PSAQ-ViT V2: Toward Accurate and General Data-Free Quantization for Vision Transformers.

Area of Science:

• Computer Vision
• Machine Learning
• Model Compression

Background:

• Data-free quantization addresses privacy concerns in model compression, crucial for sensitive data.
• Existing methods like PSAQ-ViT pioneered data-free quantization for Vision Transformers (ViTs) using patch similarity.
• There is a need for more accurate and generalizable data-free quantization techniques for ViTs.

Purpose of the Study:

• To introduce PSAQ-ViT V2, an advanced data-free quantization framework for Vision Transformers (ViTs).
• To improve the accuracy and generality of ViT quantization without access to real-world data.
• To provide a robust baseline for data-free quantization applicable to diverse vision tasks and models.

Main Methods:

• Introduced an adaptive teacher-student strategy for cyclic evolution of generated samples and the quantized student model.
• Utilized patch similarity metric from PSAQ-ViT, supervised by a full-precision teacher model.
• Employed task- and model-independent prior information, removing the need for auxiliary category guidance.

Main Results:

• PSAQ-ViT V2 achieved competitive results across image classification, object detection, and semantic segmentation tasks.
• Demonstrated strong performance on Swin-S backbone: 82.13 top-1 accuracy on ImageNet, 50.9 box AP on COCO, and 47.2 mIoU on ADE20K.
• Consistently achieved high accuracy with naive quantization strategies, without real-world data access.

Conclusions:

• PSAQ-ViT V2 offers a significant advancement in accurate and general data-free quantization for ViTs.
• The framework shows potential as a practical solution for real-world applications with sensitive data constraints.
• The released code and competitive results establish PSAQ-ViT V2 as a powerful baseline for future research.