Improving Post Training Neural Quantization: Layer-wise Calibration and Integer Programming

TL;DR

This paper introduces a novel layer-wise calibration and integer programming approach to improve post-training neural quantization, enabling effective 4-bit quantization with minimal accuracy loss on vision and text models.

Contribution

It proposes a layer-wise optimization and integer programming method for better quantization, surpassing previous dynamic range setting techniques, and demonstrates state-of-the-art results.

Findings

Achieves less than 1% accuracy degradation with 4-bit quantization on ResNet50.

Less susceptible to over-fitting, effective on small calibration sets.

State-of-the-art results on vision and text models.

Abstract

Lately, post-training quantization methods have gained considerable attention, as they are simple to use, and require only a small unlabeled calibration set. This small dataset cannot be used to fine-tune the model without significant over-fitting. Instead, these methods only use the calibration set to set the activations' dynamic ranges. However, such methods always resulted in significant accuracy degradation, when used below 8-bits (except on small datasets). Here we aim to break the 8-bit barrier. To this end, we minimize the quantization errors of each layer separately by optimizing its parameters over the calibration set. We empirically demonstrate that this approach is: (1) much less susceptible to over-fitting than the standard fine-tuning approaches, and can be used even on a very small calibration set; and (2) more powerful than previous methods, which only set the activations' dynamic ranges. Furthermore, we demonstrate how to optimally allocate the bit-widths for each layer, while constraining accuracy degradation or model compression by proposing a novel integer programming formulation. Finally, we suggest model global statistics tuning, to correct biases introduced during quantization. Together, these methods yield state-of-the-art results for both vision and text models. For instance, on ResNet50, we obtain less than 1\% accuracy degradation --- with 4-bit weights and activations in all layers, but the smallest two. We open-sourced our code.