Specifying a model

Before diving into code, read this page to understand how PySDK represents and handles models.

Estimated read time: 3 minutes

At the core of PySDK are models. Specifying a model means declaring its artifacts (such as compiled binaries, config files, labels, and any pre/post-processing resources), where those artifacts are stored, and where inference will run—along with optional runtime properties and metadata.

We package all of this into a reusable object: ModelSpec.

Visualizing the ModelSpec

Here's an example of how to define and load a model:

from degirum_tools import ModelSpec

# Example ModelSpec
model_spec = ModelSpec(
    model_name="yolov8n_coco--640x640_quant_hailort_multidevice_1",
    zoo_url="degirum/hailo",
    inference_host_address="@local",
    model_properties={"device_type": ["HAILORT/HAILO8L", "HAILORT/HAILO8"]},
    # metadata={"accuracy_top1": 0.55, "fps": 120}  # optional
)
# Loading a model
model = model_spec.load_model()

What each field means

• model_name: unique model identifier, pointing to the model’s JSON file and its artifacts

• zoo_url: where the model’s artifacts are stored (e.g., Model Zoo or local path)

• inference_host_address: where inference runs (cloud, local, or AI Server)

• model_properties: optional runtime parameters such as device type, thresholds, batch size

• metadata: optional dictionary with descriptive info like accuracy, model size, throughput/FPS, or notes

1

What are the artifacts? → model_name

Each model is defined by a JSON file named after it (<model_name>.json). This file:

• defines preprocessing and postprocessing parameters

• references compiled binaries and label files

• may point to optional postprocessing scripts (e.g., postprocess.py) or postprocessing settings

2

Where are the artifacts stored? → zoo_url

Artifacts can be stored in:

• AI Hub: e.g., degirum/hailo

• Local folders: e.g., file:///path/to/degirum-model-zoo/

• AI server: e.g., aiserver://

3

Where will the model run? → inference_host_address

The inference engine can run in different environments:

• @cloud: inference runs on AI Hub

• @local: inference runs on your machine

• localhost: local AI Server

• <server-ip>:<port>: remote AI Server

4

What are the model's properties? → model_properties (optional)

Optional runtime parameters applied to the model.

5

What metadata is associated with the model? → metadata (optional)

Optional metadata dictionary for free-form model info, including:

• accuracy

• model size

• FPS

See the PySDK Core Concepts guide for details on:

• the model JSON structure

• postprocessing files

• supported inference and storage combinations