arm-ai-model-optimizer

Arm AI Model Optimizer command-line client and Python API.

This package installs the np command, a client that submits model optimization work to the Arm AI Model Optimizer service and displays the results. Model optimization is performed by the service. Profiling jobs can execute on target devices.

Requirements

• Python 3.11 or later
• Network access to the Arm AI Model Optimizer service
• For np profile, network access from the target device to the configured broker and S3 model storage

Installation

pip install arm-ai-model-optimizer

If you received a wheel file, install that file instead. Replace the quoted placeholder below with the actual path to the supplied wheel:

python -m pip install "<path-to-provided-wheel.whl>"

Getting started

1. Sign in

Issue a credential from the Arm AI Portal, then register it once on this machine:

np auth login <credential>
np auth status

The credential carries the address of the authentication server, so no further configuration is needed. np auth status reports the state of the stored token.

2. Check the connection

np health ready

3. Run an optimization

np workspace init      # create a workspace with a sample project
np models ls           # browse registered models
np validate            # check run.yaml against the backend
np run                 # submit the run
np status              # follow progress
np report              # inspect the results

Run np --help, or np <command> --help, for the full reference.

Commands

Some commands may be unavailable depending on the features enabled for your deployment.

Profiling and network access

When using np profile, the target device connects to a broker to receive work and to S3 storage to retrieve models. This is a networked workflow.

The separately supplied executor_runner profiling binaries can also be invoked directly on a target device without a server connection or authentication. Their standalone, local-processing description applies to direct execution of those binaries, not to the full np profile workflow. See the README supplied with the profiling binaries for supported models, devices and direct-execution instructions.

Python API

The same package provides a Python API. Every command above has a counterpart, with the same behaviour: where the CLI exits non-zero, the call raises.

The example below refers to registered model, device and dataset identifiers. Replace these identifiers with resources available in your deployment.

from netspresso.api import NetsPresso

client = NetsPresso()  # same environment and stored credential as the CLI
project = client.projects.create(name="resnet50-rpi5", model="resnet50_pt2_pytorch_static_fp32",
                                 device="raspberry_pi_5_16gb", runtime="executorch", backend="xnnpack",
                                 calibration_dataset="imagenet1k_calib_seed42_n1000",  # required for quantization
                                 evaluation_dataset="imagenet1k_eval_seed42_n5000")
experiment = client.experiments.create(project_path=project.project_path, steps="go,gq")
client.run(experiment_path=experiment.experiment_path, wait=True)
report = client.report(experiment_path=experiment.experiment_path)

See the Python API reference in the documentation below for every call.

Documentation

https://developer.arm.com/dev-docs/arm-ai-model-optimizer

License

This software is licensed under the End User License Agreement for Arm AI Model Optimizer CLI, SDK and Related Software, provided in Arm AI Model Optimizer CLI SDK - EULA.txt. Read it before you install or use the software.

Third-party components are subject to their respective license terms, as identified in the accompanying third-party notices.

Notices for the open source software this package depends on are provided in licenses/Open Source Software Notice.md, with the full license texts in licenses/Appendix 1/ and the copyright notices in licenses/Appendix 2 - Copyright Attribution.md.

The software bill of materials (SBOM) is provided in CycloneDX and SPDX formats in sboms/.

In the source distribution, these paths are relative to the top level. In an installed package, the EULA and the licenses/ directory are under arm_ai_model_optimizer-<version>.dist-info/licenses/, and the SBOM files are under arm_ai_model_optimizer-<version>.dist-info/sboms/.