Source code for anndata._io.h5ad

import re
import as cabc
from functools import _find_impl, partial
from warnings import warn
from pathlib import Path
from types import MappingProxyType
from typing import Callable, Type, TypeVar, Union
from typing import Collection, Sequence, Mapping

    from importlib.metadata import version
except ImportError:
    from importlib_metadata import version

import h5py
import numpy as np
import pandas as pd
from pandas.api.types import is_categorical_dtype
from scipy import sparse

from .._core.sparse_dataset import SparseDataset
from .._core.file_backing import AnnDataFileManager
from .._core.anndata import AnnData
from .._core.raw import Raw
from ..compat import (
from .utils import (

# For allowing h5py v3
H5PY_V3 = version("h5py") >= "3"

H5Group = Union[h5py.Group, h5py.File]
H5Dataset = Union[h5py.Dataset]
T = TypeVar("T")

def _to_hdf5_vlen_strings(value: np.ndarray) -> np.ndarray:
    """This corrects compound dtypes to work with hdf5 files."""
    new_dtype = []
    for dt_name, (dt_type, _) in value.dtype.fields.items():
        if dt_type.kind in ("U", "O"):
            new_dtype.append((dt_name, h5py.special_dtype(vlen=str)))
            new_dtype.append((dt_name, dt_type))
    return value.astype(new_dtype)

def write_h5ad(
    filepath: Union[Path, str],
    adata: AnnData,
    force_dense: bool = None,
    as_dense: Sequence[str] = (),
    dataset_kwargs: Mapping = MappingProxyType({}),
) -> None:
    if force_dense is not None:
            "The `force_dense` argument is deprecated. Use `as_dense` instead.",
    if force_dense is True:
        if adata.raw is not None:
            as_dense = ("X", "raw/X")
            as_dense = ("X",)
    if isinstance(as_dense, str):
        as_dense = [as_dense]
    if "raw.X" in as_dense:
        as_dense = list(as_dense)
        as_dense[as_dense.index("raw.X")] = "raw/X"
    if any(val not in {"X", "raw/X"} for val in as_dense):
        raise NotImplementedError(
            "Currently, only `X` and `raw/X` are supported values in `as_dense`"
    if "raw/X" in as_dense and adata.raw is None:
        raise ValueError("Cannot specify writing `raw/X` to dense if it doesn’t exist.")

    if adata.raw is not None:
    dataset_kwargs = {**dataset_kwargs, **kwargs}
    filepath = Path(filepath)
    mode = "a" if adata.isbacked else "w"
    if adata.isbacked:  # close so that we can reopen below
    with h5py.File(filepath, mode) as f:
        if "X" in as_dense and isinstance(adata.X, (sparse.spmatrix, SparseDataset)):
            write_sparse_as_dense(f, "X", adata.X, dataset_kwargs=dataset_kwargs)
        elif not (adata.isbacked and Path(adata.filename) == Path(filepath)):
            # If adata.isbacked, X should already be up to date
            write_attribute(f, "X", adata.X, dataset_kwargs=dataset_kwargs)
        if "raw/X" in as_dense and isinstance(
            adata.raw.X, (sparse.spmatrix, SparseDataset)
                f, "raw/X", adata.raw.X, dataset_kwargs=dataset_kwargs
            write_attribute(f, "raw/var", adata.raw.var, dataset_kwargs=dataset_kwargs)
                f, "raw/varm", adata.raw.varm, dataset_kwargs=dataset_kwargs
            write_attribute(f, "raw", adata.raw, dataset_kwargs=dataset_kwargs)
        write_attribute(f, "obs", adata.obs, dataset_kwargs=dataset_kwargs)
        write_attribute(f, "var", adata.var, dataset_kwargs=dataset_kwargs)
        write_attribute(f, "obsm", adata.obsm, dataset_kwargs=dataset_kwargs)
        write_attribute(f, "varm", adata.varm, dataset_kwargs=dataset_kwargs)
        write_attribute(f, "obsp", adata.obsp, dataset_kwargs=dataset_kwargs)
        write_attribute(f, "varp", adata.varp, dataset_kwargs=dataset_kwargs)
        write_attribute(f, "layers", adata.layers, dataset_kwargs=dataset_kwargs)
        write_attribute(f, "uns", adata.uns, dataset_kwargs=dataset_kwargs)
    if adata.isbacked:, "r+")

def _write_method(cls: Type[T]) -> Callable[[H5Group, str, T], None]:
    return _find_impl(cls, H5AD_WRITE_REGISTRY)

def write_attribute_h5ad(f: H5Group, key: str, value, *args, **kwargs):
    if key in f:
        del f[key]
    _write_method(type(value))(f, key, value, *args, **kwargs)

def write_raw(f, key, value, dataset_kwargs=MappingProxyType({})):
    group = f.create_group(key)
    group.attrs["encoding-type"] = "raw"
    group.attrs["encoding-version"] = EncodingVersions.raw.value
    group.attrs["shape"] = value.shape
    write_attribute(f, "raw/X", value.X, dataset_kwargs=dataset_kwargs)
    write_attribute(f, "raw/var", value.var, dataset_kwargs=dataset_kwargs)
    write_attribute(f, "raw/varm", value.varm, dataset_kwargs=dataset_kwargs)

def write_not_implemented(f, key, value, dataset_kwargs=MappingProxyType({})):
    # If it’s not an array, try and make it an array. If that fails, pickle it.
    # Maybe rethink that, maybe this should just pickle,
    # and have explicit implementations for everything else
    raise NotImplementedError(
        f"Failed to write value for {key}, "
        f"since a writer for type {type(value)} has not been implemented yet."

def write_basic(f, key, value, dataset_kwargs=MappingProxyType({})):
    f.create_dataset(key, data=value, **dataset_kwargs)

def write_list(f, key, value, dataset_kwargs=MappingProxyType({})):
    write_array(f, key, np.array(value), dataset_kwargs=dataset_kwargs)

def write_none(f, key, value, dataset_kwargs=MappingProxyType({})):

def write_scalar(f, key, value, dataset_kwargs=MappingProxyType({})):
    # Can’t compress scalars, error is thrown
    if "compression" in dataset_kwargs:
        dataset_kwargs = dict(dataset_kwargs)
    write_array(f, key, np.array(value), dataset_kwargs=dataset_kwargs)

def write_array(f, key, value, dataset_kwargs=MappingProxyType({})):
    # Convert unicode to fixed length strings
    if value.dtype.kind in {"U", "O"}:
        value = value.astype(h5py.special_dtype(vlen=str))
    elif value.dtype.names is not None:
        value = _to_hdf5_vlen_strings(value)
    f.create_dataset(key, data=value, **dataset_kwargs)

def write_sparse_compressed(
    f, key, value, fmt: Literal["csr", "csc"], dataset_kwargs=MappingProxyType({})
    g = f.create_group(key)
    g.attrs["encoding-type"] = f"{fmt}_matrix"
    g.attrs["encoding-version"] = EncodingVersions[f"{fmt}_matrix"].value
    g.attrs["shape"] = value.shape

    # Allow resizing
    if "maxshape" not in dataset_kwargs:
        dataset_kwargs = dict(maxshape=(None,), **dataset_kwargs)

    g.create_dataset("data",, **dataset_kwargs)
    g.create_dataset("indices", data=value.indices, **dataset_kwargs)
    g.create_dataset("indptr", data=value.indptr, **dataset_kwargs)

write_csr = partial(write_sparse_compressed, fmt="csr")
write_csc = partial(write_sparse_compressed, fmt="csc")

def write_sparse_dataset(f, key, value, dataset_kwargs=MappingProxyType({})):
        f, key, value.to_backed(), fmt=value.format_str, dataset_kwargs=dataset_kwargs

def write_sparse_as_dense(f, key, value, dataset_kwargs=MappingProxyType({})):
    real_key = None  # Flag for if temporary key was used
    if key in f:
        if (
            isinstance(value, (h5py.Group, h5py.Dataset, SparseDataset))
            and value.file.filename == f.filename
        ):  # Write to temporary key before overwriting
            real_key = key
            # Transform key to temporary, e.g. raw/X -> raw/_X, or X -> _X
            key = re.sub(r"(.*)(\w(?!.*/))", r"\1_\2", key.rstrip("/"))
            del f[key]  # Wipe before write
    dset = f.create_dataset(key, shape=value.shape, dtype=value.dtype, **dataset_kwargs)
    compressed_axis = int(isinstance(value, sparse.csc_matrix))
    for idx in idx_chunks_along_axis(value.shape, compressed_axis, 1000):
        dset[idx] = value[idx].toarray()
    if real_key is not None:
        del f[real_key]
        f[real_key] = f[key]
        del f[key]

def write_dataframe(f, key, df, dataset_kwargs=MappingProxyType({})):
    # Check arguments
    for reserved in ("__categories", "_index"):
        if reserved in df.columns:
            raise ValueError(f"{reserved!r} is a reserved name for dataframe columns.")
    group = f.create_group(key)
    group.attrs["encoding-type"] = "dataframe"
    group.attrs["encoding-version"] = EncodingVersions.dataframe.value
    group.attrs["column-order"] = list(df.columns)

    if is not None:
        index_name =
        index_name = "_index"
    group.attrs["_index"] = index_name

    write_series(group, index_name, df.index, dataset_kwargs=dataset_kwargs)
    for colname, series in df.items():
        write_series(group, colname, series, dataset_kwargs=dataset_kwargs)

def write_series(group, key, series, dataset_kwargs=MappingProxyType({})):
    # group here is an h5py type, otherwise categoricals won’t write
    if series.dtype == object:  # Assuming it’s string
    elif is_categorical_dtype(series):
        # This should work for categorical Index and Series
        categorical: pd.Categorical = series.values
        categories: np.ndarray = categorical.categories.values
        codes: np.ndarray =
        category_key = f"__categories/{key}"

        write_array(group, category_key, categories, dataset_kwargs=dataset_kwargs)
        write_array(group, key, codes, dataset_kwargs=dataset_kwargs)

        group[key].attrs["categories"] = group[category_key].ref
        group[category_key].attrs["ordered"] = categorical.ordered
        group[key] = series.values

def write_mapping(f, key, value, dataset_kwargs=MappingProxyType({})):
    for sub_key, sub_value in value.items():
        write_attribute(f, f"{key}/{sub_key}", sub_value, dataset_kwargs=dataset_kwargs)

    Raw: write_raw,
    object: write_not_implemented,
    h5py.Dataset: write_basic,
    list: write_list,
    type(None): write_none,
    str: write_scalar,
    float: write_scalar,
    np.floating: write_scalar,
    bool: write_scalar,
    np.bool_: write_scalar,
    int: write_scalar,
    np.integer: write_scalar,
    np.ndarray: write_array,
    sparse.csr_matrix: write_csr,
    sparse.csc_matrix: write_csc,
    SparseDataset: write_sparse_dataset,
    pd.DataFrame: write_dataframe,
    cabc.Mapping: write_mapping,

def read_h5ad_backed(filename: Union[str, Path], mode: Literal["r", "r+"]) -> AnnData:
    d = dict(filename=filename, filemode=mode)

    f = h5py.File(filename, mode)

    attributes = ["obsm", "varm", "obsp", "varp", "uns", "layers"]
    df_attributes = ["obs", "var"]

    d.update({k: read_attribute(f[k]) for k in attributes if k in f})
    for k in df_attributes:
        if k in f:  # Backwards compat
            d[k] = read_dataframe(f[k])

    d["raw"] = _read_raw(f, attrs={"var", "varm"})

    X_dset = f.get("X", None)
    if X_dset is None:
    elif isinstance(X_dset, h5py.Group):
        d["dtype"] = X_dset["data"].dtype
    elif hasattr(X_dset, "dtype"):
        d["dtype"] = f["X"].dtype
        raise ValueError()


    return AnnData(**d)

[docs]def read_h5ad( filename: Union[str, Path], backed: Union[Literal["r", "r+"], bool, None] = None, *, as_sparse: Sequence[str] = (), as_sparse_fmt: Type[sparse.spmatrix] = sparse.csr_matrix, chunk_size: int = 6000, # TODO, probably make this 2d chunks ) -> AnnData: """\ Read `.h5ad`-formatted hdf5 file. Parameters ---------- filename File name of data file. backed If `'r'`, load :class:`~anndata.AnnData` in `backed` mode instead of fully loading it into memory (`memory` mode). If you want to modify backed attributes of the AnnData object, you need to choose `'r+'`. as_sparse If an array was saved as dense, passing its name here will read it as a sparse_matrix, by chunk of size `chunk_size`. as_sparse_fmt Sparse format class to read elements from `as_sparse` in as. chunk_size Used only when loading sparse dataset that is stored as dense. Loading iterates through chunks of the dataset of this row size until it reads the whole dataset. Higher size means higher memory consumption and higher (to a point) loading speed. """ if backed not in {None, False}: mode = backed if mode is True: mode = "r+" assert mode in {"r", "r+"} return read_h5ad_backed(filename, mode) if as_sparse_fmt not in (sparse.csr_matrix, sparse.csc_matrix): raise NotImplementedError( "Dense formats can only be read to CSR or CSC matrices at this time." ) if isinstance(as_sparse, str): as_sparse = [as_sparse] else: as_sparse = list(as_sparse) for i in range(len(as_sparse)): if as_sparse[i] in {("raw", "X"), "raw.X"}: as_sparse[i] = "raw/X" elif as_sparse[i] not in {"raw/X", "X"}: raise NotImplementedError( "Currently only `X` and `raw/X` can be read as sparse." ) rdasp = partial( read_dense_as_sparse, sparse_format=as_sparse_fmt, axis_chunk=chunk_size ) with h5py.File(filename, "r") as f: d = {} for k in f.keys(): # Backwards compat for old raw if k == "raw" or k.startswith("raw."): continue if k == "X" and "X" in as_sparse: d[k] = rdasp(f[k]) elif k == "raw": assert False, "unexpected raw format" elif k in {"obs", "var"}: d[k] = read_dataframe(f[k]) else: # Base case d[k] = read_attribute(f[k]) d["raw"] = _read_raw(f, as_sparse, rdasp) X_dset = f.get("X", None) if X_dset is None: pass elif isinstance(X_dset, h5py.Group): d["dtype"] = X_dset["data"].dtype elif hasattr(X_dset, "dtype"): d["dtype"] = f["X"].dtype else: raise ValueError() _clean_uns(d) # backwards compat return AnnData(**d)
def _read_raw( f: Union[h5py.File, AnnDataFileManager], as_sparse: Collection[str] = (), rdasp: Callable[[h5py.Dataset], sparse.spmatrix] = None, *, attrs: Collection[str] = ("X", "var", "varm"), ): if as_sparse: assert rdasp is not None, "must supply rdasp if as_sparse is supplied" raw = {} if "X" in attrs and "raw/X" in f: read_x = rdasp if "raw/X" in as_sparse else read_attribute raw["X"] = read_x(f["raw/X"]) for v in ("var", "varm"): if v in attrs and f"raw/{v}" in f: raw[v] = read_attribute(f[f"raw/{v}"]) return _read_legacy_raw(f, raw, read_dataframe, read_attribute, attrs=attrs) @report_read_key_on_error def read_dataframe_legacy(dataset) -> pd.DataFrame: """Read pre-anndata 0.7 dataframes.""" if H5PY_V3: df = pd.DataFrame( _decode_structured_array( _from_fixed_length_strings(dataset[()]), dtype=dataset.dtype ) ) else: df = pd.DataFrame(_from_fixed_length_strings(dataset[()])) df.set_index(df.columns[0], inplace=True) return df @report_read_key_on_error def read_dataframe(group) -> pd.DataFrame: if not isinstance(group, h5py.Group): return read_dataframe_legacy(group) columns = list(group.attrs["column-order"]) idx_key = group.attrs["_index"] df = pd.DataFrame( {k: read_series(group[k]) for k in columns}, index=read_series(group[idx_key]), columns=list(columns), ) if idx_key != "_index": = idx_key return df @report_read_key_on_error def read_series(dataset) -> Union[np.ndarray, pd.Categorical]: if "categories" in dataset.attrs: categories = dataset.attrs["categories"] if isinstance(categories, h5py.Reference): categories_dset = dataset.parent[dataset.attrs["categories"]] categories = read_dataset(categories_dset) ordered = bool(categories_dset.attrs.get("ordered", False)) else: # TODO: remove this code at some point post 0.7 # TODO: Add tests for this warn( f"Your file {str(!r} has invalid categorical " "encodings due to being written from a development version of " "AnnData. Rewrite the file ensure you can read it in the future.", FutureWarning, ) return pd.Categorical.from_codes( read_dataset(dataset), categories, ordered=ordered ) else: return read_dataset(dataset) # @report_read_key_on_error # def read_sparse_dataset_backed(group: h5py.Group) -> sparse.spmatrix: # return SparseDataset(group) @read_attribute.register(h5py.Group) @report_read_key_on_error def read_group(group: h5py.Group) -> Union[dict, pd.DataFrame, sparse.spmatrix]: if "h5sparse_format" in group.attrs: # Backwards compat return SparseDataset(group).to_memory() encoding_type = group.attrs.get("encoding-type") if encoding_type: EncodingVersions[encoding_type].check(, group.attrs["encoding-version"] ) if encoding_type in {None, "raw"}: pass elif encoding_type == "dataframe": return read_dataframe(group) elif encoding_type in {"csr_matrix", "csc_matrix"}: return SparseDataset(group).to_memory() else: raise ValueError(f"Unfamiliar `encoding-type`: {encoding_type}.") d = dict() for sub_key, sub_value in group.items(): d[sub_key] = read_attribute(sub_value) return d @read_attribute.register(h5py.Dataset) @report_read_key_on_error def read_dataset(dataset: h5py.Dataset): if H5PY_V3: string_dtype = h5py.check_string_dtype(dataset.dtype) if (string_dtype is not None) and (string_dtype.encoding == "utf-8"): dataset = dataset.asstr() value = dataset[()] if not hasattr(value, "dtype"): return value elif isinstance(value.dtype, str): pass elif issubclass(value.dtype.type, np.string_): value = value.astype(str) # Backwards compat, old datasets have strings as one element 1d arrays if len(value) == 1: return value[0] elif len(value.dtype.descr) > 1: # Compound dtype # For backwards compat, now strings are written as variable length dtype = value.dtype value = _from_fixed_length_strings(value) if H5PY_V3: value = _decode_structured_array(value, dtype=dtype) if value.shape == (): value = value[()] return value @report_read_key_on_error def read_dense_as_sparse( dataset: h5py.Dataset, sparse_format: sparse.spmatrix, axis_chunk: int ): if sparse_format == sparse.csr_matrix: return read_dense_as_csr(dataset, axis_chunk) elif sparse_format == sparse.csc_matrix: return read_dense_as_csc(dataset, axis_chunk) else: raise ValueError(f"Cannot read dense array as type: {sparse_format}") def read_dense_as_csr(dataset, axis_chunk=6000): sub_matrices = [] for idx in idx_chunks_along_axis(dataset.shape, 0, axis_chunk): dense_chunk = dataset[idx] sub_matrix = sparse.csr_matrix(dense_chunk) sub_matrices.append(sub_matrix) return sparse.vstack(sub_matrices, format="csr") def read_dense_as_csc(dataset, axis_chunk=6000): sub_matrices = [] for idx in idx_chunks_along_axis(dataset.shape, 1, axis_chunk): sub_matrix = sparse.csc_matrix(dataset[idx]) sub_matrices.append(sub_matrix) return sparse.hstack(sub_matrices, format="csc")