Improving Generative Adversarial Networks via Adversarial Learning in Latent Space

Authors: Yang Li, Yichuan Mo, Liangliang Shi, Junchi Yan

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on visual data show that our method can effectively achieve improvement in both quality and diversity. The implementation is publicly available at https://github.com/yangco-le/Adv Lat GAN.
Researcher Affiliation Academia Department of Computer Science and Engineering, Mo E Key Lab of Artificial Intelligence, Shanghai Jiao Tong University 2Key Lab. of Machine Perception (Mo E), School of Intelligence Science and Technology, Peking University {yanglily,shiliangliang,yanjunchi}@sjtu.edu.cn mo666666@stu.pku.edu.cn Junchi Yan is the correspondence author who is also with Shanghai AI Laboratory.
Pseudocode Yes We present the algorithm in Alg. 1 in the appendix which we call Adversarial Latent GAN with a post meaning quality (Adv Lat GAN-qua)... we finally develop our new diversity enhancing GAN training approach Adv Lat GAN-div as shown in Alg. 2 in the appendix...
Open Source Code Yes The implementation is publicly available at https://github.com/yangco-le/Adv Lat GAN.
Open Datasets Yes Experiments on Image Net and Celeb A are performed on GPUs of Tesla V100. Other public benchmark results are performed on a single GPU of Ge Force RTX 3090. We test Adv Lat GAN-z on Grid and Ring datasets... Results on MNIST. Results on STL-10. Results on AFHQ and FFHQ. Table 4 shows the results on Cifar-10 and STL-10... Table. 5 presents results on larger datasets: LSUN Church [59], Celeb A [60] and Image Net [61]...
Dataset Splits No No explicit mention of a 'validation set' or its specific split percentage/count was found in the provided text.
Hardware Specification Yes Experiments on Image Net and Celeb A are performed on GPUs of Tesla V100. Other public benchmark results are performed on a single GPU of Ge Force RTX 3090.
Software Dependencies No No specific software dependencies with version numbers were explicitly mentioned.
Experiment Setup Yes The number of latent iterations in training is set as one per discriminator step... The iteration step size ϵ of Adv Lat GAN-z is set to 0.05 and we conduct 20 steps each time. The iteration step size ϵ of Adv Lat GAN-z is set to 0.01 for SNGAN and 0.002 for WGANGP and we conduct 100 steps each time.