LEIP 2.8 Release

The Latent AI Efficient Inference Platform (LEIP) is a Software Development Kit (SDK) that simplifies and accelerates delivery of optimized models to edge devices. LEIP delivers better, faster, more efficient models and scales performance while easing hardware target challenges. LEIP 2.8 is now available. LEIP 2.8 includes significant improvements that help move optimized neural networks from development to device faster and with better results. New features in LEIP 2.8 include:

EfficientDet Recipes

Recipes are pre-qualified hardware configuration and settings templates that let users rapidly prototype models for different targets. In addition to Classifier recipes, LEIP 2.8 adds Recipes for popular new models including EfficientDet. EfficientDet is an increasingly popular object detection model created by Google that can fit many different problem sets. Multiple EfficientDet Recipes are now available and let you find the model that works best for your target hardware and configuration. 

MobileNet SSD Recipes

The MobileNet SSD is a robust detection model that can fit on even the smallest devices. Multiple models are now included in LEIP to support rapid prototyping of models for different hardware targets. All new LEIP models also include the ‘bring your own data’ path, pre and post processors, and evaluation support.

LEIP Package + Latent Runtime Engine (LRE)

LEIP 2.8 includes a new LEIP Package and LRE release. The LRE is a Python C++ wrapper that allows for near instant deployment of your model to any device. You no longer have to track tricky dependencies as they are now all contained within the LRE object. The LRE will soon support additional plug-ins and custom services for additional features like encryption, model integrity, model update and quality checks that ensure models are running as fast as when they were deployed. 

LEIP Evaluate

LEIP Evaluate (our CLI) is now model agnostic and supports both Classifier and Detection models. LEIP Evaluate is also data agnostic and supports several different data types including MS COCO data types, KITTI, and PASCAL VOC. We also support custom pre- and post- processors so that your evaluation with Latent AI can be customized. 

Fuji Backend 

Our new SDK backend is codenamed Fuji to match the stability and support of the mountain it is named for. Fuji provides improvements across the board for both compilation and quantization. Several new quantization and calibration schemes including symmetric per channel quantization on ARM devices are also available. Other notable functionality includes bias correction, model coverage and introspection, and support for and batch sizes greater than one on all compilation forms: TRT and CPU.

For Documentation of our LEIP SDK User Guide and LEIP Recipes, visit our Resource Center. For more information, contact us at info@latentai.com

Watch an overview of LEIP 2.8 in the video below. For more videos, click here.