ProbGAN: Towards Probabilistic GAN with Theoretical Guarantees

Authors: Hao He, Hao Wang, Guang-He Lee, Yonglong Tian

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evidence on synthetic high-dimensional multi-modal data and image databases (CIFAR-10, STL-10, and Image Net) demonstrates the superiority of our method over both start-of-the-art multi-generator GANs and other probabilistic treatment for GANs. In this section, we evaluate our model with two inference algorithms proposed in Section 3.3 (denoted as Prob GAN-GMA and Prob GAN-PSA).
Researcher Affiliation Academia Hao He, Hao Wang, Guang-He Lee, Yonglong Tian Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology {haohe,hwang87,guanghe,yonglong}@mit.edu
Pseudocode Yes Algorithm 1: Our Adapted SGHMC Inference Algorithm
Open Source Code No We will release our evaluation code soon.
Open Datasets Yes We evaluate our method on 3 widely-adopted datasets: CIFAR-10 (Krizhevsky et al., 2010), STL-10 (Coates et al., 2011) and Image Net (Deng et al., 2009).
Dataset Splits No The paper mentions 'CIFAR-10 has 50k training and 10k test' but does not specify details of a validation set split (e.g., percentages, sample counts, or methodology for creating a validation split).
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or cloud computing instance specifications used for running the experiments.
Software Dependencies No The paper mentions using 'Tensorflow' and 'Py Torch (Paszke et al., 2017)' but does not provide specific version numbers for these or any other software dependencies crucial for replication.
Experiment Setup Yes For a fair comparison with baselines, we use the same settings as MGAN. We resize the STL-10 and Image Net images down to 48x48 and 32x32 respectively. ... All models are optimized by Adam(Kingma & Ba, 2014) with a learning rate of 2 104. For probabilistic methods, the SGHMC noise factor is set as 3 102. Following the configuration in MGAN, the batch size of generators and discriminators are 120 and 64.