HDF5 matrices

One 2-D numeric dataset of a plain HDF5 file, as anndata’s read_hdf reads it. --key names it; without --key nothing is written, and the refusal lists the file’s 2-D numeric datasets with their shapes and dtypes, and the 1-D ones that could name their rows and columns.

adata convert matrix.hdf5 -o matrix.h5ad
adata convert matrix.hdf5 -o matrix.h5ad --key data/counts
adata convert export.h5 -o export.h5ad --from hdf5 --key exprs --transpose
adata convert matrix.hdf5 -o matrix.h5ad --key counts --obs-names-key barcodes --var-names-key genes

HDF5 file	AnnData
the dataset --key names (its rows are cells; genes with --transpose)	X, CSR (--dense keeps it dense), its dtype kept
row_names and col_names at the root, naming the dataset’s rows and columns	obs and var names
the key and the names’ datasets	uns["hdf5"]

Flag	Meaning
--key	The 2-D dataset that becomes X
--transpose	The dataset holds genes as rows and cells as columns
--obs-names-key, --var-names-key	The 1-D datasets naming the cells and the genes, in place of row_names and col_names
--dense	Keep X dense (default: CSR)
--make-unique	Suffix repeated var names, as var_names_make_unique does

A plain .h5 is claimed by 10x Genomics files, so one that is no 10x matrix needs --from hdf5; convert says so. A row_names or col_names whose length does not fit is left out with a warning, and that axis is numbered. The dataset is read a slab of cells at a time, a whole number of its chunks along the cell axis (columns, with --transpose), as a loom’s matrix is, so each chunk is decompressed once and memory is set by the slab.