration matrix and the ability to concurrently run experiments across multiple threads.

If you need to run a large number of time-consuming experiments MEMENTO can help:

• Structure your configuration
• Parallelize experiments across CPUs
• Save and restore results
• Checkpoint in-progress experiments
• Send notifications when experiments fail or finish

Getting Started

MEMENTO is officially available on PyPl. To install the package:

Install

pip install memento-ml

The Configuration Matrix

The core of MEMENTO is a configuration matrix that describes the list of experiments you want MEMENTO to run. This must contain a key parameters which is itself a dict, this describes each paramter you want to vary for your experiments and their values.

As an example let's say you wanted to test a few simple linear classifiers on a number of image recognition datasets. You might write something like this:

Don't worry if you're not working on machine learning, this is just an example.

matrix = {
  "parameters": {
    "model": [
      sklearn.svm.SVC,
      sklearn.linear_model.Perceptron,
      sklearn.linear_model.LogisticRegression
    ],
    "dataset": ["imagenet", "mnist", "cifar10", "quickdraw"]
  }
}

MEMENTO would then generate 12 configurations by taking the cartesian product of the parameters.

Frequently you might also want to set some global configuration values, such as a regularization parameter or potentially even change your preprocessing pipeline. In this case MEMENTO also accepts a "settings" key. These settings apply to all experiments and can be accessed from the configuration list as well as individual configurations.

matrix = {
  "parameters": ...,
  "settings": {
    "regularization": 1e-1,
    "preprocessing": make_preprocessing_pipeline()
  }
}

You can also exclude specific parameter configurations. Returning to our machine learning example, if you know SVCs perform poorly on cifar10 you might decide to skip that experiment entirely. This is done with the "exclude" key:

matrix = {
  "parameters": ...,
  "exclude": [
    {"model": sklearn.svm.SVC, "dataset": "cifar10"}
  ]
}

Running an experiment

Along with a configuration matrix you need some code to run your experiments. This can be any Callable such as a function, lambda, class, or class method.

from memento import Memento, Config, Context

def experiment(context: Context, config: Config):
  classifier = config.model()
  dataset = fetch_dataset(config.dataset)

  classifier.fit(*dataset)

  return classifier

Memento(experiment).run(matrix)

You can also perform a dry run to check you've gotten the matrix correct.

Memento(experiment).run(matrix, dry_run=True)

Running configurations:
  {'model': sklearn.svm.SVC, 'dataset': 'imagenet'}
  {'model': sklearn.svm.SVC, 'dataset': 'mnist'}
  {'model': sklearn.svm.SVC, 'dataset': 'cifar10'}
  {'model': sklearn.svm.SVC, 'dataset': 'quickdraw'}
  {'model': sklearn.linear_model.Perceptron, 'dataset': 'imagenet'}
  ...
Exiting due to dry run

Code demo

• Code demo can be found here.
• MEMENTO does not depend on any ML packages, e.g., scikit-learn. The scikit-learn and jupyterlab packages are required to run the demo (./demo/*).

pip install memento-ml scikit-learn jupyterlab

Roadmap

• Finish HPC support
• Improve result serialisation
• Improve customization for notification

Contributors

• Zac Pullar-Strecker
• Feras Albaroudi
• Liam Scott-Russell
• Joshua de Wet
• Nipun Jasti
• James Lamberton
• Xinglong (Luke) Chang
• Liam Brydon
• Ioannis Ziogas
• Katharina Dost
• Joerg Wicker

License

MEMENTO is licensed under the 3-Clause BSD License license.