First draft of SEP-13 for a sliceable Metadata API.

Author: DanRyanIrish

SEP-0013.md

@@ -0,0 +1,136 @@
+# SEP-0013 -- A Metadata Class that Supports Axis-aligned Metadata
+
+| SEP           | num                                                  |
+|---------------|------------------------------------------------------|
+| title         | A Metadata Class that Supports Axis-aligned Metadata |
+| author(s)     | [Daniel Ryan](https://orcid.org/0000-0001-8661-3825) |
+| contact email | [email protected]                                        |
+| date-creation | 2021-11-21                                           |
+| type          | standard                                             |
+| discussion    | link to discussion if available                      |
+| status        | discussion                                           |
+
+# Introduction
+Currently in `ndcube` and `sunpy`, the metadata object attached to data classes
+like `NDCube` and `Map` can be of any `dict`-like type.
+While this level of generality allows users to use the metadata object however
+they want, it prevents the data classes interacting with the metadata object
+because there is no standardized API.
+One important use-case is supporting axis-aligned metadata, i.e. a piece of
+metadata that has a different value for different indices in the data array,
+e.g. exposure time.
+This SEP strives to define a `dict`-like metadata API that is independent of the
+type of data it describes and enables, but does not require, axis-aligned metadata.
+
+The API enables:
+- Access to the metadata values via their name key;
+- Comments associated with specific pieces of metadata;
+- Editing of the metadata values and/or comments;
+- The preservation of a copy of the original metadata values;
+- Addition and removal of metadata;
+- Association of metadata with one or more data axes;
+- Slicing of the metadata object given a slice item valid for the data array the metadata object describes.
+
+It is anticipated that this API will be implemented in `ndcube`.
+The motivating use-case storing exposure time and enabling sunraster data objects
+to normalize their intensity values.
+
+This metadata API is not designed as a standardization of names of specific metadata
+associated with different data types, e.g. images, or instruments.
+Instead it is hoped that such metadata schemes can be implemented on top of this API.
+
+# MetaABC API
+Below is the class definition of an Abstract Base Class (ABC) for a `Meta` object.
+
+```python
+class Meta(dict):
+
+    def __init__(self,
+                 header: dict-like = None,
+                 comments: dict-like = None,
+                 axes: dict-like = None,
+                 data_shape: Iterable[int] = None):
+    """
+    A sliceable object for storing metadata.
+
+    Metdata can be linked to a data array axis. This causes it to be sliced when the
+    standard Python numeric slicing API is applied to the object.
+    Specific pieces of metadata can be obtained using the dict-like str slicing API.
+    Metadata associated with an axis/axes must have the same length/shape as those axes.
+
+    Parameters
+    ----------
+    header:
+        The names and values of metadata.
+    comments:
+        Comments associated with any of the above pieces of metadata.
+        This does not have to include a key for every key in header,
+        but every key in comments must also be in header.
+    axes:
+        The axis/axes associated with the above metadata values.
+        Each axis value must be None (for no axis association), `int`
+        or an iterable of `int` if the metadata is associated with multiple axes.
+        Metadata in header without a corresponding entry here are assumed to not
+        be associated with an axis.
+        All keys in axes must also be in header.
+    data_shape:
+        The shape of the data array with which this metadata is associated.
+        Must be set if axes input is set.
+    """
+
+    @abc.abstractmethod
+    @property
+    def original_header(self) -> Any:
+        """A copy of the original header input."""
+
+    @abc.abstractmethod
+    @property
+    def comments(self) -> dict:
+        """Returns a dict of comments associated with specific pieces of metadata."""
+
+    @abc.abstractmethod
+    @property
+    def axes(self) -> dict:
+        """Returns a dict of array axes with which specific pieces of metadata are associated.
+
+        Axes are returned as tuples of ints. If the metadata is not associated with
+        array axis, None is returned. Metadata keys absent from this dict are assumed
+        to be associated with no array axes.
+        """
+
+    @abc.abstractmethod
+    @property
+    def shape(self) -> numpy.ndarray:
+        """1-D array giving the length of each array axis of the associated data array."""
+
+    @abc.abstractmethod
+    def add(self,
+            name: str,
+            value: Any,
+            comment: str = None,
+            axis: Union[int, Iterable[int]] = None,
+            overwrite: bool = False):
+    """Add a new piece of metadata to instance."""
+
+    @abc.abstractmethod
+    def remove(self,
+               name: str):
+    """Remove a new piece of metadata to instance."""
+
+    @abc.abstractmethod
+    def __getitem__(self, item: Union[int, slice, Iterable[Union[slice, int]]]) -> "MetaABC":
+        """Slice the Meta object.
+
+        The input item must be a valid slice item for an array with a shape given by
+        MetaABC.shape. If MetaABC.shape is None, then this method should error.
+        This method will alter the following things:
+          * The values of axis-aligned metadata;
+          * The axes of axis-aligned metadata if any data axes are dropped by the slicing;
+          * MetaABC.shape.
+        Data axes can be dropped by slicing if an integer is applied to an axis. This means
+        that axis indices greater than the dropped axis, will be decremented by 1.
+        Metadata, all of whose axes are dropped by the slicing, will have a new axis value of None.
+        """
+
+# Decision Rationale
+This is a great idea because...