Fisher GAN
Authors: Youssef Mroueh, Tom Sercu
NeurIPS 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our claims on both image sample generation and semi-supervised classification using Fisher GAN. We experimentally validate the proposed Fisher GAN. We report results on three benchmark datasets: CIFAR-10 [20], LSUN [21] and Celeb A [22]. |
| Researcher Affiliation | Industry | Youssef Mroueh , Tom Sercu mroueh@us.ibm.com, tom.sercu1@ibm.com Equal Contribution AI Foundations, IBM Research AI IBM T.J Watson Research Center |
| Pseudocode | Yes | Algorithm 1 Fisher GAN |
| Open Source Code | Yes | 1 Code is available at https://github.com/tomsercu/Fisher GAN |
| Open Datasets | Yes | We report results on three benchmark datasets: CIFAR-10 [20], LSUN [21] and Celeb A [22]. |
| Dataset Splits | No | The paper mentions using "validation loss" but does not provide specific details on the dataset splits, such as percentages or sample counts for training, validation, or test sets in the main text. "We show both train and validation loss on LSUN and CIFAR-10" implies validation was used but without specific split information. |
| Hardware Specification | Yes | All timings are obtained by running on a single K40 GPU on the same cluster. |
| Software Dependencies | No | The paper mentions using "ADAM [19] as optimizer" and that "All inception scores are computed from the same tensorflow codebase", but it does not provide specific version numbers for these or any other software components or libraries, which would be necessary for reproducible software setup. |
| Experiment Setup | Yes | For LSUN we use a higher number of D updates (nc = 5) , since we see similarly to WGAN that the loss shows large fluctuations with lower nc values. For CIFAR-10 and Celeb A we use reduced nc = 2 with no negative impact on loss stability. Algorithm 1 Fisher GAN Input: penalty weight, Learning rate, nc number of iterations for training the critic, N batch size. We used Adam [19] as optimizer for all our experiments, hyper-parameters given in Appendix F. We found λD = λG = 0.1 to be optimal. In this setting we found λD = 1.5, λG = 0.1 to be optimal. |