Dynamically Grown Generative Adversarial Networks

Authors: Lanlan Liu, Yuting Zhang, Jia Deng, Stefano Soatto8680-8687

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate new state-of-the-art of image generation. We evaluate DGGAN against manually designed Prog GAN and other recent GAN models on CIFAR-10 and LSUN.
Researcher Affiliation Collaboration Lanlan Liu*, 1 Yuting Zhang, 2 Jia Deng, 3 Stefano Soatto 2 1 University of Michigan, Ann Arbor 2 Amazon Web Services 3 Princeton University
Pseudocode Yes Algorithm 1: Top-K Greedy Pruning Algorithm
Open Source Code No The paper refers to using "the most popular Py Torch implementation1of Prog GAN" (Footnote 1: https://github.com/facebookresearch/pytorchGANzoo) to implement their DGGAN, but it does not explicitly state that their own DGGAN source code is released or available at this or any other link.
Open Datasets Yes CIFAR-10 (Krizhevsky 2009) contains 50k 32 32 training images. LSUN (Yu et al. 2015) has over a million 256 256 bedroom images for training.
Dataset Splits No The paper mentions datasets used for training but does not specify training, validation, or test splits or percentages for reproducing the data partitioning.
Hardware Specification No The paper states: "The resulting computational cost is 580 GPU days for 2k+ CIFAR-10 models and 1720 GPU days for 1k+ LSUN models." This indicates the use of GPUs but does not specify exact models, manufacturers, or other hardware details.
Software Dependencies No The paper states: "We use the most popular Py Torch implementation1of Prog GAN to obtain comprehensive Prog GAN results and to implement our DGGAN." It mentions PyTorch but does not provide a specific version number for PyTorch or any other software dependency.
Experiment Setup Yes We train initial candidates for 100k iterations and train each new candidate with 100k iterations after weight inheritance. We gradually increase the resolution from d0 = 8 to 32. After reaching the final resolution, following (Karras et al. 2018), we further train the fixed architecture longer to achieve convergence. We follow the same training schedule as Prog GAN.