This guide aims to provide an edge-cloud synergy training and inference framework based on eLxr system. First a typical two-node K8s cluster environment is set up on eLxr system, then KubeEdge is deployed to establish extending native containerized application orchestration capabilities to hosts at Edge. After the cloud-edge infrastructure is created, Sedna is applied on it to enable edge-cloud synergy capabilities which bring great experiences on federated inference, incremental learning, federated learning, and lifelong learning with the popular AI frameworks like TensorFlow, Pytorch, PaddlePaddle or MindSpore.
An edge node that enables edge-cloud synergy capabilities for AI
Let’s start the interesting example of joint inference in helmet detection scenario:
Comparison between Sedna’s official edge-side inference results and the edge-cloud collaborative inference results：
edge inference vs joint inference that I got on eLxr:
- Missing detection by the edge inference:
2. Detected by joint inference
The big model ‘yolov3_darknet.pb’ with higher accuracy runs by the cloud and the lightweight model ‘yolov3_resnet18.pb’ runs by the edge. The edge-cloud synergy capability makes the inference results more accurate and the inference speed faster.