ZERO-SHOT LEARNING OF A CONDITIONAL GENERATIVE ADVERSARIAL NETWORK FOR DATA-FREE NETWORK QUANTIZATION Yoojin Choi Mostafa El-Khamy Jungwon Lee

2025-05-06 0 0 4.44MB 5 页 10玖币

侵权投诉

ZERO-SHOT LEARNING OF A CONDITIONAL GENERATIVE ADVERSARIAL NETWORK

FOR DATA-FREE NETWORK QUANTIZATION

Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

SoC R&D, Samsung Semiconductor Inc., San Diego, CA 92121, USA

ABSTRACT

We propose a novel method for training a conditional gener-

ative adversarial network (CGAN) without the use of train-

ing data, called zero-shot learning of a CGAN (ZS-CGAN).

Zero-shot learning of a conditional generator only needs a

pre-trained discriminative (classiﬁcation) model and does not

need any training data. In particular, the conditional gener-

ator is trained to produce labeled synthetic samples whose

characteristics mimic the original training data by using the

statistics stored in the batch normalization layers of the pre-

trained model. We show the usefulness of ZS-CGAN in data-

free quantization of deep neural networks. We achieved the

state-of-the-art data-free network quantization of the ResNet

and MobileNet classiﬁcation models trained on the ImageNet

dataset. Data-free quantization using ZS-CGAN showed a

minimal loss in accuracy compared to that obtained by con-

ventional data-dependent quantization.

Index Terms—Zero-shot learning, conditional genera-

tive adversarial networks, data-free training, quantization

1. INTRODUCTION

Generative adversarial networks (GANs) [1] are of great in-

terest in deep learning for image or speech synthesis prob-

lems. Two neural networks, called generator and discrim-

inator, play a zero-sum game to learn the mapping from a

random noise distribution to the target data distribution. The

generator is trained to fool the discriminator by making its

fake samples as similar as possible to the real training data,

while the discriminator is trained to distinguish the fake sam-

ples produced by the generator from the real training data.

Conditional GANs (CGANs) [2] are the conditional version

of GANs, where both the generator and the discriminator are

conditioned on some extra information, such as classes or at-

tributes. CGANs have potential in various conditional gener-

ation tasks such as labeled image generation, image-to-image

translation, and so on [3]. In this paper, we propose a novel

method of training a conditional GAN without any training

data, called zero-shot learning of a CGAN (ZS-CGAN).

Network quantization is an important procedure for efﬁ-

cient inference when deploying pre-trained deep neural net-

works on resource-limited platforms [4]. By using quantized

Student

(quantized)

Teacher/Discriminator

(pre-trained)

Conditional

generator

Random

noise

Label

Constraint to match BN statistics

Cross-entropy ↓

Data-free

knowledge

distillation

Fig. 1: Zero-shot learning of a CGAN (ZS-CGAN) and data-

free knowledge distillation. First, we train a conditional gen-

erator without any training data. A pre-trained classiﬁcation

model (called teacher) plays the role of a (ﬁxed) discrimina-

tor that evaluates generated samples. In particular, the cross-

entropy between the generator input and the teacher output is

minimized. The generator is also constrained to produce syn-

thetic samples similar to the original training data by match-

ing the statistics at the batch normalization (BN) layers of the

teacher. Second, we transfer knowledge from the pre-trained

teacher to a (quantized) student via data-free knowledge dis-

tillation using the synthetic samples from the generator.

weights and activations, we not only reduce the computational

cost, but also lower the memory footprint required for infer-

ence. The improved efﬁciency after quantization is usually

traded off against an accuracy loss. To minimize the accu-

racy loss from quantization, the quantization parameters are

optimized with some calibration data in post-training quan-

tization (PTQ). To recover the accuracy loss, the quantized

model can also be re-trained with some training data, which

is called quantization-aware training (QAT).

Most of the existing network quantization methods are

data-dependent, implying that a large number of training data,

which were already used in training the ﬂoating-point model,

are assumed to be available and used in the quantization pro-

cedure (cf. [5–10]). However, it becomes more and more dif-

ﬁcult and expensive to share the training data due to their large

size, restriction by proprietary rights, and to preserve data pri-

vacy. The regulations and compliance requirements around

privacy and security complicate both data sharing by the orig-

inal model trainer and data collection by the model quantizer,

for example, in the case of medical and bio-metric data.

The conditional generator trained with our proposed ZS-

CGAN can generate samples whose statistics match those of

the training dataset. Hence, we propose utilizing it for data-

arXiv:2210.14392v1 [cs.CV] 26 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ZERO-SHOTLEARNINGOFACONDITIONALGENERATIVEADVERSARIALNETWORKFORDATA-FREENETWORKQUANTIZATIONYoojinChoi,MostafaEl-Khamy,JungwonLeeSoCR&D,SamsungSemiconductorInc.,SanDiego,CA92121,USAABSTRACTWeproposeanovelmethodfortrainingaconditionalgener-ativeadversarialnetwork(CGAN)withouttheuseoftrain-ingdata,calle...

展开>> 收起<<

ZERO-SHOT LEARNING OF A CONDITIONAL GENERATIVE ADVERSARIAL NETWORK FOR DATA-FREE NETWORK QUANTIZATION Yoojin Choi Mostafa El-Khamy Jungwon Lee.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

ZERO-SHOT LEARNING OF A CONDITIONAL GENERATIVE ADVERSARIAL NETWORK FOR DATA-FREE NETWORK QUANTIZATION Yoojin Choi Mostafa El-Khamy Jungwon Lee

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: