TernScope co-designs ternary models for efficient segmentation
Original title: TernScope: Model-Hardware Co-Design for Efficient Ternary Endoscopic Segmentation
TernScope delivers a 0.2483M parameter model with 88.46% lower compute area than prior accelerators while retaining accuracy.
TernScope reconciles small size, accuracy, and speed for edge medical segmentation via model-hardware co-design. Split-channel blocks and branch fusion preserve context. Ternary Progressive Quantization-aware Training adapts the network from deep to shallow layers using ternary weights and 8-bit activations. Results hold on in-domain and external datasets.
- 0.2483 M parameters
- 0.0949 GMACs
- 88.46% compute-array area reduction
- All convolutions ternary weights
- TPQ adapts across decoder layers
From the original article
1 min read at huggingface.co
Edge medical segmentation must reconcile small model size, high accuracy, and fast inference. TernScope addresses this tradeoff through model--hardware co-design, learning a compact segmentation network whose low-bit operators also simplify hardware execution. Split-channel large-kernel blocks preserve spatial context with limited computation, while branch fusion removes redundant inference paths.
By Zhuoyu Wu, Wenhui Ou, Qiawei Zheng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.