Sharpness-Aware Data Generation for Zero-shot Quantization
This work addresses the challenge of quantizing neural networks without access to real training data, which is incremental by incorporating sharpness minimization into synthetic data generation.
The paper tackles the problem of zero-shot quantization by generating synthetic data that minimizes the sharpness of the quantized model, resulting in improved generalization and superior performance over state-of-the-art methods on CIFAR-100 and ImageNet datasets in low-bit settings.
Zero-shot quantization aims to learn a quantized model from a pre-trained full-precision model with no access to original real training data. The common idea in zero-shot quantization approaches is to generate synthetic data for quantizing the full-precision model. While it is well-known that deep neural networks with low sharpness have better generalization ability, none of the previous zero-shot quantization works considers the sharpness of the quantized model as a criterion for generating training data. This paper introduces a novel methodology that takes into account quantized model sharpness in synthetic data generation to enhance generalization. Specifically, we first demonstrate that sharpness minimization can be attained by maximizing gradient matching between the reconstruction loss gradients computed on synthetic and real validation data, under certain assumptions. We then circumvent the problem of the gradient matching without real validation set by approximating it with the gradient matching between each generated sample and its neighbors. Experimental evaluations on CIFAR-100 and ImageNet datasets demonstrate the superiority of the proposed method over the state-of-the-art techniques in low-bit quantization settings.