Searching towards Class-Aware Generators for Conditional Generative Adversarial Networks

Zhou, Peng; Xie, Lei; Zhang, Xiaopeng; Ni, Bingbing

doi:10.48550/arxiv.2006.14208

Cited by 5 publications

(5 citation statements)

References 25 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The mode collapse problem has been observed in GAN training [14], [15], [16], [17]. The mode collapse problem causes the generator to stick to some distributions (modes) of the real data.…”

Section: Log(1 − D(g(z)))]mentioning

confidence: 99%

See 1 more Smart Citation

Stabilizing and Improving Training of Generative Adversarial Networks Through Identity Blocks and Modified Loss Function

2023

View full text Add to dashboard Cite

Generative adversarial networks (GANs) are a powerful tool for synthesizing realistic images, but they can be difficult to train and are prone to instability and mode collapse. This paper proposes a new model called Identity Generative Adversarial Network (IGAN) that addresses these issues. This model is based on three modifications to the baseline deep convolutional generative adversarial network (DCGAN). The first change is to add a non-linear identity block to the architecture. This will make it easier for the model to fit complex data types and cut down on the time it takes to train. The second change is to smooth out the standard GAN loss function by using a modified loss function and label smoothing. The third and final change is to use minibatch training to let the model use other examples from the same minibatch as side information to improve the quality and variety of generated images. These changes help to stabilize the training process and improve the model's performance. The performance of the GAN models is compared using the inception score (IS) and the Fréchet inception distance (FID), which are widely used metrics for evaluating the quality and diversity of generated images. The effectiveness of our approach was tested by comparing an IGAN model with other GAN models on the CelebA and stacked MNIST datasets. Results show that IGAN outperforms all the other models, achieving an IS of 13.95 and an FID of 43.71 after traning for 200 epochs. In addition to demonstrating the improvement in the performance of the IGAN, the instabilities, diversity, and fidelity of the models were investigated. The results showed that the IGAN was able to converge to a distribution of the real data more quickly. Furthermore, the experiments revealed that IGAN is capable of producing more stable and high-quality images. This suggests that IGAN is a promising approach for improving the training and performance of GANs and may have a range of applications in image synthesis and other areas.INDEX TERMS Generative adversarial network, deep learning, mode collapse, label smoothing, identity block.

show abstract

“…The mode collapse problem has been observed in GAN training [14], [15], [16], [17]. The mode collapse problem causes the generator to stick to some distributions (modes) of the real data.…”

Section: Log(1 − D(g(z)))]mentioning

confidence: 99%

“…The mode collapse is frequently sacrificed for more realistic individual samples [17], [18]. This trade-off of mode collapse for high-quality, realistic samples can lead to a biased model that produces a racial or gender-biased image [19].…”

Section: Log(1 − D(g(z)))]mentioning

confidence: 99%

Stabilizing and Improving Training of Generative Adversarial Networks Through Identity Blocks and Modified Loss Function

2023

View full text Add to dashboard Cite

show abstract

“…NAS is also applied for image generation, in particular, designing network architectures for the generative adversarial networks (GAN) [269], [270], [353]. Recently, this method has been extended to conditional GAN where the generator varies among classes and thus the search space becomes exponentially larger [274]. Beyond the search spaces that are extendable (see Sections 2.2.2 and 2.2.4), there were efforts that directly search for powerful architectures for RNN design [354], [355], [356], [357] or language modeling [307], [308], [309], [314].…”

Section: Other Search Spacesmentioning

confidence: 99%

Weight-Sharing Neural Architecture Search: A Battle to Shrink the Optimization Gap

Xie¹,

Chen

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

Neural architecture search (NAS) has attracted increasing attentions in both academia and industry. In the early age, researchers mostly applied individual search methods which sample and evaluate the candidate architectures separately and thus incur heavy computational overheads. To alleviate the burden, weight-sharing methods were proposed in which exponentially many architectures share weights in the same super-network, and the costly training procedure is performed only once. These methods, though being much faster, often suffer the issue of instability. This paper provides a literature review on NAS, in particular the weight-sharing methods, and points out that the major challenge comes from the optimization gap between the super-network and the sub-architectures. From this perspective, we summarize existing approaches into several categories according to their efforts in bridging the gap, and analyze both advantages and disadvantages of these methodologies. Finally, we share our opinions on the future directions of NAS and AutoML. Due to the expertise of the authors, this paper mainly focuses on the application of NAS to computer vision problems and may bias towards the work in our group.

show abstract

“…Conditional GAN (cGAN) [23] adds conditional information to the generator and discriminator of GANs. There are some ways to incorporate class information into the generator, such as conditional batch normalization (CBN) [6], conditional instance normalization (CIN) [9,12], classmodulated convolution (CMConv) [44], etc. There are also different ways to add class information to the discriminator.…”

Section: Conditional Gansmentioning

confidence: 99%

Omni-GAN: On the Secrets of cGANs and Beyond

Zhou¹,

Xie²,

Ni³

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

It has been an important problem to design a proper discriminator for conditional generative adversarial networks (cGANs). In this paper, we investigate two popular choices, the projection-based and classification-based discriminators, and reveal that both of them suffer some kind of drawbacks that affect the learning ability of cGANs. Then, we present our solution that trains a powerful discriminator and avoids over-fitting with regularization. In addition, we unify multiple targets (class, domain, reality, etc.) into one loss function to enable a wider range of applications. Our algorithm, named Omni-GAN, achieves competitive performance on a few popular benchmarks. More importantly, Omni-GAN enjoys both high generation quality and low risks in mode collapse, offering new possibilities for optimizing cGANs. Code is available 1 .

show abstract

Searching towards Class-Aware Generators for Conditional Generative Adversarial Networks

Cited by 5 publications

References 25 publications

Stabilizing and Improving Training of Generative Adversarial Networks Through Identity Blocks and Modified Loss Function

Stabilizing and Improving Training of Generative Adversarial Networks Through Identity Blocks and Modified Loss Function

Weight-Sharing Neural Architecture Search: A Battle to Shrink the Optimization Gap

Omni-GAN: On the Secrets of cGANs and Beyond

Contact Info

Product

Resources

About