# Licensed under a 3-clause BSD style license - see LICENSE.rst
"""
This module contains helper functions and classes for handling metadata.
"""
from functools import wraps
import warnings
from collections import OrderedDict
from collections.abc import Mapping
from copy import deepcopy
import numpy as np
from astropy.utils.exceptions import AstropyWarning
from astropy.utils.misc import dtype_bytes_or_chars
__all__ = ['MergeConflictError', 'MergeConflictWarning', 'MERGE_STRATEGIES',
'common_dtype', 'MergePlus', 'MergeNpConcatenate', 'MergeStrategy',
'MergeStrategyMeta', 'enable_merge_strategies', 'merge', 'MetaData',
'MetaAttribute']
class MergeConflictError(TypeError):
pass
class MergeConflictWarning(AstropyWarning):
pass
MERGE_STRATEGIES = []
def common_dtype(arrs):
"""
Use numpy to find the common dtype for a list of ndarrays.
Only allow arrays within the following fundamental numpy data types:
``np.bool_``, ``np.object_``, ``np.number``, ``np.character``, ``np.void``
Parameters
----------
arrs : list of ndarray
Arrays for which to find the common dtype
Returns
-------
dtype_str : str
String representation of dytpe (dtype ``str`` attribute)
"""
def dtype(arr):
return getattr(arr, 'dtype', np.dtype('O'))
np_types = (np.bool_, np.object_, np.number, np.character, np.void)
uniq_types = set(tuple(issubclass(dtype(arr).type, np_type) for np_type in np_types)
for arr in arrs)
if len(uniq_types) > 1:
# Embed into the exception the actual list of incompatible types.
incompat_types = [dtype(arr).name for arr in arrs]
tme = MergeConflictError(f'Arrays have incompatible types {incompat_types}')
tme._incompat_types = incompat_types
raise tme
arrs = [np.empty(1, dtype=dtype(arr)) for arr in arrs]
# For string-type arrays need to explicitly fill in non-zero
# values or the final arr_common = .. step is unpredictable.
for i, arr in enumerate(arrs):
if arr.dtype.kind in ('S', 'U'):
arrs[i] = [('0' if arr.dtype.kind == 'U' else b'0') *
dtype_bytes_or_chars(arr.dtype)]
arr_common = np.array([arr[0] for arr in arrs])
return arr_common.dtype.str
class MergeStrategyMeta(type):
"""
Metaclass that registers MergeStrategy subclasses into the
MERGE_STRATEGIES registry.
"""
def __new__(mcls, name, bases, members):
cls = super().__new__(mcls, name, bases, members)
# Wrap ``merge`` classmethod to catch any exception and re-raise as
# MergeConflictError.
if 'merge' in members and isinstance(members['merge'], classmethod):
orig_merge = members['merge'].__func__
@wraps(orig_merge)
def merge(cls, left, right):
try:
return orig_merge(cls, left, right)
except Exception as err:
raise MergeConflictError(err)
cls.merge = classmethod(merge)
# Register merging class (except for base MergeStrategy class)
if 'types' in members:
types = members['types']
if isinstance(types, tuple):
types = [types]
for left, right in reversed(types):
MERGE_STRATEGIES.insert(0, (left, right, cls))
return cls
class MergeStrategy(metaclass=MergeStrategyMeta):
"""
Base class for defining a strategy for merging metadata from two
sources, left and right, into a single output.
The primary functionality for the class is the ``merge(cls, left, right)``
class method. This takes ``left`` and ``right`` side arguments and
returns a single merged output.
The first class attribute is ``types``. This is defined as a list of
(left_types, right_types) tuples that indicate for which input types the
merge strategy applies. In determining whether to apply this merge
strategy to a pair of (left, right) objects, a test is done:
``isinstance(left, left_types) and isinstance(right, right_types)``. For
example::
types = [(np.ndarray, np.ndarray), # Two ndarrays
(np.ndarray, (list, tuple)), # ndarray and (list or tuple)
((list, tuple), np.ndarray)] # (list or tuple) and ndarray
As a convenience, ``types`` can be defined as a single two-tuple instead of
a list of two-tuples, e.g. ``types = (np.ndarray, np.ndarray)``.
The other class attribute is ``enabled``, which defaults to ``False`` in
the base class. By defining a subclass of ``MergeStrategy`` the new merge
strategy is automatically registered to be available for use in
merging. However, by default the new merge strategy is *not enabled*. This
prevents inadvertently changing the behavior of unrelated code that is
performing metadata merge operations.
In most cases (particularly in library code that others might use) it is
recommended to leave custom strategies disabled and use the
`~astropy.utils.metadata.enable_merge_strategies` context manager to locally
enable the desired strategies. However, if one is confident that the
new strategy will not produce unexpected behavior, then one can globally
enable it by setting the ``enabled`` class attribute to ``True``.
Examples
--------
Here we define a custom merge strategy that takes an int or float on
the left and right sides and returns a list with the two values.
>>> from astropy.utils.metadata import MergeStrategy
>>> class MergeNumbersAsList(MergeStrategy):
... types = ((int, float), (int, float)) # (left_types, right_types)
...
... @classmethod
... def merge(cls, left, right):
... return [left, right]
"""
# Set ``enabled = True`` to globally enable applying this merge strategy.
# This is not generally recommended.
enabled = False
# types = [(left_types, right_types), ...]
class MergePlus(MergeStrategy):
"""
Merge ``left`` and ``right`` objects using the plus operator. This
merge strategy is globally enabled by default.
"""
types = [(list, list), (tuple, tuple)]
enabled = True
@classmethod
def merge(cls, left, right):
return left + right
class MergeNpConcatenate(MergeStrategy):
"""
Merge ``left`` and ``right`` objects using np.concatenate. This
merge strategy is globally enabled by default.
This will upcast a list or tuple to np.ndarray and the output is
always ndarray.
"""
types = [(np.ndarray, np.ndarray),
(np.ndarray, (list, tuple)),
((list, tuple), np.ndarray)]
enabled = True
@classmethod
def merge(cls, left, right):
left, right = np.asanyarray(left), np.asanyarray(right)
common_dtype([left, right]) # Ensure left and right have compatible dtype
return np.concatenate([left, right])
def _both_isinstance(left, right, cls):
return isinstance(left, cls) and isinstance(right, cls)
def _not_equal(left, right):
try:
return bool(left != right)
except Exception:
return True
class _EnableMergeStrategies:
def __init__(self, *merge_strategies):
self.merge_strategies = merge_strategies
self.orig_enabled = {}
for left_type, right_type, merge_strategy in MERGE_STRATEGIES:
if issubclass(merge_strategy, merge_strategies):
self.orig_enabled[merge_strategy] = merge_strategy.enabled
merge_strategy.enabled = True
def __enter__(self):
pass
def __exit__(self, type, value, tb):
for merge_strategy, enabled in self.orig_enabled.items():
merge_strategy.enabled = enabled
def enable_merge_strategies(*merge_strategies):
"""
Context manager to temporarily enable one or more custom metadata merge
strategies.
Examples
--------
Here we define a custom merge strategy that takes an int or float on
the left and right sides and returns a list with the two values.
>>> from astropy.utils.metadata import MergeStrategy
>>> class MergeNumbersAsList(MergeStrategy):
... types = ((int, float), # left side types
... (int, float)) # right side types
... @classmethod
... def merge(cls, left, right):
... return [left, right]
By defining this class the merge strategy is automatically registered to be
available for use in merging. However, by default new merge strategies are
*not enabled*. This prevents inadvertently changing the behavior of
unrelated code that is performing metadata merge operations.
In order to use the new merge strategy, use this context manager as in the
following example::
>>> from astropy.table import Table, vstack
>>> from astropy.utils.metadata import enable_merge_strategies
>>> t1 = Table([[1]], names=['a'])
>>> t2 = Table([[2]], names=['a'])
>>> t1.meta = {'m': 1}
>>> t2.meta = {'m': 2}
>>> with enable_merge_strategies(MergeNumbersAsList):
... t12 = vstack([t1, t2])
>>> t12.meta['m']
[1, 2]
One can supply further merge strategies as additional arguments to the
context manager.
As a convenience, the enabling operation is actually done by checking
whether the registered strategies are subclasses of the context manager
arguments. This means one can define a related set of merge strategies and
then enable them all at once by enabling the base class. As a trivial
example, *all* registered merge strategies can be enabled with::
>>> with enable_merge_strategies(MergeStrategy):
... t12 = vstack([t1, t2])
Parameters
----------
*merge_strategies : `~astropy.utils.metadata.MergeStrategy`
Merge strategies that will be enabled.
"""
return _EnableMergeStrategies(*merge_strategies)
def _warn_str_func(key, left, right):
out = ('Cannot merge meta key {0!r} types {1!r}'
' and {2!r}, choosing {0}={3!r}'
.format(key, type(left), type(right), right))
return out
def _error_str_func(key, left, right):
out = f'Cannot merge meta key {key!r} types {type(left)!r} and {type(right)!r}'
return out
def merge(left, right, merge_func=None, metadata_conflicts='warn',
warn_str_func=_warn_str_func,
error_str_func=_error_str_func):
"""
Merge the ``left`` and ``right`` metadata objects.
This is a simplistic and limited implementation at this point.
"""
if not _both_isinstance(left, right, dict):
raise MergeConflictError('Can only merge two dict-based objects')
out = deepcopy(left)
for key, val in right.items():
# If no conflict then insert val into out dict and continue
if key not in out:
out[key] = deepcopy(val)
continue
# There is a conflict that must be resolved
if _both_isinstance(left[key], right[key], dict):
out[key] = merge(left[key], right[key], merge_func,
metadata_conflicts=metadata_conflicts)
else:
try:
if merge_func is None:
for left_type, right_type, merge_cls in MERGE_STRATEGIES:
if not merge_cls.enabled:
continue
if (isinstance(left[key], left_type) and
isinstance(right[key], right_type)):
out[key] = merge_cls.merge(left[key], right[key])
break
else:
raise MergeConflictError
else:
out[key] = merge_func(left[key], right[key])
except MergeConflictError:
# Pick the metadata item that is not None, or they are both not
# None, then if they are equal, there is no conflict, and if
# they are different, there is a conflict and we pick the one
# on the right (or raise an error).
if left[key] is None:
# This may not seem necessary since out[key] gets set to
# right[key], but not all objects support != which is
# needed for one of the if clauses.
out[key] = right[key]
elif right[key] is None:
out[key] = left[key]
elif _not_equal(left[key], right[key]):
if metadata_conflicts == 'warn':
warnings.warn(warn_str_func(key, left[key], right[key]),
MergeConflictWarning)
elif metadata_conflicts == 'error':
raise MergeConflictError(error_str_func(key, left[key], right[key]))
elif metadata_conflicts != 'silent':
raise ValueError('metadata_conflicts argument must be one '
'of "silent", "warn", or "error"')
out[key] = right[key]
else:
out[key] = right[key]
return out
class MetaData:
"""
A descriptor for classes that have a ``meta`` property.
This can be set to any valid `~collections.abc.Mapping`.
Parameters
----------
doc : `str`, optional
Documentation for the attribute of the class.
Default is ``""``.
.. versionadded:: 1.2
copy : `bool`, optional
If ``True`` the the value is deepcopied before setting, otherwise it
is saved as reference.
Default is ``True``.
.. versionadded:: 1.2
"""
def __init__(self, doc="", copy=True):
self.__doc__ = doc
self.copy = copy
def __get__(self, instance, owner):
if instance is None:
return self
if not hasattr(instance, '_meta'):
instance._meta = OrderedDict()
return instance._meta
def __set__(self, instance, value):
if value is None:
instance._meta = OrderedDict()
else:
if isinstance(value, Mapping):
if self.copy:
instance._meta = deepcopy(value)
else:
instance._meta = value
else:
raise TypeError("meta attribute must be dict-like")
class MetaAttribute:
"""
Descriptor to define custom attribute which gets stored in the object
``meta`` dict and can have a defined default.
This descriptor is intended to provide a convenient way to add attributes
to a subclass of a complex class such as ``Table`` or ``NDData``.
This requires that the object has an attribute ``meta`` which is a
dict-like object. The value of the MetaAttribute will be stored in a
new dict meta['__attributes__'] that is created when required.
Classes that define MetaAttributes are encouraged to support initializing
the attributes via the class ``__init__``. For example::
for attr in list(kwargs):
descr = getattr(self.__class__, attr, None)
if isinstance(descr, MetaAttribute):
setattr(self, attr, kwargs.pop(attr))
The name of a ``MetaAttribute`` cannot be the same as any of the following:
- Keyword argument in the owner class ``__init__``
- Method or attribute of the "parent class", where the parent class is
taken to be ``owner.__mro__[1]``.
:param default: default value
"""
def __init__(self, default=None):
self.default = default
def __get__(self, instance, owner):
# When called without an instance, return self to allow access
# to descriptor attributes.
if instance is None:
return self
# If default is None and value has not been set already then return None
# without doing touching meta['__attributes__'] at all. This helps e.g.
# with the Table._hidden_columns attribute so it doesn't auto-create
# meta['__attributes__'] always.
if (self.default is None
and self.name not in instance.meta.get('__attributes__', {})):
return None
# Get the __attributes__ dict and create if not there already.
attributes = instance.meta.setdefault('__attributes__', {})
try:
value = attributes[self.name]
except KeyError:
if self.default is not None:
attributes[self.name] = deepcopy(self.default)
# Return either specified default or None
value = attributes.get(self.name)
return value
def __set__(self, instance, value):
# Get the __attributes__ dict and create if not there already.
attributes = instance.meta.setdefault('__attributes__', {})
attributes[self.name] = value
def __delete__(self, instance):
# Remove this attribute from meta['__attributes__'] if it exists.
if '__attributes__' in instance.meta:
attrs = instance.meta['__attributes__']
if self.name in attrs:
del attrs[self.name]
# If this was the last attribute then remove the meta key as well
if not attrs:
del instance.meta['__attributes__']
def __set_name__(self, owner, name):
import inspect
params = [param.name for param in inspect.signature(owner).parameters.values()
if param.kind not in (inspect.Parameter.VAR_KEYWORD,
inspect.Parameter.VAR_POSITIONAL)]
# Reject names from existing params or best guess at parent class
if name in params or hasattr(owner.__mro__[1], name):
raise ValueError(f'{name} not allowed as {self.__class__.__name__}')
self.name = name
def __repr__(self):
return f'<{self.__class__.__name__} name={self.name} default={self.default}>'