The Qualcomm® AI Hub Models repository provides a curated collection of state-of-the-art machine learning models optimized for deployment on Qualcomm® devices. It targets developers and researchers aiming for efficient on-device inference, offering a wide range of pre-trained models across computer vision, audio, multimodal, and generative AI tasks.

How It Works

The project leverages Qualcomm's AI optimization expertise to deliver models with improved latency and memory footprints. Models can be compiled and profiled using the Qualcomm® AI Hub platform, which supports various Qualcomm® runtimes (AI Engine Direct, LiteRT) and hardware targets (CPU, GPU, NPU). This allows for efficient deployment and performance evaluation on specific Qualcomm chipsets and devices.

Quick Start & Requirements

• Install: pip install qai_hub_models (or pip install "qai_hub_models[yolov7]" for specific model dependencies).
• Prerequisites: Python, Qualcomm® AI Hub account and API token for full functionality.
• Setup: Requires configuration of AI Hub access via qai-hub configure --api_token API_TOKEN.
• Resources: Model compilation and profiling are performed on cloud-hosted Qualcomm® devices.
• Docs: https://aihub.qualcomm.com/

Highlighted Details

• Extensive model library covering image classification, object detection, segmentation, super-resolution, speech recognition, and generative AI.
• Support for various hardware targets including CPU, GPU, and NPU with different precision levels (FP32, FP16, INT16, INT8).
• Includes end-to-end CLI demos and Python application examples for easy integration and testing.
• Models can be compiled and quantized for optimized on-device performance.

Maintenance & Community

• Community support available via Slack: https://aihub.qualcomm.com/community/slack.
• Issues can be reported on GitHub: https://github.com/quic/ai-hub-models/issues.

Licensing & Compatibility

• Licensed under BSD-3.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

Some models may require specific hardware features or have limited precision support on older chipsets. The full suite of features, including compilation and profiling, necessitates an account and API token for the Qualcomm® AI Hub.