New lightweight detection transformer enables integer-only inference

Researchers have developed I-LW-DETR, a novel lightweight detection transformer that enables fully integer-only inference. This is a significant advancement for deploying such models on NPUs and microcontrollers, which traditionally struggle with floating-point operations common in transformers. The system incorporates a scale-preserving split convolution, a sign-dependent GELU approximation (SD-ShiftGELU), and a constrained Shiftmax function to ensure all operations, including nonlinearities, are executed using integer arithmetic. This approach results in models that are approximately 3.6 times smaller and over ten times more computationally efficient, with only a moderate loss in accuracy. AI

IMPACT Enables deployment of advanced vision transformers on resource-constrained devices like NPUs and microcontrollers.

RANK_REASON This is a research paper detailing a new technical approach for model inference. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• CNNS
• Gelu
• I-LW-DETR
• LayerNorm
• National Pingtung University of Science and Technology
• SD-ShiftGELU
• Shiftmax
• Softmax

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →