When, Why, and Which Pretrained GANs Are Useful?

Authors: Timofey Grigoryev, Andrey Voynov, Artem Babenko

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our conclusions are confirmed by experiments with the state-of-the-art Style GAN2 (Karras et al., 2020b), chosen due to its practical importance and a variety of open-sourced checkpoints, which can be used as pretrained sources.
Researcher Affiliation Collaboration Timofey Grigoryev Yandex grigorev.ta@phystech.edu Andrey Voynov Yandex an.voynov@yandex.ru Artem Babenko Yandex artem.babenko@phystech.edu
Pseudocode No The paper describes methods in text and uses figures, but no formal pseudocode or algorithm blocks are present.
Open Source Code Yes The code and pretrained models are available online at https://github.com/yandex-research/gan-transfer
Open Datasets Yes Table 5: Datasets information. CIFAR-104 https://www.cs.toronto.edu/~kriz/cifar.html FFHQ5 https://github.com/NVlabs/ffhq-dataset Flowers6 https://www.robots.ox.ac.uk/~vgg/data/flowers/102/index.html Grumpy-Cat7 https://hanlab.mit.edu/projects/data-efficient-gans/datasets/ Imagenet8 https://image-net.org/index.html LSUN Bedroom9 https://www.yf.io/p/lsun LSUN Cat9 https://www.yf.io/p/lsun LSUN Church9 https://www.yf.io/p/lsun LSUN Dog9 https://www.yf.io/p/lsun Satellite-Buildings10 https://www.aicrowd.com/challenges/mapping-challenge-old Satellite-Landscapes11 https://earthview.withgoogle.com Simpsons12 https://www.kaggle.com/c/cmx-simpsons/data Bre Ca HAD13 https://figshare.com/articles/dataset/Bre_Ca_HAD_A_Dataset_for_Breast_Cancer_Histopathological_Annotation_and_Diagnosis/7379186
Dataset Splits Yes As for LSUN-Bedroom, we split the original data into a train and a test subset in the proportion 9 : 1 and train e4e on the train set and evaluate on the test set.
Hardware Specification Yes Training is performed on eight Tesla V100 GPUs and takes approximately three hours per 1M real images shown to the discriminator.
Software Dependencies No We always use the official Py Torch implementation of Style GAN2-ADA (Karras et al., 2020a) provided by the authors1. (No specific version numbers for PyTorch or StyleGAN2-ADA are mentioned.)
Experiment Setup Yes We use the stylegan2 configuration in the ADA implementation with the default hyperparameters (same for all datasets). [...] In these experiments, we train for 25M real images shown to the discriminator. [...] We use batch size 64 and Adam optimizers with learning rate 0.0002 and β1, β2 = 0.5, 0.999.