Towards Accurate Post-Training Quantization of Vision Transformers via Error Reduction.

Area of Science:

• Computer Vision
• Machine Learning
• Model Optimization

Background:

• Post-training quantization (PTQ) is crucial for efficient deployment of vision transformers (ViTs).
• Existing PTQ methods struggle with complex weight-activation interactions, leading to performance degradation.
• There's a need for PTQ techniques that minimize quantization errors in ViTs.

Purpose of the Study:

• To introduce ERQ, a two-step PTQ method for ViTs.
• To reduce quantization errors from both activations and weights.
• To enhance the performance of quantized ViTs.

Main Methods:

• ERQ employs a sequential two-step approach: Activation quantization error reduction (Aqer) and Weight quantization error reduction (Wqer).
• Aqer uses Reparameterization Initialization and Ridge Regression for activation errors.
• Wqer utilizes Dual Uniform Quantization and iterative Rounding Refinement with Ridge Regression for weight errors.

Main Results:

• ERQ significantly reduces quantization errors in ViTs.
• The method demonstrates superior performance across various ViT models and tasks.
• ERQ achieves a 36.81% accuracy improvement over GPTQ for W3A4 ViT-S.

Conclusions:

• ERQ effectively minimizes quantization errors in ViTs through its sequential two-step strategy.
• The proposed method offers a robust solution for deploying quantized ViTs with high accuracy.
• ERQ represents a significant advancement in PTQ techniques for vision transformers.