Getting Started

Installation

alp-data is available on PyPI.

uv add alp-data

or with pip:

pip install alp-data

from alp_data import Beans

# Load 'train' split of BEANS dataset at 16kHz sample rate
# Resampling is done on the fly with librosa.resample
beans = Beans(split="train", sample_rate=16000)

print(len(beans))

# Iterate over all samples
for sample in beans:
    print(sample["audio"].shape)
    break

# Fetch a single sample
sample = beans[0]
print(sample["audio"].shape)

# Streaming only
beans_streaming = Beans(split="train", streaming=True)  # loads faster
print(len(beans_streaming))  # Throws an error since length is unknown in streaming mode

# Iterate over all samples in streaming mode
for sample in beans_streaming:
    print(sample["audio"].shape)
    break

PyTorch DataLoader with num_workers > 0 requires spawn

When wrapping an alp-data Dataset in a PyTorch DataLoader with num_workers > 0, you must use the "spawn" multiprocessing start method instead of the default "fork" on Linux. alp-data datasets hold fsspec / gcsfs / s3fs handles that are not safe to inherit across a fork — workers may deadlock or return corrupted audio. Either call torch.multiprocessing.set_start_method("spawn", force=True) once at program start, or pass multiprocessing_context=mp.get_context("spawn") to the DataLoader.

Check out the datasets documentation for more details here.