Energy-based Generative Adversarial Networks

Authors: Junbo Zhao, Michael Mathieu, Yann LeCun

ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we study the training stability of EBGANs over GANs on a simple task of MNIST digit generation with fully-connected networks. We run an exhaustive grid search over a set of architectural choices and hyper-parameters for both frameworks.
Researcher Affiliation Collaboration Junbo Zhao, Michael Mathieu and Yann Le Cun Department of Computer Science, New York University Facebook Artificial Intelligence Research {jakezhao, mathieu, yann}@cs.nyu.edu
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access to source code, such as a repository link or an explicit statement of code release.
Open Datasets Yes In this section, we study the training stability of EBGANs over GANs on a simple task of MNIST digit generation with fully-connected networks. using the LSUN bedroom dataset (Yu et al., 2015) and the large-scale face dataset Celeb A under alignment (Liu et al., 2015). We trained EBGANs to generate high-resolution images on Image Net (Russakovsky et al., 2015).
Dataset Splits No The paper does not provide specific dataset split information, such as exact percentages, sample counts, or a detailed splitting methodology for reproducibility across all datasets.
Hardware Specification No The paper does not provide specific hardware details used for running its experiments.
Software Dependencies No The paper mentions software components like Adam optimizer and Batch Normalization, but it does not provide specific version numbers for any ancillary software dependencies needed to replicate the experiment.
Experiment Setup Yes Formally, we specify the search grid in table 1. We impose the following restrictions on EBGAN models: (i)-using learning rate 0.001 and Adam (Kingma & Ba, 2014) for both G and D; (ii)-n Layer D represents the total number of layers combining Enc and Dec. For simplicity, we fix Dec to be one layer and only tune the Enc #layers; (iii)-the margin is set to 10 and not being tuned. Batch normalization (Ioffe & Szegedy, 2015) is applied after each weight layer, except for the generator output layer and the discriminator input layer (Radford et al., 2015). Training images are scaled into range [-1,1]. Correspondingly the generator output layer is followed by a Tanh function. Re LU is used as the non-linearity function. Initialization: the weights in D from N(0, 0.002) and in G from N(0, 0.02). The bias are initialized to be 0.