ModelPack is offered as an outperforming Object Detector that is fully compatible with Au-Zone technologies middleware offerings (VisionPack, DeepViewRT and eIQ Toolkit). MPK for short, is based on one of the best state-of-the-art algorithms, which is highly optimized to improve detection rates while reducing inference latency at the edge. ModelPack allows user to achieve the highest accuracy while maintaining high throughput, over 30 frames per second on the i.MX 8M Plus.
This guide provides a quick walkthrough to demonstrate how to setup and train a ModelPack project. Au-Zone provides in the DeepView Dataset Zoo a few ready-to-use eIQ Projects, which include ModelPack checkpoints and pre-trained models. This Zoo is a free resource to quickly demonstrate diverse use cases using ModelPack.
Note: This section covers ModelPack with eIQ Portal. If you wish to use the stand-alone training pipeline please refer to ModelPack with Docker.
Note: In case this is your first time with eIQ Portal then the eIQ Portal User Guide should be reviewed to get familiar with the general workflow.
To make it possible for MPK to run on a stand alone computer system with eIQ Portal, there are some requirements that need to be fulfilled first:
- A PC with either of Windows, Linux, or Mac OS
- eIQ Toolkit 2.5.12 or higher
- ModelPack dataset - Playing Cards V7 Dataset
ModelPack v1.x is included in eIQ Toolkit 2.5.12 and above. Before starting, install the eIQ Toolkit on your desktop machine.
Before getting started, we need to download the ModelPack dataset. Copy the dataset into a custom directory with more than 5 GB of free space, ~1 GB is used for storing the dataset and ~4 GB extra space will be needed for all the test you might perform over the dataset.
For the first ModelPack project, we will open a sample project, obtain the evaluation license, train, validate and deploy the model to the target device.
Get Started - ModelPack Activation