Group Equivariant Generative Adversarial Networks

Authors: Neel Dey, Antong Chen, Soheil Ghafurian

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

Reproducibility Variable Result LLM Response
Research Type Experimental 3 EXPERIMENTS Common setups. In each subsection, we list specific experimental design choices with full details available in App. C. For each comparison, the number of group-filters in each layer is divided by the square root of the cardinality of the symmetry set to ensure a similar number of parameters to the standard CNNs to enable fair comparison. We skew towards stabilizing training over absolute performance to compare models under the same settings to obviate extensive checkpointing typically required for Big GAN-like models.
Researcher Affiliation Collaboration Neel Dey New York University neel.dey@nyu.edu Antong Chen & Soheil Ghafurian Data Science & Scientific Informatics, Merck & Co., Inc. antong.chen@merck.com, soheilghafurian@gmail.com
Pseudocode No The paper illustrates architectural components using diagrams and describes steps in text and tables, but it does not include any formal pseudocode or algorithm blocks.
Open Source Code Yes Implementations are available at https://github.com/neel-dey/equivariant-gans.
Open Datasets Yes Table 1: A summary of the datasets considered in this paper. The right-most column indicates whether the dataset has a preferred pose. Dataset Resolution nclasses ntraining nvalidation Pose Preference Rotated MNIST (28, 28) 10 12,000 50,000 No ANHIR (128, 128, 3) 5 28,407 9,469 No LYSTO (256, 256, 3) 3 20,000 No CIFAR-10 (32, 32, 3) 10 50,000 10,000 Yes Food-101 (64, 64, 3) 101 75,747 25,250 Yes
Dataset Splits Yes Table 1: A summary of the datasets considered in this paper. The right-most column indicates whether the dataset has a preferred pose. Dataset Resolution nclasses ntraining nvalidation Pose Preference Rotated MNIST (28, 28) 10 12,000 50,000 No ANHIR (128, 128, 3) 5 28,407 9,469 No LYSTO (256, 256, 3) 3 20,000 No CIFAR-10 (32, 32, 3) 10 50,000 10,000 Yes Food-101 (64, 64, 3) 101 75,747 25,250 Yes
Hardware Specification No The paper mentions general terms like 'compute' but does not specify any particular hardware components such as GPU models, CPU models, or memory specifications used for experiments.
Software Dependencies No The paper mentions software like Adam and TensorFlow Inception-v3, but it does not provide specific version numbers for any software dependencies.
Experiment Setup Yes Optimization is performed via Adam (Kingma & Ba, 2014) with β1 = 0.0 and β2 = 0.9, as in Zhang et al. (2018); Brock et al. (2018). Unless otherwise noted, all discriminators are updated twice per generator update and employ unequal learning rates for the generator and discriminator following Heusel et al. (2017). We use an exponential moving average (α = 0.9999) of generator weights across iterations when sampling images as in Brock et al. (2018). All initializations use the same random seed, except for Rot MNIST where we average over 3 random seeds. An overview of the small datasets considered here is presented in Table 1.