IMPROVING LOW-BIT POST TRAINING QUANTIZATION: A DATA-FREE APPROACH
Keywords: Deep Neural Network, Quantization Aware Training, Post Training Quantization, Data Free Quantization
TL;DR: A data free post training quantization agnostic to batch normalization statistics from full precision model
Abstract: Post-training quantization (PTQ) without access to real data is enabling efficient model optimization and deployment in scenarios where privacy or proprietary constraints restrict the use of original datasets. Traditional data free quantization methods rely on Batch Normalization (BN) statistics from the trained full-precision model to generate calibration dataset for quantization. However, this reliance on BN statistics limits their applicability to deep neural networks (DNNs) without BN layer such as AlexNet. In this paper, we propose a calibration dataset generation algorithm that is agnostic to BN statistics, leveraging just the backpropagation to create synthetic images for PTQ. We also demonstrate that it is not necessary to include samples from every target category in the calibration dataset to get the representative activation ranges for quantization. Extensive experiments with both large and lightweight models on large-scale image classification tasks demonstrate that our method consistently improves quantization performance across various DNN architectures, especially in low-bit settings. Notably, in 4-bit quantization, we achieve an improvement of 3.42\% in top-1 accuracy for the ResNet18 model and 3.14\% for the InceptionV3 model compared to the state-of-the-art (SOTA) DSG method. Importantly, we use very few synthetic samples for quantization compared to other methods.
Primary Area: infrastructure, software libraries, hardware, systems, etc.
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2674
Loading