Training with Quantization Noise for Extreme Model Compression

TL;DR

This paper introduces a novel quantization noise training method that enables extreme model compression while preserving accuracy, surpassing previous techniques especially in NLP and image classification tasks.

Contribution

The authors propose a new approach to quantization-aware training that allows for unbiased gradients during extreme compression, extending beyond traditional int8 quantization.

Findings

Achieved 82.5% accuracy on MNLI with a 14MB RoBERTa model.

Reached 80.0% top-1 accuracy on ImageNet with a 3.3MB EfficientNet-B3.

Established new state-of-the-art accuracy-compression trade-offs.

Abstract

We tackle the problem of producing compact models, maximizing their accuracy for a given model size. A standard solution is to train networks with Quantization Aware Training, where the weights are quantized during training and the gradients approximated with the Straight-Through Estimator. In this paper, we extend this approach to work beyond int8 fixed-point quantization with extreme compression methods where the approximations introduced by STE are severe, such as Product Quantization. Our proposal is to only quantize a different random subset of weights during each forward, allowing for unbiased gradients to flow through the other weights. Controlling the amount of noise and its form allows for extreme compression rates while maintaining the performance of the original model. As a result we establish new state-of-the-art compromises between accuracy and model size both in natural language processing and image classification. For example, applying our method to state-of-the-art Transformer and ConvNet architectures, we can achieve 82.5% accuracy on MNLI by compressing RoBERTa to 14MB and 80.0 top-1 accuracy on ImageNet by compressing an EfficientNet-B3 to 3.3MB.