Sinkhorn Natural Gradient for Generative Models

Authors: Zebang Shen, Zhenfu Wang, Alejandro Ribeiro, Hamed Hassani

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

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, we quantitatively compare Si NG with state-of-the-art SGD-type solvers on generative tasks to demonstrate its efficiency and efficacy of our method.
Researcher Affiliation Academia Department of Electrical and Systems Engineering :Department of Mathematics University of Pennsylvania {zebang@seas,zwang423@math,aribeiro@seas,hassani@seas}.upenn.edu
Pseudocode No The paper describes algorithms conceptually and refers to their implementation, but it does not include any explicitly labeled pseudocode blocks or algorithm listings.
Open Source Code No The paper refers to an Appendix E for implementation details but does not provide an explicit statement about open-sourcing the code or a link to a code repository.
Open Datasets Yes In our experiments, we pretrain the discriminators for the celeb A and cifar10 datasets. [...] Our experiment considers a specific instance of problem (30) where we take the measure β to be the distribution of the images in the Celeb A dataset. [...] Two specific instances are considered: we take the measure β to be the distribution of the images in either the Celeb A or the Cifar10 dataset.
Dataset Splits No The paper mentions using Celeb A and Cifar10 datasets but does not specify the training, validation, or test splits (e.g., percentages, sample counts, or references to standard splits with citations) within the provided text.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, or cloud resources) used for running the experiments.
Software Dependencies No The paper mentions 'modern auto-differential mechanisms such as Py Torch' and 'ADAM as the optimizer' but does not specify version numbers for any software dependencies, which are necessary for reproducibility.
Experiment Setup Yes The entropy regularization parameter γ is set to 0.01 for both the objective and the constraint. [...] The entropy regularization of the Sinkhorn divergence objective is set to γ 100 as suggested in Table 2 of [Genevay et al., 2018]. The regularization for the constraint is set to γ 1 in Si NG. We used ADAM as the optimizer for the discriminators (with step size 10 3 and batch size 4000).