AutoGAN: Neural Architecture Search for Generative Adversarial Networks

By Xinyu Gong, Shiyu Chang, Zhangyang Wang. 2019

289 citations recorded by the source corpus.

Neural architecture search (NAS) has witnessed prevailing success in image classification and (very recently) segmentation tasks. In this paper, we present the first preliminary study on introducing the NAS algorithm to generative adversarial networks (GANs), dubbed AutoGAN. The marriage of NAS and GANs faces its unique challenges. We define the search space for the generator architectural variations and use an RNN controller to guide the search, with parameter sharing and dynamic-resetting to accelerate the process. Inception score is adopted as the reward, and a multi-level search strategy is introduced to perform NAS in a progressive way. Experiments validate the effectiveness of AutoGAN on the task of unconditional image generation. Specifically, our discovered architectures achieve highly competitive performance compared to current state-of-the-art hand-crafted GANs, e.g., setting new state-of-the-art FID scores of 12.42 on CIFAR-10, and 31.01 on STL-10, respectively. We also conclude with a discussion of the current limitations and future potential of AutoGAN. The code is available at https://github.com/TAMU-VITA/AutoGAN

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Generative Models

Citations in the corpus

Score-based Generative Modeling in Latent Space (2021, background): “[83] Xinyu Gong, Shiyu Chang, Yifan Jiang, and Zhangyang Wang.”
Dual Contradistinctive Generative Autoencoder (2020, positive): “The decoder in Figure 15 matches the generator in [15].”
AutoML: A Survey of the State-of-the-Art (2019, methodology): “Another novel technique is Generative Adversarial Network (GAN) [54], which can be used to generate images 3https://datasetsearch.research.google.com/ 4https://www.datasearch.elsevier.com/ 5https://pytorch.org/docs/stable/torchvision/transforms.html and text data.”
AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks (2020, positive): “• (Gong et al., 2019) only demonstrated to find relatively small GAN models that synthesize low-resolution images from noise (e.g., a 6-layer generator on 32 × 32 CIFAR-10 images), and also did not consider image-to-image GANs.”
A Comprehensive Survey of Neural Architecture Search (2020, methodology): “With the improvement of NAS search efficiency, NAS is also quickly applied in the fields of object detection [65, 75, 111, 118], semantic segmentation [63, 64, 120], adversarial learning [53], architectural scaling [114, 122, 124], multi-objective optimization [39, 115, 125], platform-aware [28, 34,”
DARTS+: Improved Differentiable Architecture Search with Early Stopping (2019, background): “1 Introduction Neural Architecture Search (NAS) plays an important role in Automated Machine Learning (AutoML), which has attracted a lot of attention recently [3, 8, 20–22, 26, 35, 42].”
InfoMax-GAN: Improved Adversarial Image Generation via Information Maximization and Contrastive Learning (2020, background): “SNGAN is an unconditional GAN utilizing spectral normalization to stabilize GAN training by constraining the discriminator to be 1Lipschitz, and has been the basis of recent state-of-the-art GANs [30]–[33].”
Best Practices for Scientific Research on Neural Architecture Search (2019, background): “We note that application papers in NAS have already started tackling non-standard applications, such as image restoration (Suganuma et al., 2018), semantic segmentation (Chen et al., 2018; Nekrasov et al., 2018; Liu et al., 2019a), disparity estimation (Saikia et al., 2019), machine translation (So ”
FasterSeg: Searching for Faster Real-time Semantic Segmentation (2019, background): “To free human experts from this challenging trade-off, neural architecture search (NAS) has been recently introduced and drawn a booming interest (Zoph & Le, 2016; Brock et al., 2017; Pham et al., 2018; Liu et al., 2018a; Chen et al., 2018a; Bender et al., 2018; Chen et al., 2018c; Gong et al., 2019”
NAS-DIP: Learning Deep Image Prior with Neural Architecture Search (2020, methodology): “Motivated by the Neural Architecture Search (NAS) algorithm [24, 26, 89] which has been shown effective in discovering networks with top performance in a large search space, we leverage reinforcement learning (RL) with a recurrent neural network (RNN) controller [24, 26, 89] and use the PSNR as the ”
Neural Architecture Search for Deep Image Prior (2020, methodology): “For example, by exploring alternative policies such as proximal policy optimization [31] or Q-learning [2], RL approaches over RNN now scale to contemporary datasets e.g. NASNet for ImageNet classification [26, 47] and was recently explored for GAN over CIFAR-10 [16].”
TransGAN: Two Pure Transformers Can Make One Strong GAN, and That Can Scale Up (2021, methodology): “After its origin, GANs quickly embraced fully convolutional backbones [26], and inherited most successful designs from CNNs such as batch normalization, pooling, (Leaky) ReLU and more [38, 39, 40, 18].”
CARS: Continuous Evolution for Efficient Neural Architecture Search (2019, methodology): “To accelerate this process and make it automated, network architecture search (NAS) [41, 13, 33] has been proposed, and the learned architectures have exceeded those human-designed architectures on a variety of tasks.”
Efficient Architecture Search for Continual Learning (2020, background): “NAS is another promising research topic in the AI community [24], [25].”
AdaBits: Neural Network Quantization With Adaptive Bit-Widths (2019, unknown): “The NAS algorithms also benefit other tasks, such as generative adversarial networks [11], object detection [5] and segmentation [23].”
TransGAN: Two Transformers Can Make One Strong GAN (2021, methodology): “Those works are consistently able to discover stronger GAN architectures beyond the standard ResNet topology (Gong et al., 2019; Gao et al., 2020; Tian et al., 2020).”