Learning Semantic-aware Normalization for Generative Adversarial Networks

Authors: Heliang Zheng, Jianlong Fu, Yanhong Zeng, Jiebo Luo, Zheng-Jun Zha

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our approach outperforms the SOTA style-based approaches in both unconditional image generation and conditional image inpainting tasks.
Researcher Affiliation Collaboration Heliang Zheng1 , Jianlong Fu2, Yanhong Zeng3*, Jiebo Luo4, Zheng-Jun Zha1 1University of Science and Technology of China, Hefei, China 2Microsoft Research, Beijing, China 3Sun Yat-sen University, Guangzhou, China 4University of Rochester, Rochester, NY
Pseudocode No The paper describes algorithms and modules in text and equations, but does not include a clearly labeled pseudocode or algorithm block.
Open Source Code Yes More details can be found in our code https://github.com/researchmm/Sari GAN.
Open Datasets Yes We conduct experiments on both unconditional image generation and conditional image inpainting. For unconditional image generation, we evaluate our Sari GAN on three datasets, including LSUN CATS [31], LSUN CARS [31], and FFHQ [2] dataset. For image inpainting, we use Paris Street View [32] for evaluation.
Dataset Splits No The paper mentions 'training images' and evaluates on datasets with standard splits, but does not explicitly provide specific percentages, sample counts, or detailed methodology for its own dataset splits (train/validation/test).
Hardware Specification Yes We use Py Torch [34] as our codebase and run each experiment on 8 Tesla V100 GPUs for 7 days.
Software Dependencies No The paper mentions using 'PyTorch' as the codebase, but does not provide a specific version number for PyTorch or any other software dependencies.
Experiment Setup Yes The choice of hyperparameters: a smaller t in Equation 2 would better approximate one-hot vector, while also makes the optimizing harder since the gradient would vanish. We defer such a trade-off to the network by making t learnable, and the learned t is around 0.3. The setting of the threshold r (e.g.,0.1) in Equation 2 is conditioned on the group number g (e.g.,16), and the principle is that each channel would have c/g similar channels after activated on average. The λ1 and λ2 in Equation 7 is experimentally set to be 2 and 10, respectively.