ACAI: Alert-Classifying AI for the Zwicky Transient Facility

(presented at the NeurIPS 2021 ML4PS workshop)

Astronomy has been experiencing an explosive increase in the data volumes, doubling every two years or so. At the forefront of this revolution, the Zwicky Transient Facility (ZTF) – a robotic sky survey – registers millions of transient events (supernova explosions, asteroid detections, variable stars changing their brightness etc.) in the dynamic sky every (clear) night.

Alert-Classifying AI (ACAI) is an open-source deep-learning framework for the phenomenological classification of ZTF astronomical event alerts.

For more information and context, please see the Report on ACAI made with W&B.

Command-Line Interface (CLI)

Use acai.py to execute actions such as fetching training data, training a model, running hyper-parameter tuning or linting sources. Requires a config file, config.yaml - please use config.defaults.yaml as an example.

Example usages of the CLI:

• Fetch training data including the individual alert packets converted to the json format:

./acai.py fetch-datasets --fetch_json_alerts=true

• Train a "hosted" classifier overloading the default parameters set in config.yaml:

./acai.py train \
  --tag=acai_h \
  --path-labels=data/d1.csv --path-data=data/alerts \
  --optimizer=adam --lr=0.001 \
  --epochs=300 --patience=50 \
  --batch_size=128 --balance=2.5 \
  --dense_block_units=64 --conv_block_filters=16 --head_block_units=16 \
  --parallel-data-load \
  --gpu=1 \
  --verbose

• Running hyper-parameter optimization:

./acai.py sweep --tag=acai_h --path-labels=data/labels.mini1k.csv --path-data=data/alerts