sql-grain 0.2.2

Sqlite databases as Grain datasets.

🌾 SQL Grain

SQLite databases as Grain data sources.

sql-grain lets you prototype ML data pipelines using SQL queries before committing to a production data format. Define your training examples with expressive SQL—joins, window functions, filtering—and iterate quickly without preprocessing. When you're ready to scale, convert to ArrayRecord or similar formats; sql-grain is not designed for large-scale training.

from sqlgrain import Sqlite3DataSource
import grain

source = Sqlite3DataSource(
    "data.db",
    key_query="SELECT id FROM users",
    record_query="SELECT item FROM purchases WHERE user_id = :id ORDER BY timestamp",
)
dataset = grain.MapDataset.source(source).shuffle().batch(32)

Converting to ArrayRecord

Once you're ready to run larger experiments, convert the dataset to ArrayRecords using the to_array_record function which serializes records in msgpack format with native support for NumPy arrays.

from sqlgrain import to_array_record

to_array_record(source, "output/", shard_every=1000)

Grain's ArrayRecordDataSource reads the raw bytes, and we need to integrate a decoder into the data pipeline. SQL Grain's decode_record takes care of that, matching the behavior of to_array_record.

from sqlgrain import from_array_record, decode_record

ar_source, metadata = from_array_record("output/")
dataset = grain.MapDataset.source(ar_source).map(decode_record).batch(32)

to_array_record is agnostic to the type of the records, and you can also serialize datasets, e.g., for pre-batching.