Adavis: adaptive training framework and system optimization for large-scale visual data

Aiming at the pain points of low training efficiency and poor generalisation of existing computer vision (CV) models in large-scale heterogeneous data scenarios, this paper proposes the AdaVis framework, which constructs a cross-layer optimisation closure loop of data-model-system through the collaborative design of AdaSelect, HAP and distributed gradient optimisation (GradReduce). The core innovations include a two-dimensional data screening mechanism based on gradient value and hardware cost, a hardware-aware real-time compression algorithm, and a communication optimisation protocol with gradient hierarchical sparsification. Experimental results show that AdaVis achieves a Top-1 accuracy of 86.7% on the ImageNet-21K dataset (a 2.1% improvement over ViT-Base), a 3.2-fold increase in training speed, a 22% reduction in communication volume compared to the traditional solution, and a 41% reduction in energy consumption per sample. In the industrial PCB defect detection scenario, the defect detection F1-score reaches 95.3%, and the inference speed reaches 83 FPS, which verifies the effectiveness of the framework in large-scale training and industrial implementation.