Files
cellxgene/server/converters/cxgtool.py
2020-04-10 12:58:30 -07:00

475 lines
17 KiB
Python

"""
This program converts an [AnnData H5AD](https://anndata.readthedocs.io/en/stable/)
into a cellxgene TileDB structure, aka a 'CXG'.
The organization of the TileDB structure is:
the.cxg TileDB Group
|-- obs TileDB array containing cell (row) attributes, one attribute per
| dataframe columm, shape (n_obs,)
|-- var TileDB array containing gene (column) attributes, with one attribute per
| dataframe column, shape (n_var,)
|-- X Main count matrix as a 2D TileDB array, single unnanmed numeric attribute
|-- emb TileDB group, storing optional embeddings (group may be empty)
| |-- <name1> TileDB Array, single anon attribute, ND numeric array, shape (n_obs, N)
|-- cxg_group_metadata Empty array used only to stash metadata about the overall object.
...
All arrays are defined to have a uint32 domain, zero based. All X counds and embedding
coordinates are coerced to float32, which is ample precision for visualization purposes.
Dataframe (metadata) types are generally preserved, or where that is not possible,
converted to somemthing with equal representative value in the cellxgene application
(eg, categorical types are converted to string, bools to uint8, etc).
The following objects are also decorated with auxilliary metadata using TileDB
array metadata:
* cxg_group_metadata: minimally, will contain a 'cxg_version' field, which
is a semver string identifing the version number of the CXG layout.
It may also contain 'cxg_parameters', a JSON-encoded parameter list
describing CXG-wide dataset parameters.
* obs, var: both contain an optional 'cxg_schema' field that is a json string,
containing per-column (attribute) schema hinting. This is used where the TileDB
native typing information is insufficient to reconstruct useful information
such as categorical typing from Pandas DataFrames, and to communicate which column
is the preferred human-readable index for obs & var.
This file also embodies a number of empirically derived tiledb schema parameters,
including the global data layout, spatial tile size, and the like. The CXG is
self-describing in these areas, and the actual values (eg, tile size) are empirically
derived from benchmarking. They may change in the future.
---
TODO/ISSUES:
* add sub-command structure to argparse, for future sub-commands
* Possible future work: accept Loom files
"""
import re
import anndata
import tiledb
import argparse
import numpy as np
from os.path import splitext, basename
import json
# the CXG container version number. Must be a semver string.
CXG_VERSION = "0.1"
def log(level, *args):
global log_level
if log_level and level <= log_level:
print(*args)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("h5ad", nargs="?", help="H5AD file name")
parser.add_argument(
"--backed", action="store_true", help="loaded in file backed mode. Will be slower, but use less memory."
)
parser.add_argument(
"--obs-names", help="Name of annotation to use for observations. If not specified, will use the obs index."
)
parser.add_argument(
"--var-names", help="Name of annotation to use for variables. If not specified, will use the var index."
)
parser.add_argument("--verbose", "-v", action="count", default=0, help="verbose output")
parser.add_argument("--title", help="Human readable dataset title. If omitted, will use filename")
parser.add_argument(
"--about",
metavar="<URL>",
help="URL providing more information about the dataset (hint: must be a fully specified absolute URL).",
)
parser.add_argument("--out", "--output", "-o", help="output CXG file name")
args = parser.parse_args()
global log_level
log_level = args.verbose
adata = anndata.read_h5ad(args.h5ad, backed="r" if args.backed else None)
log(1, f"{basename(args.h5ad)} loaded...")
basefname = splitext(basename(args.h5ad))[0]
out = args.out if args.out is not None else basefname
container = out if splitext(out)[1] == ".cxg" else out + ".cxg"
title = args.title if args.title is not None else basefname
write_cxg(adata, container, title, var_names=args.var_names, obs_names=args.obs_names, about=args.about)
log(1, "done")
def write_cxg(adata, container, title, var_names=None, obs_names=None, about=None):
if not adata.var.index.is_unique:
raise ValueError("Variable index is not unique - unable to convert.")
if not adata.obs.index.is_unique:
raise ValueError("Observation index is not unique - unable to convert.")
"""
TileDB bug TileDB-Inc/TileDB#1575 requires that we sanitize all column names
prior to saving. This can be reverted when the bug is fixed.
"""
log(0, "Warning: sanitizing all dataframe column names.")
clean_all_column_names(adata)
ctx = tiledb.Ctx(
{
"sm.num_reader_threads": 32,
"sm.num_writer_threads": 32,
"sm.consolidation.buffer_size": 1 * 1024 * 1024 * 1024,
}
)
tiledb.group_create(container, ctx=ctx)
log(1, f"\t...group created, with name {container}")
# dataset metadata
save_metadata(container, {"title": title, "about": about})
log(1, "\t...dataset metadata saved")
# var/gene dataframe
save_dataframe(container, "var", adata.var, var_names, ctx=ctx)
log(1, "\t...var dataframe created")
# obs/cell dataframe
save_dataframe(container, "obs", adata.obs, obs_names, ctx=ctx)
log(1, "\t...obs dataframe created")
# embeddings
e_container = f"{container}/emb"
tiledb.group_create(e_container, ctx=ctx)
save_embeddings(e_container, adata, ctx)
log(1, "\t...embeddings created")
# X matrix
save_X(container, adata, ctx)
log(1, "\t...X created")
"""
TODO: the code used to handle type inferencing should not be duplicated between
this tool and the server/common/utils code. When this tool is merged into
the cellxgene CLI, consolidate.
"""
def dtype_to_schema(dtype):
if dtype == np.float32:
return (np.float32, {})
elif dtype == np.int32:
return (np.int32, {})
elif dtype == np.bool_:
return (np.uint8, {type: "boolean"})
elif dtype == np.str:
return (np.unicode, {"type": "string"})
elif dtype == "category":
typ, hint = cxg_type(dtype.categories)
return (typ, {"type": "categorical", "categories": dtype.categories.tolist()})
else:
raise TypeError(f"Annotations of type {dtype} are unsupported.")
def _can_cast_to_float32(array):
if array.dtype.kind == "f":
# force downcast for all floats
return True
return False
def _can_cast_to_int32(array):
if array.dtype.kind in ["i", "u"]:
if np.can_cast(array.dtype, np.int32):
return True
ii32 = np.iinfo(np.int32)
if array.min() >= ii32.min and array.max() <= ii32.max:
return True
return False
def cxg_type(array):
try:
return dtype_to_schema(array.dtype)
except TypeError:
dtype = array.dtype
data_kind = dtype.kind
if _can_cast_to_float32(array):
return (np.float32, {})
elif _can_cast_to_int32(array):
return (np.int32, {})
elif data_kind == "O" and dtype == "object":
return (np.unicode, {"type": "string"})
else:
raise TypeError(f"Annotations of type {dtype} are unsupported.")
def cxg_dtype(array):
return cxg_type(array)[0]
def create_dataframe(name, df, ctx):
"""
Current access patterns are oriented toward reading very large slices of
the dataframe, one attribute at a time. Attribute data also tends to be
(often) repetitive (bools, categories, strings).
Given this, we use:
* a large tile size (1000)
* very aggressive compression levels
"""
filter = tiledb.FilterList(
[
# attempt aggressive compression as many of these dataframes are very repetitive
# strings, bools and other non-float data.
tiledb.ZstdFilter(level=22),
]
)
attrs = [tiledb.Attr(name=col, dtype=cxg_dtype(df[col]), filters=filter) for col in df]
domain = tiledb.Domain(tiledb.Dim(domain=(0, df.shape[0] - 1), tile=min(df.shape[0], 1000), dtype=np.uint32))
schema = tiledb.ArraySchema(
domain=domain, sparse=False, attrs=attrs, cell_order="row-major", tile_order="row-major"
)
tiledb.DenseArray.create(name, schema)
def create_unique_column_name(df_cols, col_name_prefix):
"""
given the columns of a dataframe, and a name prefix, return a column name which
does not exist in the dataframe, AND which is prefixed by `prefix`
The approach is to append a numeric suffix, starting at zero and increasing by
one, until an unused name is found (eg, prefix_0, prefix_1, ...).
"""
suffix = 0
while f"{col_name_prefix}{suffix}" in df_cols:
suffix += 1
return f"{col_name_prefix}{suffix}"
def alias_index_col(df, df_name, index_col_name):
"""
We rely in the existance of a unique, human-readable index for
any dataframe (eg, var is typically gene name, obs the cell name).
The user can specify these via the --obs-names and --var-names config.
If they are not specified, use the existing index to create them, giving
the resulting column a unique name (eg, "name").
In both cases, enforce that the result is unique, and communicate the
index column name via the 'index' field in the schema hints.
"""
if index_col_name is None:
if not df.index.is_unique:
raise KeyError(
f"Values in {df_name}.index must be unique. "
"Please prepare data to contain unique index values, or specify an "
"alternative with --{ax_name}-name."
)
index_col_name = create_unique_column_name(df.columns, "name_")
# turn the index into a normal column
df.rename_axis(index_col_name, inplace=True)
df.reset_index(inplace=True)
elif index_col_name in df.columns:
# User has specified alternative column for unique names, and it exists
if not df[index_col_name].is_unique:
raise KeyError(
f"Values in {df_name}.{index_col_name} must be unique. Please prepare data to contain unique values."
)
else:
raise KeyError(f"Annotation {index_col_name}, specified in --{df_name}-name, does not exist.")
return (df, index_col_name)
def save_dataframe(container, name, df, index_col_name, ctx):
A_name = f"{container}/{name}"
(df, index_col_name) = alias_index_col(df, name, index_col_name)
create_dataframe(A_name, df, ctx=ctx)
with tiledb.DenseArray(A_name, mode="w", ctx=ctx) as A:
value = {}
schema_hints = {}
for k, v in df.items():
dtype, hints = cxg_type(v)
value[k] = v.to_numpy(dtype=dtype)
if hints:
schema_hints.update({k: hints})
schema_hints.update({"index": index_col_name})
A[:] = value
A.meta["cxg_schema"] = json.dumps(schema_hints)
tiledb.consolidate(A_name, ctx=ctx)
def create_emb(e_name, emb):
"""
Embeddings are typically accessed with very large slices (or all of the embedding),
and do not benefit from overly aggressive compression due to their format. Given
this, we use:
* large tile size (1000)
* default compression level
"""
filters = tiledb.FilterList([tiledb.ZstdFilter()])
attrs = [tiledb.Attr(dtype=emb.dtype, filters=filters)]
dims = []
for d in range(emb.ndim):
shape = emb.shape
dims.append(tiledb.Dim("", domain=(0, shape[d] - 1), tile=min(shape[d], 1000), dtype=np.uint32))
domain = tiledb.Domain(*dims)
schema = tiledb.ArraySchema(
domain=domain, sparse=False, attrs=attrs, capacity=1_000_000, cell_order="row-major", tile_order="row-major"
)
tiledb.DenseArray.create(e_name, schema)
def is_valid_embedding(adata, name, arr):
""" return True if this layout data is a valid array for front-end presentation:
* ndarray, with shape (n_obs, >= 2), dtype float/int/uint
* contains only finite values
* follows ScanPy embedding naming conventions
"""
is_valid = type(name) == str and name.startswith("X_") and len(name) > 2
is_valid = is_valid and type(arr) == np.ndarray and arr.dtype.kind in "fiu"
is_valid = is_valid and arr.shape[0] == adata.n_obs and arr.shape[1] >= 2
is_valid = is_valid and np.all(np.isfinite(arr))
return is_valid
def save_embeddings(container, adata, ctx):
for (name, value) in adata.obsm.items():
if is_valid_embedding(adata, name, value):
e_name = f"{container}/{name[2:]}"
create_emb(e_name, value)
with tiledb.DenseArray(e_name, mode="w", ctx=ctx) as A:
A[:] = value
tiledb.consolidate(e_name, ctx=ctx)
log(1, f"\t\t...{name} embedding created")
def create_X(X_name, shape):
"""
Dense, always. Future task: explore if sparse encoding is worth the trouble
below a sparsity threshold.
The X matrix is access in both row and column oriented patterns, depending on the
particular operation. Because of the data type, default compression works best.
The tile size (50, 100) and global layout (row/col) was choosen empirically, by benchmarking
the current cellxgene backend.
"""
filters = tiledb.FilterList([tiledb.ZstdFilter()])
attrs = [tiledb.Attr(dtype=np.float32, filters=filters)]
domain = tiledb.Domain(
tiledb.Dim(name="obs", domain=(0, shape[0] - 1), tile=min(shape[0], 50), dtype=np.uint32),
tiledb.Dim(name="var", domain=(0, shape[1] - 1), tile=min(shape[1], 100), dtype=np.uint32),
)
schema = tiledb.ArraySchema(
domain=domain, sparse=False, attrs=attrs, cell_order="row-major", tile_order="col-major"
)
tiledb.DenseArray.create(X_name, schema)
def save_X(container, adata, ctx):
# Save X count matrix
X_name = f"{container}/X"
shape = adata.X.shape
create_X(X_name, shape)
stride = min(int(np.power(10, np.around(np.log10(1e9 / shape[1])))), 10_000)
with tiledb.DenseArray(X_name, mode="w", ctx=ctx) as X:
for row in range(0, shape[0], stride):
lim = min(row + stride, shape[0])
a = adata.X[row:lim, :]
if type(a) is not np.ndarray:
a = a.toarray()
X[row:lim, :] = a
log(2, "\t...rows", row, "to", lim)
tiledb.consolidate(X_name, ctx=ctx)
tiledb.consolidate(X_name, ctx=ctx)
def save_metadata(container, metadata):
"""
Save all dataset-wide metadata. This includes:
* CXG version
* dataset metadata, such as title and about link.
Longer term, tiledb will have support for metadata on groups. Until
such feature exists, create an empty array and annotate that array.
https://github.com/TileDB-Inc/TileDB-Py/issues/254
"""
a_name = f"{container}/cxg_group_metadata"
with tiledb.from_numpy(a_name, np.zeros((1,))) as A:
pass
with tiledb.DenseArray(a_name, mode="w") as A:
A.meta["cxg_version"] = CXG_VERSION
A.meta["cxg_properties"] = json.dumps(metadata)
def sanitize_keys(keys):
"""
We need names to be safe to use as attribute names in tiledb. See:
TileDB-Inc/TileDB#1575
TileDB-Inc/TileDB-Py#294
This can be entirely removed once they add proper escaping.
Args: list of keys
Returns: dict of {old_key: new_key, ...}
Returned new keys will be both safe and unique.
Masking out [~/.] and anything outside the ASCII range.
"""
p = re.compile(r"[^ -\.0-\[\]-\}]")
clean_keys = {k: p.sub("_", k) for k in keys}
used_keys = set()
clean_unique_keys = {}
for k, v in clean_keys.items():
if v not in used_keys:
used_keys.add(v)
clean_unique_keys[k] = v
continue
# else, needs deduping.
counter = 1
while True:
candidate_name = v + "-" + str(counter)
if candidate_name not in used_keys:
used_keys.add(candidate_name)
clean_unique_keys[k] = candidate_name
break
counter += 1
for k, v, in clean_unique_keys.items():
if k != v:
log(1, f"Renaming {k} to {v}")
return clean_unique_keys
def sanitize_df(df):
df.rename(columns=sanitize_keys(df.keys().tolist()), inplace=True)
def sanitize_mapping(mapping):
clean_keys = sanitize_keys([k for k in mapping.keys()])
for old_key, new_key in clean_keys.items():
if old_key != new_key:
mapping[new_key] = mapping[old_key]
del mapping[old_key]
def clean_all_column_names(adata):
sanitize_df(adata.obs)
sanitize_df(adata.var)
sanitize_mapping(adata.obsm)
if __name__ == "__main__":
main()