ADaPS: Adaptive Data Partitioning to Parallelize CNN Inference on Resource-Constrained(…)
summary
ADaPS is a framework that partitions CNN height/width on heterogeneous embedded devices with an analytical model that accounts for compute, memory and comms and looks for layer fusion; it builds a procedural partition tree on demand and searches with a tree-based hybrid optimization (Alpha‑Beta pruning + dynamic programming). Implementation uses PyTorch distributed and Raspberry Pi 4/5; evaluation on 2–4 nodes shows around 15–18% latency improvement (up to 1.2×) and much lower data transfer than prior work, with sync overheads rising at 4 nodes. DOI: 10.1145/3735452.3735532
AI-written summary · about
chapters (12)
- 0:00ADaPS approach
- 1:07CNN inference on edge devices
- 2:10Partitioning CNN layers
- 3:33Decentralized systems comparison
- 4:58ADaPS algorithm steps
- 6:45Partition tree structure
- 8:08Dynamic programming optimizations
- 9:12Evaluation setup
- 10:17Results for two and four node systems
- 10:38Pruning impact on search space
- 10:59Conclusions and future work
- 11:21Q&A
related talks
06mContext-aware Adaptive Surgery: A Fast and Effective Framework for Adaptative Model Partition
13mLightning Talk: Exploring PiPPY, Tensor Parallel and Torchserve for Large...
09mHyperParallel: A SuperPoD-Aware Distributed Acceleration Library
24mFast training and near-linear scaling with DataParallel in Amazon SageMaker
10mPyTorch Distributed Data Parallel (DDP)
12mEfficient Workflow Serving for Diffusion Models with Many Adapters
12mElastic Resource Sharing for Distributed Deep Learning
08mPyTorch Distributed: Towards Large Scale Training