YOLOX - Train and Deploy on Ambarella

The algorithm delivers a fully automated, end-to-end pipeline for training and deploying YOLOX small which is a high-accuracy, anchor-free object detection network optimized for Ambarella edge AI processors. Users supply a custom labeled dataset and configure key hyperparameters (image resolution, number of classes, epochs, batch size, learning rate, and target chipset), and the algorithm manages the complete training workflow on AWS GPU instances.

Upon training completion, the best checkpoint is automatically compiled into Ambarella CVFlow artifacts. The tools generate an ambapb checkpoint (compiled artifacts represented in ONNX format) for host evaluation and a cavalry binary to deploy on Ambarella CV-series chipsets. All compiled outputs are stored as SageMaker model artifacts, ready for deployment.

This SageMaker algorithm provides an end-to-end workflow to train and deploy a YOLOX small variant, a high-accuracy, anchor-free object detection network, for Ambarella edge AI processors. Users can bring their own labeled dataset, configure core hyperparameters (image resolution, number of classes, epochs, batch size, learning rate, and chipset), and run training on AWS GPU infrastructure using a managed pipeline. After training, the best checkpoint is automatically converted into Ambarella CVFlow deployment artifacts. The toolchain produces an AmbaPB checkpoint for host-side evaluation and a cavalry binary for deployment on Ambarella CV-series chipsets, including CV72, CV75, CV7, and N1-655. Generated outputs are packaged as SageMaker model artifacts for straightforward deployment and lifecycle management.

For inference, the bundled server loads CVFlow artifacts and exposes REST endpoints compatible with SageMaker real-time and asynchronous inference. Input images are preprocessed, executed through Ambarella runtime, and returned as JSON detections, enabling low-latency and scalable object detection workflows from model training to production deployment.