autonml 0.1.2

AutonML : CMU's AutoML System

CMU TA2 (Built using DARPA D3M ecosystem)

Autoⁿ ML is an automated machine learning system developed by CMU Auton Lab to power data scientists with efficient model discovery and advanced data analytics. Auton ML also powers the D3M Subject Matter Expert (SME) User Interfaces such as Two Ravens http://2ra.vn/.

Taking your machine learning capacity to the nth power.

D3M dataset

• Any dataset to be used should be in D3M dataset format (directory structure with TRAIN, TEST folders and underlying .json files).
• Example available of a single dataset here
• More datasets available here
• Any non-D3M data can be converted to D3M dataset. (See section below on "Convert raw dataset to D3M dataset").

Convert raw dataset to D3M dataset

pip install d3m
python create_d3m_dataset.py <train_data.csv> <test_data.csv> <label> <metric> -t classification <-t ...>

Detailed description of dataset type(s), task type(s) and metrics provided here.

Run in search mode

We can run the AutonML pipeline in two ways. It be run as a standalone CLI command, accessed via the autonml_main command. This command takes four arguments, listed below:

• Path to the data directory (must be in D3M format)
• Output directory where results are to be stored. This directory will be dynamically created if it does not exist.
• Timeout (measured in minutes)
• Number of CPUs to be used
• Path to problemDoc.json (see example below)

INPUT_DIR=/home/<user>/d3m/datasets/185_baseball_MIN_METADATA
OUTPUT_DIR=/output
TIMEOUT=2
NUMCPUS=8
PROBLEMPATH=${INPUT_DIR}/TRAIN/problem_TRAIN/problemDoc.json

autonml_main ${INPUT_DIR} ${OUTPUT_DIR} ${TIMEOUT} ${NUMCPUS} ${PROBLEMPATH} 

The above script will do the following-

1. Run search for best pipelines for the specified dataset using TRAIN data.
2. JSON pipelines (with ranks) will be output in JSON format at /output/<search_dir>/pipelines_ranked/
3. CSV prediction files of the pipelines trained on TRAIN data and predicted on TEST data will be available at /output/<search_dir>/predictions/
4. Training data predictions (cross-validated mostly) are produced in the current directory as /output/<search_dir>/training_predictions/<pipeline_id>_train_predictions.csv.
5. Python code equivalent of executing a JSON pipeline on a dataset produced at /output/<search_dir>/executables/

An example -

python ./output/6b92f2f7-74d2-4e86-958d-4e62bbd89c51/executables/131542c6-ea71-4403-9c2d-d899e990e7bd.json.code.py 185_baseball predictions.csv