A Unified Generative Adversarial Network Training via Self-Labeling and Self-Attention

Authors: Tomoki Watanabe, Paolo Favaro

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on CIFAR-10, STL-10 and SVHN, and show that both self-labeling and self-attention consistently improve the quality of generated data.
Researcher Affiliation Collaboration Tomoki Watanabe 1 Paolo Favaro 2 [...] 1Toshiba Corporation, Kawasaki, Japan 2University of Bern, Bern, Switzerland.
Pseudocode Yes We show the pseudo code of the training in Algorithm 1.
Open Source Code No The paper does not provide any explicit statement about releasing open-source code for the methodology, nor does it include any links to a code repository.
Open Datasets Yes We evaluate our method on CIFAR-10 (Krizhevsky & Hinton, 2009), STL-10 (Coates et al., 2011), and SVHN (Netzer et al., 2011) datasets using the Big GAN model (Brock et al., 2019) and show that our method improves the quality of the generated images in terms of the FID (Fr echet Inception Distance) score (Heusel et al., 2017).
Dataset Splits No The paper mentions 'FID-5ep, which is the FID averaged over the last 5 epochs of 5 evaluation runs' and discusses 'unlabeled' and 'labeled' datasets. However, it does not specify a distinct validation dataset split with percentages, sample counts, or a reference to a predefined split for their experiments.
Hardware Specification Yes We train the network on a single GPU: Ge Force RTX 2080ti for CIFAR-10 and SVHN and Quadro RTX 6000 for STL-10.
Software Dependencies No The paper mentions several components and techniques like 'Resnet18', 'Rand Augment', 'Big GAN', 'Adam s optimizer', and 'hinge loss'. However, it does not provide specific version numbers for any programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes To train the teacher we use Nesterov s momentum optimizer with a batch size of 64, momentum 0.9, K = 10, EMA decay of 0.999, and the number of epochs is 40 and 60 on the unlabeled and labeled datasets respectively. To train the c GAN we use Adam s optimizer with a batch size of 128 for the loss with the artificial labels and of 64 for the other losses, Th = 0.95, and the gradient penalty weight is 10.