We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computer Vision and Pattern Recognition

Title: ClusterQ: Semantic Feature Distribution Alignment for Data-Free Quantization

Abstract: Network quantization has emerged as a promising method for model compression and inference acceleration. However, tradtional quantization methods (such as quantization aware training and post training quantization) require original data for the fine-tuning or calibration of quantized model, which makes them inapplicable to the cases that original data are not accessed due to privacy or security. This gives birth to the data-free quantization with synthetic data generation. While current DFQ methods still suffer from severe performance degradation when quantizing a model into lower bit, caused by the low inter-class separability of semantic features. To this end, we propose a new and effective data-free quantization method termed ClusterQ, which utilizes the semantic feature distribution alignment for synthetic data generation. To obtain high inter-class separability of semantic features, we cluster and align the feature distribution statistics to imitate the distribution of real data, so that the performance degradation is alleviated. Moreover, we incorporate the intra-class variance to solve class-wise mode collapse. We also employ the exponential moving average to update the centroid of each cluster for further feature distribution improvement. Extensive experiments across various deep models (e.g., ResNet-18 and MobileNet-V2) over the ImageNet dataset demonstrate that our ClusterQ obtains state-of-the-art performance.
Subjects: Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2205.00179 [cs.CV]
  (or arXiv:2205.00179v1 [cs.CV] for this version)

Submission history

From: Zhao Zhang [view email]
[v1] Sat, 30 Apr 2022 06:58:56 GMT (4748kb,D)

Link back to: arXiv, form interface, contact.