Small-GAN: Speeding up GAN Training using Core-Sets

Authors: Samarth Sinha, Han Zhang, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, Augustus Odena

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments showing that this technique substantially reduces training time and memory usage for modern GAN variants, that it reduces the fraction of dropped modes in a synthetic dataset, and that it allows GANs to reach a new state of the art in anomaly detection.
Researcher Affiliation Collaboration 1University of Toronto, Vector Institute 2Google Brain 3Mila, Universite de Montreal.
Pseudocode Yes Algorithm 1 Greedy Coreset
Open Source Code No The paper references using and modifying existing open-source code for SAGAN ('Using the code at https://github.com/heykeetae/Self-Attention-GAN'), but does not explicitly state that their specific implementation of Small-GAN or their modifications are publicly released.
Open Datasets Yes We conduct experiments on the CIFAR, LSUN, and Image Net data sets
Dataset Splits No The paper uses well-known datasets like CIFAR, LSUN, and ImageNet and evaluates performance on generated samples, but does not explicitly state the training, validation, and test splits (e.g., percentages or specific partitioning methodology) for these datasets, nor does it explicitly mention using a distinct validation set for hyperparameter tuning.
Hardware Specification Yes All the experiments were performed on a single NVIDIA Titan-XP GPU.
Software Dependencies No The paper mentions using 'PyTorch version of FID scores' and specific GAN architectures, but does not provide a reproducible list of software dependencies with specific version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions) that would allow for replication of the experiments.
Experiment Setup Yes For our Core-set algorithm, the distance function, d( , ) is the Euclidean distance for both the prior and target distributions. The only hyper-parameter altered is the batch-size, which is stated for all experiments. For over-sampling, we use a factor of 4 for the prior p(z) and a factor of 8 for the target, p(x), unless otherwise stated.