bixverse

Streams a bixverse binary cell-major store into minibatches for neural network training. The Rust crate does the work: a feeder thread decompresses and materialises batches ahead of your training step, and the GIL is released while it does. No materialising the experiment in memory, no Python-side decode.

Install

uv pip install bixverse            # numpy only
uv pip install "bixverse[sparse]"  # adds scipy, for Batch.scipy_csr
uv pip install "bixverse[torch]"   # torch, for bixverse.nn
uv pip install "bixverse[mlx]"     # MLX, Apple Silicon only

Thirty seconds

import bixverse as bv

reader = bv.CellReader("counts.bin")
loader = bv.CellLoader(reader, batch_size=256, layer="norm", seed=42)

for batch in loader:
    x = batch.scipy_csr()  # (256, n_genes) CSR

Cell-major stores only. A gene-major file raises on construction.

Next: the quickstart, then the API reference.