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

Download:

Current browse context:

cs.CV

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: Pruning Ternary Quantization

Abstract: Inference time, model size, and accuracy are three key factors in deep model compression.
Most of the existing work addresses these three key factors separately as it is difficult to optimize them all at the same time.
For example, low-bit quantization aims at obtaining a faster model; weight sharing quantization aims at improving compression ratio and accuracy; and mixed-precision quantization aims at balancing accuracy and inference time. To simultaneously optimize bit-width, model size, and accuracy, we propose pruning ternary quantization (PTQ): a simple, effective, symmetric ternary quantization method. We integrate L2 normalization, pruning, and the weight decay term to reduce the weight discrepancy in the gradient estimator during quantization, thus producing highly compressed ternary weights. Our method brings the highest test accuracy and the highest compression ratio. For example, it produces a 939kb (49$\times$) 2bit ternary ResNet-18 model with only 4\% accuracy drop on the ImageNet dataset. It compresses 170MB Mask R-CNN to 5MB (34$\times$) with only 2.8\% average precision drop. Our method is verified on image classification, object detection/segmentation tasks with different network structures such as ResNet-18, ResNet-50, and MobileNetV2.
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
Cite as: arXiv:2107.10998 [cs.CV]
  (or arXiv:2107.10998v2 [cs.CV] for this version)

Submission history

From: Dan Liu [view email]
[v1] Fri, 23 Jul 2021 02:18:00 GMT (545kb,D)
[v2] Wed, 26 Jan 2022 18:21:52 GMT (1659kb,D)

Link back to: arXiv, form interface, contact.