Three-Player Wasserstein GAN via Amortised Duality

Authors: Nhan Dam, Quan Hoang, Trung Le, Tu Dinh Nguyen, Hung Bui, Dinh Phung

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present our experimental results on a synthetic and 2 real-world datasets (i.e. CIFAR-10 [Krizhevsky and Hinton, 2009] and Celeb A [Liu et al., 2015]). The synthetic experiment empirically demonstrates the stable convergence property of our proposed 3P-WGAN. On the other hand, experiments on real-world datasets show that 3P-WGAN outperforms WGAN [Arjovsky et al., 2017] and DCGAN [Radford et al., 2015], and yields comparable results to WGAN-GP [Gulrajani et al., 2017].
Researcher Affiliation Collaboration Nhan Dam1 , Quan Hoang1 , Trung Le1 , Tu Dinh Nguyen1 , Hung Bui2 and Dinh Phung1 1 Monash University 2 Google Deep Mind {nhan.dam, quan.hoang, trunglm, tu.dinh.nguyen}@monash.edu, buih@google.com, dinh.phung@monash.edu
Pseudocode Yes Algorithm 1 Update scheme of 3P-WGAN.
Open Source Code Yes We use Tensor Flow [Abadi et al., 2016] and our code is available on Git Hub1. 1https://github.com/nhandam/p3_wgan
Open Datasets Yes In this section, we present our experimental results on a synthetic and 2 real-world datasets (i.e. CIFAR-10 [Krizhevsky and Hinton, 2009] and Celeb A [Liu et al., 2015]).
Dataset Splits No The paper mentions dataset sizes but does not explicitly provide training/validation/test splits (percentages, counts, or explicit mention of a validation set).
Hardware Specification No The paper does not specify any particular hardware components (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions 'Tensor Flow [Abadi et al., 2016]' and 'Adam optimiser [Kingma and Ba, 2014]' but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We employ Adam optimiser [Kingma and Ba, 2014] with learning rate of 0.0002 and exponential decay rates β1, β2 of 0.0, 0.9. The learning rate is decayed linearly over 100,000 generator updates. Due to its critical role, the mover h is updated 5 times for each weight update of the critic and generator. Other settings include: (i) weights are randomly initialised from Gaussian distribution N (0, 0.02I) with zero bias; (ii) mini-batch size for training each of 3 players is 64; (iii) and the generator is fed with 128 noise units drawn from a uniform distribution.