Disconnected Manifold Learning for Generative Adversarial Networks

Authors: Mahyar Khayatkhoei, Maneesh K. Singh, Ahmed Elgammal

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct several experiments to illustrate the aforementioned shortcoming of GANs, its consequences in practice, and the effectiveness of our proposed modifications in alleviating these issues.
Researcher Affiliation Collaboration Mahyar Khayatkhoei Department of Computer Science Rutgers University m.khayatkhoei@cs.rutgers.edu Ahmed Elgammal Department of Computer Science Rutgers University elgammal@cs.rutgers.edu Maneesh Singh Verisk Analytics maneesh.singh@verisk.com
Pseudocode Yes See Appendix A for details of our algorithm and the DMGAN objectives. (Appendix A contains Algorithm 1 Training DMWGAN)
Open Source Code No The paper does not contain any explicit statement or link providing access to the source code for the described methodology.
Open Datasets Yes MNIST [16] is particularly suitable since samples with different class labels can be reasonably interpreted as lying on disjoint manifolds... We combine 20K face images from Celeb A dataset [17] and 20K bedroom images from LSUN Bedrooms dataset [27] to construct a natural image dataset supported on a disconnected manifold.
Dataset Splits No The paper does not explicitly provide information about training, validation, and test dataset splits.
Hardware Specification No The paper describes network architectures and training parameters but does not specify the hardware (e.g., GPU/CPU models) used for running the experiments.
Software Dependencies No The paper mentions using Adam optimizer and DCGAN-like networks but does not provide specific version numbers for software dependencies or libraries.
Experiment Setup Yes In all experiments, we train each model for a total of 200 epochs with a five to one update ratio between discriminator and generator... See Appendix B for details of our networks and the hyperparameters. (Appendix B states: We use Adam optimizer with β1 = 0 and β2 = 0.9 for both generator and discriminator. Learning rate for generator and discriminator is 1e-4, and for Q and prior is 1e-5. We also use a learning rate decay of 0.5 per 10000 iterations for the prior training. We use batch size of 64 for all experiments. We use 20 generators for MNIST and 5 for Face-Bed, unless otherwise stated. We train all models for 200 epochs.)