跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 2026-10-21AI 評分37

TernScope:面向高效三元內窺鏡分割的模型—硬件協同設計

TernScope: Model-Hardware Co-Design for Efficient Ternary Endoscopic Segmentation

AI 導讀

TernScope 協同設計內窺鏡分割模型與配套加速器,令全部 25 個卷積層採用三元權重,並以 8-bit 激活值進行運算。模型有 0.2483M 個參數,在 256×256 下運算量為 0.0949 GMACs;在匹配的 45 nm 綜合條件下,計算陣列面積較原始 INT8 Gemmini 低 88.46%。

正文

Hi everyone! I'm the first author of TernScope. Can a tiny segmentation network use ternary weights throughout—including the input and prediction layers—and translate that into smaller hardware?

We co-design an endoscopic segmentation model and a matching accelerator:
• 0.2483M parameters and 0.0949 GMACs at 256×256.
• All 25 convolutions use ternary weights {−1, 0, +1} and 8-bit activations, trained through branch fusion and progressive QAT.
• External evaluation on ClinicDB and ColonDB without fine-tuning.
• 88.46% lower compute-array area than original INT8 Gemmini under matched 45 nm synthesis.

Code, a W2A8 checkpoint, and accelerator RTL are available:
https://github.com/ReaganWu/TernScope

Happy to discuss full-network ternary quantization, efficient medical vision, and model–hardware co-design!

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co