ATM is an open source software library under the Human Data Interaction project (HDI) at MIT. It is a distributed, scalable AutoML system designed with ease of use in mind.

For a given classification problem, ATM's goal is to find

1. a classification method, such as decision tree, support vector machine, or random forest, and
2. a set of hyperparameters for that method

which generate the best classifier possible.

ATM takes in a dataset with pre-extracted feature vectors and labels as a CSV file. It then begins training and testing classifiers (machine learning models) in parallel. As time goes on, ATM will use the results of previous classifiers to intelligently select which methods and hyperparameters to try next. Along the way, ATM saves data about each classifier it trains, including the hyperparameters used to train it, extensive performance metrics, and a serialized version of the model itself.

ATM has the following features:

• It allows users to run the system for multiple datasets and multiple problem configurations in parallel.
• It can be run locally, on AWS*, or on a custom compute cluster*
• It can be configured to use a variety of AutoML approaches for hyperparameter tuning and selection, available in the accompanying library btb
• It stores models, metrics and cross validated accuracy information about each classifier it has trained.

*work in progress! See issue #40

ATM and the accompanying library BTB are under active development. We have made the transition and improvements.

This section describes the quickest way to get started with ATM on a machine running Ubuntu Linux. We hope to have more in-depth guides in the future, but for now, you should be able to substitute commands for the package manager of your choice to get ATM up and running on most modern Unix-based systems.

ATM is compatible with and has been tested on Python 2.7, 3.5, and 3.6.

1. Clone the project

   $ git clone https://github.com/hdi-project/atm.git /path/to/atm
   $ cd /path/to/atm
   

2. Install a database You will need to install the libmysqlclient-dev package (for sqlalchemy) and at least one of the following databases.

$ sudo apt install libmysqlclient-dev

• for SQLite (simpler):

$ sudo apt install sqlite3

• for MySQL:

$ sudo apt install mysql-server mysql-client

3. Install python dependencies. This will also install btb, the core AutoML library in development under the HDI project, as an egg which will track changes to the git repository. Here, usage of virtualenv is shown, but you can substitute conda or your preferred environment manager as well.

   $ virtualenv venv
   $ . venv/bin/activate
   $ pip install -r requirements.txt
   $ python setup.py install
   

Below we will give a quick tutorial of how to run ATM on your desktop. We will use a featurized dataset, already saved in data/test/pollution_1.csv. This is one of the datasets available on openml.org. More details can be found here. In this problem the goal is predict mortality using the metrics associated with the air pollution. Below we show a snapshot of the csv file. The data has 15 features and the last column is the class label.

1. Create a datarun

   $ python scripts/enter_data.py
   

   This command will create a datarun. In ATM, a "datarun" is a single logical machine learning task. If you run the above command without any arguments, it will use the default settings found in atm/config.py to create a new SQLite3 database at ./atm.db, create a new dataset instance which refers to the data above, and create a datarun instance which points to that dataset. More about what is stored in this database and what is it used for can be found here.

   The command should produce a lot of output, the end of which looks something like this:

   ========== Summary ==========
   Training data: data/test/pollution_1.csv
   Test data: <None>
   Dataset ID: 1
   Frozen set selection strategy: uniform
   Parameter tuning strategy: gp_ei
   Budget: 100 (classifier)
   Datarun ID: 1
   

   The most important piece of information is the datarun ID.

2. Start a worker

   $ python scripts/worker.py
   

   This will start a process that builds classifiers, tests them, and saves them to the ./models/ directory. The output should show which hyperparameters are being tested and the performance of each classifier (the "judgment metric"), plus the best overall performance so far.

   Classifier type: classify_logreg
   Params chosen:
           C = 8904.06127554
           _scale = True
           fit_intercept = False
           penalty = l2
           tol = 4.60893080631
           dual = True
           class_weight = auto
   
   Judgment metric (f1): 0.536 +- 0.067
   Best so far (classifier 21): 0.716 +- 0.035
   

   Occasionally, a worker will encounter an error in the process of building and testing a classifier. When this happens, the worker will print error data to the terminal, log the error in the database, and move on to the next classifier.

And that's it! You can break out of the worker with Ctrl+c and restart it with the same command; it will pick up right where it left off. You can also run the command simultaneously in different terminals to parallelize the work -- all workers will refer to the same ModelHub database. When all 100 classifiers in your budget have been built, all workers will exit gracefully.

ATM's default configuration is fully controlled by atm/config.py. Our documentation will cover the configuration in more detail, but this section provides a brief overview of how to specify the most important values.

If you want to use the system for your own dataset, convert your data to a CSV file similar to the example shown above. The format is:

• Each column is a feature (or the label)
• Each row is a training example
• The first row is the header row, which contains names for each column of data
• A single column (the target or label) is named class

Next, you'll need to use enter_data.py to create a dataset and datarun for your task.

The script will look for values for each configuration variable in the following places, in order:

1. Command line arguments
2. Configuration files
3. Defaults specified in atm/config.py

That means there are two ways to pass configuration to the command.

1. Using YAML configuration files

   Saving configuration as YAML files is an easy way to save complicated setups or share them with team members.

   You should start with the templates provided in atm/config/templates and modify them to suit your own needs.

   $ mkdir config
   $ cp atm/config/templates/*.yaml config/
   $ vim config/*.yaml
   

   run.yaml contains all the settings for a single dataset and datarun. Specify the train_path to point to your own dataset.

   sql.yaml contains the settings for the ModelHub SQL database. The default configuration will connect to (and create if necessary) a SQLite database at ./atm.db relative to the directory from which enter_data.py is run. If you are using a MySQL database, you will need to change the file to something like this:

   dialect: mysql
   database: atm
   username: username
   password: password
   host: localhost
   port: 3306
   query:
   

   aws.yaml should contain the settings for running ATM in the cloud. This is not necessary for local operation.

Once your YAML files have been updated, run the datarun creation script and pass it the paths to your new config files:

$ python scripts/enter_data.py --sql-config config/sql.yaml \
> --aws-config config/aws.yaml \
> --run-config config/run.yaml

2. Using command line arguments

   You can also specify each argument individually on the command line. The names of the variables are the same as those in the YAML files. SQL configuration variables must be prepended by sql-, and AWS config variables must be prepended by aws-.

   Using command line arguments is convenient for quick experiments, or for cases where you need to change just a couple of values from the default configuration. For example:

   $ python scripts/enter_data.py --train-path ./data/my-custom-data.csv --selector bestkvel
   

   You can also use a mixture of config files and command line arguments; any command line arguments you specify will override the values found in config files.

Once you've created your custom datarun, start a worker, specifying your config files and the datarun(s) you'd like to compute on.

$ python scripts/worker.py --sql-config config/sql.yaml \
> --aws-config config/aws.yaml --dataruns 1

It's important that the SQL configuration used by the worker matches the configuration you passed to enter_data.py -- otherwise, the worker will be looking in the wrong ModelHub database for its datarun!

BTB, for Bayesian Tuning and Bandits, is the core AutoML library in development under the HDI project. BTB exposes several methods for hyperparameter selection and tuning through a common API. It allows domain experts to extend existing methods and add new ones easily. BTB is a central part of ATM, and the two projects were developed in tandem, but it is designed to be implementation-agnostic and should be useful for a wide range of hyperparameter selection tasks.

Featuretools is a python library for automated feature engineering. It can be used to prepare raw transactional and relational datasets for ATM. It is created and maintained by Feature Labs and is also a part of the Human Data Interaction Project.