Generative Ratio Matching Networks

Authors: Akash Srivastava, Kai Xu, Michael U. Gutmann, Charles Sutton

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we empirically compare GRAM-nets against MMD-GANs and vanilla GANs, on the Cifar10 and Celeb A image datasets.
Researcher Affiliation Collaboration Akash Srivastava MIT-IBM Watson AI Lab akash.srivastava@ibm.com Michael U. Gutmann University of Edinburgh michael.gutmann@ed.ac.uk Kai Xu University of Edinburgh kai.xu@ed.ac.uk Charles Sutton Google AI charlessutton@google.com
Pseudocode Yes Algorithm 1: Generative ratio matching
Open Source Code Yes 1Official implementations are available at https://github.com/GRAM-nets.
Open Datasets Yes In this section we empirically compare GRAM-nets against MMD-GANs and vanilla GANs, on the Cifar10 and Celeb A image datasets.
Dataset Splits No The paper states that FID is reported on a 'held-out set that was not used to train the models', implying a train/test split, but does not provide specific details on validation splits (e.g., percentages, sample counts) or how the data was partitioned into train, validation, and test sets.
Hardware Specification No The paper does not provide specific details regarding the hardware (e.g., GPU models, CPU types, memory specifications) used for running the experiments.
Software Dependencies No The paper mentions optimizers like ADAM and RMSprop, and points to an external implementation of MMD-GANs, but it does not specify version numbers for any key software dependencies or libraries required for reproduction.
Experiment Setup Yes To facilitate fair comparison with MMD-GAN we set all the hyperparameters shared across the three methods to the values used in Li et al. (2017). Therefore, we use a learning rate of 5e 5 and set the batch size to 64.