autobazaar 0.2.1

The Machine Learning Bazaar

An open source project from Data to AI Lab at MIT.

AutoBazaar

• License: MIT
• Development Status: Pre-Alpha
• Documentation: https://HDI-Project.github.io/AutoBazaar/
• Homepage: https://github.com/HDI-Project/AutoBazaar
• Paper: https://arxiv.org/pdf/1905.08942.pdf

Overview

AutoBazaar is an AutoML system created using The Machine Learning Bazaar, a research project and framework for building ML and AutoML systems by the Data To AI Lab at MIT. See below for more references.

It comes in the form of a python library which can be used directly inside any other python project, as well as a CLI which allows searching for pipelines to solve a problem directly from the command line.

Install

Requirements

AutoBazaar has been developed and tested on Python 3.5, 3.6 and 3.7

Also, although it is not strictly required, the usage of a virtualenv is highly recommended in order to avoid interfering with other software installed in the system where AutoBazaar is run.

Install with pip

The easiest and recommended way to install AutoBazaar is using pip:

pip install autobazaar

This will pull and install the latest stable release from PyPI.

If you want to install from source or contribute to the project please read the Contributing Guide.

Data Format

AutoBazaar works with datasets in the D3M Schema Format as input.

This dataset schema, developed by MIT Lincoln Labs Laboratory for DARPA's Data-Driven Discovery of Models (D3M) Program, requires the data to be in plainly readable formats such as CSV files or JPG images, and to be set within a folder hierarchy alongside some metadata specifications in JSON format, which include information about all the data contained, as well as the problem that we are trying to solve.

For more details about the schema and about how to format your data to be compliant with it, refer to the Schema Documentation

As an example, you can browse some datasets which have been included in this repository for demonstration purposes:

• 185_baseball: Single Table Regression
• 196_autoMpg: Single Table Classification

Quickstart

In this short tutorial we will guide you through a series of steps that will help you getting started with AutoBazaar using its CLI command abz.

For more details about its usage and the available options, please execute abz --help on your command line.

1. Prepare your Data

Make sure to have your data prepared in the Data Format explained above inside and uncompressed folder in a filesystem directly accessible by AutoBazaar.

In order to check, whether your dataset is available and ready to use, you can execute the abz command in your command line with its list subcommand. If your dataset is in a different place than inside a folder called data within your current working directory, do not forget to add the -i argument to your command indicating the path to the folder that contains your dataset.

Assuming that the data is inside a folder called input within your current folder, you can run:

$ abz list -i /path/to/your/datasets/folder

The output should be a table which includes the details of all the datasets found inside the indicated directory:

             data_modality                task_type task_subtype             metric size_human  train_samples
dataset
185_baseball  single_table           classification  multi_class            f1Macro       148K           1073
196_autoMpg   single_table               regression   univariate   meanSquaredError        32K            298
30_personae           text           classification       binary                 f1       1,4M            116
32_wikiqa      multi_table           classification       binary                 f1       4,9M          23406
60_jester     single_table  collaborative_filtering               meanAbsoluteError        44M         880719

Note: If you see an error saying that No matching datasets found, please review your dataset format and make sure to have indicated the right path.

For the rest of this quickstart, we will be using the 185_baseball dataset that you can find inside the input folder contained in this repository.

2. Start the search process

Once your data is ready, you can start the AutoBazaar search process using the abz search command. To do this, you will need to provide again the path to where your datasets are contained, as well as the name of the datasets that you want to process.

For example if you want to search for the best

$ abz search -i /path/to/your/datasets/folder name_of_your_dataset

This will evaluate the default pipeline without performing additional tuning iteration on it.

In order to start an actual tuning process, you will need to provide at least one of the following additional options:

• -b, --budget: Maximum number of tuning iterations to perform.
• -t, --timeout: Maximum time that the system needs to run, in seconds.
• -c, --checkpoints: Comma separated string containing the different checkpoints where the best pipeline so far must be stored and evaluated against the test dataset. There must be no spaces between the checkpoint times. For example, to store the best pipeline every 10 minutes until 30 minutes have passed, you would use the option -c 600,1200,1800.

For example, to search process the 185_baseball dataset during 30 seconds evaluating the best pipeline so far every 10 seconds but with a maximum of 10 tuning iterations, we would use the following command:

abz search 185_baseball -c10,20,30 -b10

For further details about the available options, please execute abz search --help in your terminal.

3. Explore the results

Once the AutoBazaar has finished searching for the best pipeline, a table will be printed in stdout with a summary of the best pipeline found for each dataset. If multiple checkpoints were provided, details about the best pipeline in each checkpoint will also be included.

The output will be a table similar to this one:

                                          pipeline     score      rank  cv_score   metric data_modality       task_type task_subtype    elapsed  iterations  load_time  trivial_time  fit_time    cv_time error  step
dataset
185_baseball  fce28425-e45c-4620-9d3c-d329b8684bea  0.316961  0.682957  0.317043  f1Macro  single_table  classification  multi_class  10.024457         0.0   0.011041      0.026212       NaN        NaN  None  None
185_baseball  f7428924-79ee-439d-bc32-998a9efea619  0.675132  0.390927  0.609073  f1Macro  single_table  classification  multi_class  21.412262         1.0   0.011041      0.026212   9.99484        NaN  None  None
185_baseball  397780a5-6bf6-48c9-9a85-06b0d08c5a9d  0.675132  0.357361  0.642639  f1Macro  single_table  classification  multi_class  31.712946         2.0   0.011041      0.026212   9.99484  12.618179  None  None

Alternatively, a -r option can be passed with the name of a CSV file, and the results will be stored there:

abz search 185_baseball -c10,20,30 -b10 -r results.csv

What's next?

For more details about AutoBazaar and all its possibilities and features, please check the project documentation site!

Credits

AutoBazaar is an open-source project from the Data to AI Lab at MIT built by the following team:

• Micah Smith [e-mail]
• Carles Sala [e-mail]
• Max Kanter [e-mail]
• Kalyan Veeramachaneni [e-mail]

Citing AutoBazaar

If you use AutoBazaar for your research, please consider citing the following paper (https://arxiv.org/pdf/1905.08942.pdf):

@article{smith2019mlbazaar,
  author = {Smith, Micah J. and Sala, Carles and Kanter, James Max and Veeramachaneni, Kalyan},
  title = {The Machine Learning Bazaar: Harnessing the ML Ecosystem for Effective System Development},
  journal = {arXiv e-prints},
  year = {2019},
  eid = {arXiv:1905.08942},
  pages = {arxiv:1904.09535},
  archivePrefix = {arXiv},
  eprint = {1905.08942},
}

History

0.2.1 - 2020-06-22

This release curates some of the dependencies and updates the example datasets and documentation.

0.2.0 - 2019-11-26

Second Release:

• Improved CLI interface
• Improved Dataset support
• New Docker image
• Newer dependencies

This is the version used to generate the results explained in the third version of The Machine Learning Bazaar Paper

0.1.0 - 2019-06-24

First Release.

This is a slightly cleaned up version of the software used to generate the results explained in the first version of The Machine Learning Bazaar Paper