Converters

The Converter interface defines a mapping between tagged objects in the ASDF tree and their corresponding Python object(s). Typically a Converter will map one YAML tag to one Python type, but the interface also supports many-to-one and many-to-many mappings. A Converter provides the software support for a tag and is responsible for both converting from parsed YAML to more complex Python objects and vice versa.

The Converter interface

Every Converter implementation must provide two required properties and two required methods:

Converter.tags - a list of tag URIs or URI patterns handled by the converter. Patterns may include the wildcard character *, which matches any sequence of characters up to a /, or **, which matches any sequence of characters. The uri_match method can be used to test URI patterns.

Converter.types - a list of Python types or fully-qualified Python type names handled by the converter. For strings, the private or public path can be used. For example, if class Foo is implemented in example_package.foo.Foo but imported as example_package.Foo for convenience either example_package.foo.Foo or example_package.Foo can be used. As most libraries do not consider moving where a class is implemented it is preferred to use the “public” location where the class is imported (in this example example_package.Foo).

The string type name is recommended over a type object for performance reasons, see Entry point performance considerations.

Converter.to_yaml_tree - a method that accepts a complex Python object and returns a simple node object (typically a dict) suitable for serialization to YAML. The node is permitted to contain nested complex objects; these will in turn be passed to other to_yaml_tree methods in other Converters.

Converter.from_yaml_tree - a method that accepts a simple node object from parsed YAML and returns the appropriate complex Python object. For a non-lazy-tree, nested nodes in the received node will have already been converted to complex objects by other calls to from_yaml_tree methods, except where reference cycles are present – see Reference cycles for information on how to handle that situation. For a lazy_tree (see asdf.open) the node will contain asdf.lazy_nodes instances which act like dicts and lists but convert child objects only when they are accessed.

Additionally, the Converter interface includes a method that must be implemented when some logic is required to select the tag to assign to a to_yaml_tree result:

Converter.select_tag - an optional method that accepts a complex Python object and a list candidate tags and returns the tag that should be used to serialize the object.

Converter.lazy - a boolean attribute indicating if this converter accepts “lazy” objects (those defined in asdf.lazy_nodes). This is mostly useful for container-like classes (where the “lazy” objects can defer conversion of contained objects until they are accessed). If a converter produces a generator lazy should be set to False as asdf will need to generate nodes further out the branch to fully resolve the object returned from the generator.

A simple example

Say we have a Python class, Rectangle, that we wish to serialize to an ASDF file. A Rectangle instance has two attributes, width and height, and a convenient method that computes its area:

# in module example_package.shapes
class Rectangle:
    def __init__(self, width, height):
        self.width = width
        self.height = height

    def get_area(self):
        return self.width * self.height

We’ll need to designate a tag URI to represent this object’s type in the ASDF tree – let’s use asdf://example.com/example-project/tags/rectangle-1.0.0. Here is a simple Converter implementation for this type and tag:

from asdf.extension import Converter


class RectangleConverter(Converter):
    tags = ["asdf://example.com/shapes/tags/rectangle-1.0.0"]
    types = ["example_package.shapes.Rectangle"]

    def to_yaml_tree(self, obj, tag, ctx):
        return {
            "width": obj.width,
            "height": obj.height,
        }

    def from_yaml_tree(self, node, tag, ctx):
        from example_package.shapes import Rectangle

        return Rectangle(node["width"], node["height"])

Note that import of the Rectangle class has been deferred to inside the from_yaml_tree method. This is a performance consideration that is discussed in Entry point performance considerations.

In order to use this Converter, we’ll need to create a simple extension around it and install that extension:

import asdf
from asdf.extension import Extension


class ShapesExtension(Extension):
    extension_uri = "asdf://example.com/shapes/extensions/shapes-1.0.0"
    converters = [RectangleConverter()]
    tags = ["asdf://example.com/shapes/tags/rectangle-1.0.0"]


asdf.get_config().add_extension(ShapesExtension())

Now we can include a Rectangle object in an AsdfFile tree and write out a file:

with asdf.AsdfFile() as af:
    af["rect"] = Rectangle(5, 4)
    af.write_to("test.asdf")

The portion of the ASDF file that represents the rectangle looks like this:

rect: !<asdf://example.com/shapes/tags/rectangle-1.0.0> {height: 4, width: 5}

Multiple tags

Now say we want to map our one Rectangle class to one of two tags, either rectangle-1.0.0 or square-1.0.0. We’ll need to add square-1.0.0 to the converter’s list of tags and implement a select_tag method:

RETANGLE_TAG = "asdf://example.com/shapes/tags/rectangle-1.0.0"
SQUARE_TAG = "asdf://example.com/shapes/tags/square-1.0.0"


class RectangleConverter(Converter):
    tags = [RECTANGLE_TAG, SQUARE_TAG]
    types = ["example_package.shapes.Rectangle"]

    def select_tag(self, obj, tags, ctx):
        if obj.width == obj.height:
            return SQUARE_TAG
        else:
            return RECTANGLE_TAG

    def to_yaml_tree(self, obj, tag, ctx):
        if tag == SQUARE_TAG:
            return {
                "side_length": obj.width,
            }
        else:
            return {
                "width": obj.width,
                "height": obj.height,
            }

    def from_yaml_tree(self, node, tag, ctx):
        from example_package.shapes import Rectangle

        if tag == SQUARE_TAG:
            return Rectangle(node["side_length"], node["side_length"])
        else:
            return Rectangle(node["width"], node["height"])

Deferring to another converter

Converters only support the exact types listed in Converter.types. When a supported type is subclassed the extension will need to be updated to support the new subclass. There are a few options for supporting subclasses.

If serialization of the subclass needs to differ from the superclass a new Converter, tag and schema should be defined.

If the subclass can be treated the same as the superclass (specifically if subclass instances can be serialized as the superclass) then the subclass can be added to the existing Converter.types. Note that adding the subclass to the supported types (without making other changes to the Converter) will result in subclass instances using the same tag as the superclass. This means that any instances created during deserialization will always be of the superclass (subclass instances will never be read from an ASDF file).

Another option (useful when modifying the existing Converter is not convenient) is to define a Converter that does not tag the subclass instance being serialized and instead defers to the existing Converter. Deferral is triggered by returning None from Converter.select_tag and implementing Converter.to_yaml_tree to convert the subclass instance into an instance of the (supported) superclass.

For example, using the example Rectangle class above, let’s say we have another class, AspectRectangle, that represents a rectangle as a height and aspect ratio. We know we never need to deserialize this class for our uses and are ok with always reading Rectangle instances after saving AspectRectangle instances. In this case we can define a Converter for AspectRectangle that converts instances to Rectangle and defers to the RectangleConverter.

class AspectRectangle(Rectangle):
    def __init__(self, height, ratio):
        self.height = height
        self.ratio = ratio

    def get_area(self):
        width = self.height * self.ratio
        return width * self.height


class AspectRectangleConverter(Converter):
    tags = []
    types = [AspectRectangle]

    def select_tag(self, obj, tags, ctx):
        return None  # defer to a different Converter

    def to_yaml_tree(self, obj, tag, ctx):
        # convert the instance of AspectRectangle (obj) to
        # a supported type (Rectangle)
        return Rectangle(obj.height * obj.ratio, obj.height)

    def from_yaml_tree(self, node, tag, ctx):
        raise NotImplementedError()

Just like a non-deferring Converter this Converter will need to be added to an Extension and registered with asdf.

Reference cycles

Special considerations must be made when deserializing a tagged object that contains a reference to itself among its descendants. Consider a fractions.Fraction subclass that maintains a reference to its multiplicative inverse:

# in the example_project.fractions module
class FractionWithInverse(fractions.Fraction):
    def __init__(self, *args, **kwargs):
        self._inverse = None

    @property
    def inverse(self):
        return self._inverse

    @inverse.setter
    def inverse(self, value):
        self._inverse = value

The inverse of the inverse of a fraction is the fraction itself, we might wish to construct the objects in the following way:

f1 = FractionWithInverse(3, 5)
f2 = FractionWithInverse(5, 3)
f1.inverse = f2
f2.inverse = f1

Which creates an “infinite loop” between the two fractions. An ordinary Converter wouldn’t be able to deserialize this, since each fraction requires that the other be deserialized first! Let’s see what happens when we define our from_yaml_tree method in a naive way:

class FractionWithInverseConverter(Converter):
    tags = ["asdf://example.com/fractions/tags/fraction-1.0.0"]
    types = ["example_project.fractions.FractionWithInverse"]

    def to_yaml_tree(self, obj, tag, ctx):
        return {
            "numerator": obj.width,
            "denominator": obj.height,
            "inverse": obj.inverse,
        }

    def from_yaml_tree(self, node, tag, ctx):
        from example_project.fractions import FractionWithInverse

        obj = FractionWithInverse(tree["numerator"], tree["denominator"])
        obj.inverse = tree["inverse"]
        return obj

After adding this Converter to an Extension and installing it, the fraction will serialize correctly:

with asdf.AsdfFile({"fraction": f1}) as af:
    af.write_to("with_inverse.asdf")

But upon deserialization, we notice a problem:

with asdf.open("with_inverse.asdf") as af:
    reconstituted_f1 = af["fraction"]

assert reconstituted_f1.inverse.inverse is asdf.treeutil.PendingValue

The presence of _PendingValue is asdf’s way of telling us that the value corresponding to the key inverse was not fully deserialized at the time that we retrieved it. We can handle this situation by making our from_yaml_tree a generator function:

def from_yaml_tree(self, node, tag, ctx):
    from example_project.fractions import FractionWithInverse

    obj = FractionWithInverse(tree["numerator"], tree["denominator"])
    yield obj
    obj.inverse = tree["inverse"]

The generator version of from_yaml_tree yields the partially constructed FractionWithInverse object before setting its inverse property. This allows asdf to proceed to constructing the inverse FractionWithInverse object, and resume the original from_yaml_tree execution only when the inverse is actually available.

With this modification we can successfully deserialize our ASDF file:

with asdf.open("with_inverse.asdf") as af:
    reconstituted_f1 = ff["fraction"]

assert reconstituted_f1.inverse.inverse is reconstituted_f1

Block storage

As described above Converters can return complex objects that will be passed to other Converters. If a Converter returns a ndarray, asdf will recognize this array and store it in an ASDF block. This is the easiest and preferred means of storing data in ASDF blocks.

For applications that require more flexibility, Converters can control block storage through use of the asdf.extension.SerializationContext provided as an argument to Converter.to_yaml_tree Converter.from_yaml_tree and Converter.select_tag.

It is helpful to first review some details of how asdf stores block. Blocks are stored sequentially within a ASDF file following the YAML tree. During reads and writes, asdf will need to know the index of the block a Converter would like to use to read or write the correct block. However, the index used for reading might not be the same index for writing if the tree was modified or the file is being written to a new location. During serialization and deserialization, asdf will associate each object with the accessed block during Converter.from_yaml_tree and Converter.to_yaml_tree.

Note

Converters using multiple blocks are slightly more complicated. See: Converters using multiple blocks

A simple example of a Converter using block storage to store the payload for BlockData object instances is as follows:

import asdf
import numpy as np
from asdf.extension import Converter, Extension

class BlockData:
    def __init__(self, payload):
        self.payload = payload


class BlockConverter(Converter):
    tags = ["asdf://somewhere.org/tags/block_data-1.0.0"]
    types = [BlockData]

    def to_yaml_tree(self, obj, tag, ctx):
        block_index = ctx.find_available_block_index(
            lambda: np.ndarray(len(obj.payload), dtype="uint8", buffer=obj.payload),
        )
        return {"block_index": block_index}

    def from_yaml_tree(self, node, tag, ctx):
        block_index = node["block_index"]
        data_callback = ctx.get_block_data_callback(block_index)
        obj = BlockData(data_callback())
        return obj

class BlockExtension(Extension):
    tags = ["asdf://somewhere.org/tags/block_data-1.0.0"]
    converters = [BlockConverter()]
    extension_uri = "asdf://somewhere.org/extensions/block_data-1.0.0"

with asdf.config_context() as cfg:
    cfg.add_extension(BlockExtension())
    ff = asdf.AsdfFile({"example": BlockData(b"abcdefg")})
    ff.write_to("block_converter_example.asdf")

block_converter_example.asdf

#ASDF 1.0.0
#ASDF_STANDARD 1.5.0
%YAML 1.1
%TAG ! tag:stsci.edu:asdf/
--- !core/asdf-1.1.0
asdf_library: !core/software-1.0.0 {author: The ASDF Developers, homepage: 'http://github.com/asdf-format/asdf',
  name: asdf, version: 3.5.0}
history:
  extensions:
  - !core/extension_metadata-1.0.0
    extension_class: asdf.extension._manifest.ManifestExtension
    extension_uri: asdf://asdf-format.org/core/extensions/core-1.5.0
    manifest_software: !core/software-1.0.0 {name: asdf_standard, version: 1.1.1}
    software: !core/software-1.0.0 {name: asdf, version: 3.5.0}
  - !core/extension_metadata-1.0.0 {extension_class: builtins.BlockExtension, extension_uri: 'asdf://somewhere.org/extensions/block_data-1.0.0'}
example: !<asdf://somewhere.org/tags/block_data-1.0.0> {block_index: 0}
...

BLOCK 0:
    allocated_size: 7
    used_size: 7
    data_size: 7
    data: b'61626364656667'

#ASDF BLOCK INDEX
%YAML 1.1
---
- 792
...

During read, Converter.from_yaml_tree will be called. Within this method the Converter can prepare to access a block by calling SerializationContext.get_block_data_callback. This will return a function that when called will return the contents of the block (to support lazy loading without keeping a reference to the SerializationContext (which is meant to be a short lived and lightweight object).

During write, Converter.to_yaml_tree will be called. The Converter can use SerializationContext.find_available_block_index to find the location of an available block for writing. The data to be written to the block can be provided as an ndarray or a callable function that will return a ndarray (note that it is possible this callable function will be called multiple times and the developer should cache results from any non-repeatable sources).

Converters using multiple blocks

As discussed above, while serializing and deserializing objects that use one block, asdf will watch which block is accessed by find_available_block_index and get_block_data_callback and associate the block with the converted object. This association allows asdf to map read and write blocks during updates of ASDF files. An object that uses multiple blocks must provide a unique key for each block it uses. These keys are generated using SerializationContext.generate_block_key and must be stored by the extension code. These keys must be resupplied to the converter when writing an object that was read from an ASDF file.

import asdf
import numpy as np
from asdf.extension import Converter, Extension

class MultiBlockData:
    def __init__(self, data):
        self.data = data
        self.keys = []


class MultiBlockConverter(Converter):
    tags = ["asdf://somewhere.org/tags/multi_block_data-1.0.0"]
    types = [MultiBlockData]

    def to_yaml_tree(self, obj, tag, ctx):
        if not len(obj.keys):
            obj.keys = [ctx.generate_block_key() for _ in obj.data]
        indices = [ctx.find_available_block_index(d, k) for d, k in zip(obj.data, obj.keys)]
        return {
            "indices": indices,
        }

    def from_yaml_tree(self, node, tag, ctx):
        indices = node["indices"]
        keys = [ctx.generate_block_key() for _ in indices]
        cbs = [ctx.get_block_data_callback(i, k) for i, k in zip(indices, keys)]
        obj = MultiBlockData([cb() for cb in cbs])
        obj.keys = keys
        return obj


class MultiBlockExtension(Extension):
    tags = ["asdf://somewhere.org/tags/multi_block_data-1.0.0"]
    converters = [MultiBlockConverter()]
    extension_uri = "asdf://somewhere.org/extensions/multi_block_data-1.0.0"

with asdf.config_context() as cfg:
    cfg.add_extension(MultiBlockExtension())
    obj = MultiBlockData([np.arange(3, dtype="uint8") + i for i in range(3)])
    ff = asdf.AsdfFile({"example": obj})
    ff.write_to("multi_block_converter_example.asdf")

multi_block_converter_example.asdf

#ASDF 1.0.0
#ASDF_STANDARD 1.5.0
%YAML 1.1
%TAG ! tag:stsci.edu:asdf/
--- !core/asdf-1.1.0
asdf_library: !core/software-1.0.0 {author: The ASDF Developers, homepage: 'http://github.com/asdf-format/asdf',
  name: asdf, version: 3.5.0}
history:
  extensions:
  - !core/extension_metadata-1.0.0
    extension_class: asdf.extension._manifest.ManifestExtension
    extension_uri: asdf://asdf-format.org/core/extensions/core-1.5.0
    manifest_software: !core/software-1.0.0 {name: asdf_standard, version: 1.1.1}
    software: !core/software-1.0.0 {name: asdf, version: 3.5.0}
  - !core/extension_metadata-1.0.0 {extension_class: builtins.MultiBlockExtension,
    extension_uri: 'asdf://somewhere.org/extensions/multi_block_data-1.0.0'}
example: !<asdf://somewhere.org/tags/multi_block_data-1.0.0>
  indices: [0, 1, 2]
...

BLOCK 0:
    allocated_size: 3
    used_size: 3
    data_size: 3
    data: b'000102'

BLOCK 1:
    allocated_size: 3
    used_size: 3
    data_size: 3
    data: b'010203'

BLOCK 2:
    allocated_size: 3
    used_size: 3
    data_size: 3
    data: b'020304'

#ASDF BLOCK INDEX
%YAML 1.1
---
- 817
- 874
- 931
...

Entry point performance considerations

For the good of asdf users everywhere, it’s important that entry point methods load as quickly as possible. All extensions must be loaded before reading an ASDF file, and therefore all converters are created as well. Any converter module or __init__ method that lingers will introduce a delay to the initial call to asdf.open. For that reason, we recommend that converter authors minimize the number of imports that occur in the module containing the Converter implementation, and defer imports of serializable types to within the from_yaml_tree method. This will prevent the type from ever being imported when reading ASDF files that do not contain the associated tag.