Faucet ML

Faucet ML is a Python package that enables high speed mini-batch data reading & preprocessing from BigQuery for machine learning model training.

Faucet ML is designed for cases where:

• Datasets are too large to fit into memory
• Model training requires mini-batches of data (SGD based algorithms)

Features:

• High speed batch data reading from BigQuery
• Automatic feature identification and preprocessing via. PyTorch
• Integration with Feast feature store (coming soon)

Installation

pip install faucetml

More about Faucet

Many training datasets are too large to fit in memory, but model training would benefit from using all of the training data. Naively issuing 1 query per mini-batch of data is unnecessarily expensive due round-trip network costs. Faucet is a library that solves these issues by:

• Fetching large "chunks" of data in non-blocking background threads
  • where chunks are much larger than mini-batches, but still fit in memory
• Caching chunks locally
• Returning mini-batches from cached chunks in O(1) time

Examples

See examples for detailed ipython notebook examples on how to use Faucet.

# initialize the client
fml = get_client(
    datastore="bigquery",
    credential_path="bq_creds.json",
    table_name="my_training_table",
    ds="2020-01-20",
    epochs=2,
    batch_size=1024
    chunk_size=1024 * 10000,
    test_split_percent=20,
)

# train & test
for epoch in range(2):

    # training loop
    fml.prep_for_epoch()
    batch = fml.get_batch()
    while batch is not None:
        train(batch)
        batch = fml.get_batch()

    # evaluation loop
    fml.prep_for_eval()
    batch = fml.get_batch(eval=True)
    while batch is not None:
        test(batch)
        batch = fml.get_batch(eval=True)

Future features

• Support more data warehouses (redshift, hive, etc.)
• Support reading features & preprocessing specs from Feast

Suggestions for other features? Open an issue and let us know.