sampler = grain.IndexSampler(
    num_records=len(source),
    shuffle=True,
    num_epochs=None,   # None means infinite, useful for continuous training
    seed=42,
)

        __
        __