Topology Distance: A Topology-Based Approach for Evaluating Generative Adversarial Networks

Authors: Danijela Horak, Simiao Yu, Gholamreza Salimi-Khorshidi7721-7728

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments Datasets and Experimental Setup We compared our proposed TD (lower is better) with IS (higher is better), FID (lower is better), KID (lower is better) and GS (lower is better) as introduced in Related Work section. In addition to some simulated data, which we will introduce in the next Section, our experiments were carried out on the following four datasets: Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017), CIFAR10, corrupted CIFAR100 (CIFAR100C) (Hendrycks and Dietterich 2019) and Celeb A (Liu et al. 2015).
Researcher Affiliation Collaboration Danijela Horak,1 Simiao Yu, 1 Gholamreza Salimi-Khorshidi 1,2 1 AIG 2 University of Oxford danijela.horak@aig.com, simiao.yu@aig.com, reza.khorshidi@aig.com
Pseudocode Yes Algorithm 1 This algorithm is to compute the longevity vector l for a set of images. l is of length n, which represents living times of all n homology classes throughout filtration (see Method section for more details).
Open Source Code No The paper mentions implementing their algorithm using GUDHI and PyTorch and provides a link to the GUDHI library (http://gudhi.gforge.inria.fr/), but does not provide a link or explicit statement for their own source code.
Open Datasets Yes our experiments were carried out on the following four datasets: Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017), CIFAR10, corrupted CIFAR100 (CIFAR100C) (Hendrycks and Dietterich 2019) and Celeb A (Liu et al. 2015).
Dataset Splits No The paper mentions using specific datasets but does not provide explicit details on train, validation, or test splits, such as percentages or sample counts.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU, CPU models, or memory) used for running the experiments.
Software Dependencies No The paper mentions using 'Python version of GUDHI' and 'Py Torch' but does not specify the exact version numbers for these software components, nor for Python itself.
Experiment Setup Yes We trained a GAN model (WGANGP (Gulrajani et al. 2017)) on the training set of Celeb A; original images were cropped to be of size 64 64, and the model was then trained on them for 200 epochs with a batch size of 64.