Model Registry

DeGirum Tools API Reference Guide. YAML registry for selecting models by task, hardware, and runtime defaults.

This API Reference is based on DeGirum Tools version 0.22.4.

Model Registry Overview

Use this module when you want to use a working model without memorizing model details. ModelRegistry acts as a guided menu of DeGirum models, while ModelSpec captures the handful of settings required to load one of those choices. Together they help pick a model with confidence and launch inference in just a few lines of code.

Key Features

• Capture one model request with ModelSpec so it can be shared, reused, or versioned alongside your project.

• Browse curated registries to surface models by goal, compatible hardware, or descriptive metadata.

• Layer simple filters to shrink the catalog to only the options that fit your scenario.

• Keep common defaults (host, zoo, properties) in a single place so every run follows the same playbook.

Typical Usage

1. Point ModelRegistry at the YAML file or hosted URL that lists the models available to your team.

2. Narrow the registry with helpers such as for_task and for_hardware to focus on the models that match your intent.

3. Choose a remaining ModelSpec (or let ranking helpers do it) to represent the model you plan to run.

4. Call ModelSpec.load_model() to launch the model and start running inference.

Example:

from degirum_tools import Display, ModelRegistry, ModelSpec, remote_assets

registry = ModelRegistry(
    config_file="https://assets.degirum.com/registry/models.yaml",
)

model_spec = (
    registry
    .for_task("coco_detection")
    .top_model_spec()
)

model = model_spec.load_model()
inference_result = model(remote_assets.three_persons)

print(inference_result)

with Display("Model Registry Demo") as output_display:
    output_display.show_image(inference_result.image_overlay)

Integration Notes

• Registry files can live in source control or be hosted online; point the constructor at whichever location you maintain.

• ModelSpec.load_model() opens the connection with degirum.connect on your behalf, so you do not have to manage sessions manually.

• Override defaults as needed when experimenting, without editing the shared registry catalog.

Key Functions

• ModelSpec.load_model() turns a saved specification into a ready-to-run model.

• ModelSpec.ensure_local() downloads the assets you need for offline use and returns a spec that targets @local.

• ModelRegistry.all_model_specs() lists the prepared specs that match the filters you have applied.

• ModelRegistry.best_model_spec() helps you pick a model based on ranking fields such as accuracy scores.

Configuration Options

• Registry defaults set shared values such as the inference host, zoo URL, tokens, and model properties.

• Metadata filters accept literal values or simple callables when you want to match custom fields in your catalog.

Classes

ModelSpec

ModelSpec

dataclass

Serializable description of a single model load request.

Attributes:

Name	Type	Description
model_name	str	Exact model identifier expected by the zoo. Ignored when model_url is provided.
zoo_url	str	Zoo location that hosts this model. When left empty, registry defaults or explicit overrides must supply the value.
model_url	str	Direct URL to the model file. When set, overrides model_name and zoo_url.
inference_host_address	str	Target inference host in PySDK locator format (for example @cloud or @local).
token	str | None	Optional authentication token for the zoo.
model_properties	dict[str, Any]	Keyword arguments forwarded to dg.ZooManager.load_model.
metadata	dict | None	Free-form informational payload, typically copied from the registry entry.

Examples:

spec = ModelSpec(
    model_name="<model>",
    zoo_url="<zoo>",
    inference_host_address="@local",
)
spec.load_model()

spec = ModelSpec(model_url="<model_url>")
spec.load_model()

ModelSpec Methods

__post_init__

__post_init__()

Validate required fields, support model_url, and normalize values.

Raises:

Type	Description
ValueError	If mandatory attributes are missing or inconsistent.

download_model(destination=None, ...)

download_model(destination=None, *, cloud_sync=False)

Download the model file to a local directory.

Parameters:

Name	Type	Description	Default
destination	str | Path | None	Destination directory or path. If a directory is provided, the model is saved into a subdirectory named after the model. When None, the default application data directory is used.	None
cloud_sync	bool	When True, checks the cloud zoo for an updated model version and downloads it if the local copy is missing or out of date.	False

Returns:

Name	Type	Description
local_path	Path	Path to the downloaded model assets directory.

ensure_local(cloud_sync=False)

ensure_local(cloud_sync=False)

Ensures the model is present locally; downloads if needed. Returns a new ModelSpec with zoo_url set to local path and inference_host_address to '@local'.

Parameters:

Name	Type	Description	Default
cloud_sync	bool	When True, checks the cloud zoo for an updated model version and downloads it if the local copy is missing or out of date.	False

Returns:

Name	Type	Description
local_spec	ModelSpec	New specification pointing to the local model copy.

load_model(zoo=None)

load_model(zoo=None)

Resolve the specification into a ready-to-use model instance.

Parameters:

Name	Type	Description	Default
zoo	ZooManager | None	Optional pre-connected inference manager. When None, zoo_connect is called automatically.	None

Returns:

Name	Type	Description
model	Model	Loaded model returned by the PySDK.

zoo_connect

zoo_connect()

Create a connection to the configured model zoo.

Returns:

Name	Type	Description
inference_manager	ZooManager	Reusable inference manager suitable for repeated calls to load_model.

ModelRegistry

ModelRegistry

Queryable collection of ModelSpec entries.

Registry data is sourced from structured YAML files that describe the available models, their target tasks, and compatible hardware. Instances remain immutable, and filtering methods return new copies so intermediate views can be chained without side effects.

Examples:

registry = ModelRegistry(
    config_file="https://assets.degirum.com/registry/models.yaml",
)
filtered = registry.for_task("coco_detection")
first_spec = filtered.top_model_spec()
print(first_spec.model_name)

ModelRegistry Methods

__init__(*, ...)

__init__(*, config=None, config_file=None)

Create a registry from a configuration dictionary or YAML file.

Parameters:

Name	Type	Description	Default
config	dict[str, dict] | None	Parsed registry configuration. If provided, config_file is ignored.	None
config_file	Path | str | None	Path or URL to a YAML registry document. Defaults to models.yaml located alongside this module. The latest DeGirum-managed catalog is published at https://assets.degirum.com/registry/models.yaml.	None

Raises:

Type	Description
RuntimeError	If a remote registry cannot be retrieved.
ValidationError	When the configuration does not satisfy ModelRegistry.schema_text.

all_model_specs(*, ...)

all_model_specs(*, inference_host_address=None, zoo_url=None, token=None, model_properties=None)

Create ModelSpec objects for every model in the registry.

Parameters:

Name	Type	Description	Default
inference_host_address	str | None	Destination inference host. If omitted, the registry defaults are used.	None
zoo_url	str | None	Override for the zoo URL. A common use case is redirecting all specs to a local zoo during testing.	None
token	str | None	Authentication token to apply to each spec.	None
model_properties	dict | None	Keyword arguments merged into the default properties and applied to every resulting spec.	None

Returns:

Name	Type	Description
specs	list[ModelSpec]	Specifications matching the current filtered view.

best_model_spec(key, ...)

best_model_spec(key, compare='max', **kwargs)

Select the model with the best numeric metadata value.

Parameters:

Name	Type	Description	Default
key	str	Metadata field to inspect.	required
compare	str	Either "max" (default) for the largest value or "min" for the smallest.	'max'
**kwargs	dict[str, Any]	Overrides forwarded to all_model_specs.	{}

Returns:

Name	Type	Description
best_spec	ModelSpec	Model whose metadata matches the requested criteria.

Raises:

Type	Description
RuntimeError	If the registry contains no models.
ValueError	If none of the models expose the requested key or provide numeric values.

for_alias(alias)

for_alias(alias)

Filter models by alias.

Parameters:

Name	Type	Description	Default
alias	str	Registry alias to match exactly.	required

Returns:

Name	Type	Description
registry	ModelRegistry	Filtered registry containing only matching models.

for_hardware(hardware)

for_hardware(hardware)

Filter models to those compatible with a specific hardware target.

Parameters:

Name	Type	Description	Default
hardware	str	Hardware identifier in RUNTIME/DEVICE format, for example N2X/ORCA1.	required

Returns:

Name	Type	Description
registry	ModelRegistry	Filtered registry containing only matching models.

for_meta(meta)

for_meta(meta)

Filter models by metadata key/value pairs.

Parameters:

Name	Type	Description	Default
meta	dict[str, Any]	Dictionary describing metadata criteria. Values may be callables that receive a metadata dictionary and return True when the model should be included. Non-callable values must match exactly.	required

Returns:

Name	Type	Description
registry	ModelRegistry	Filtered registry containing only matching models.

for_task(task)

for_task(task)

Filter models by the task label declared in the registry.

Parameters:

Name	Type	Description	Default
task	str	Task identifier such as face_detection or segmentation.	required

Returns:

Name	Type	Description
registry	ModelRegistry	Filtered registry containing only matching models.

get_aliases

get_aliases()

Return every unique alias present in the registry.

Returns:

Name	Type	Description
aliases	list[str]	Sorted list of aliases.

get_hardware

get_hardware()

Return every unique hardware target present in the registry.

Returns:

Name	Type	Description
hardware	list[str]	Sorted list of hardware identifiers.

get_tasks

get_tasks()

Return every unique task label present in the registry.

Returns:

Name	Type	Description
tasks	list[str]	Sorted list of task names.

top_model_spec(**kwargs)

top_model_spec(**kwargs)

Return the first model listed in the current registry view.

Parameters:

Name	Type	Description	Default
**kwargs	dict[str, Any]	Overrides forwarded to all_model_specs.	{}

Returns:

Name	Type	Description
top_spec	ModelSpec	Specification for the top-most entry.

Raises:

Type	Description
RuntimeError	If the filtered registry is empty.

with_defaults(*, ...)

with_defaults(*, inference_host_address=None, zoo_url=None, token=None, model_properties=None)

Return a copy of the registry with overridden default settings.

Parameters:

Name	Type	Description	Default
inference_host_address	str | None	Override for the default inference target.	None
zoo_url	str | None	Fallback zoo URL applied when entries omit an explicit value.	None
token	str | None	Token injected into downstream connections.	None
model_properties	dict | None	Base keyword arguments merged into every ModelSpec emitted by the registry.	None

Returns:

Name	Type	Description
registry	ModelRegistry	New registry with updated defaults applied.