Interpretable Generative Adversarial Networks
Authors: Chao Li, Kelu Yao, Jin Wang, Boyu Diao, Yongjun Xu, Quanshi Zhang1280-1288
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments have demonstrated the effectiveness of our method. We applied our method to two state-of-the-art GANs trained on two different datasets. For qualitative evaluation, we visualized feature maps of filters to show the consistency of the visual concept represented by each filter. We also visualized the results of modifying specific visual concepts on generated images. Besides, we demonstrated that performing Langevin dynamics could improve the realism of some bad generated images and modified images. For quantitative evaluation, we conduct a user study and a face verification experiment to examine the correctness of exchanging a specific visual concept and faces between pairs of images. We also calculated the mean squared-error (MSE) between original images and modified images in terms of a certain visual concept, in order to evaluate the locality of our modifications. We calculated the Fr echet Inception Distance (FID) (Heusel et al. 2017) to measure the realism of generated images. |
| Researcher Affiliation | Academia | 1Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 2Shanghai Jiao Tong University, China 3Zhejiang Laboratory, Hangzhou 311100, China |
| Pseudocode | No | The paper describes the optimization process and mathematical formulations but does not include structured pseudocode or an algorithm block. |
| Open Source Code | No | The paper does not provide a concrete access link to its source code or explicitly state that its code is being released. |
| Open Datasets | Yes | Big GAN was trained on FFHQ dataset (Karras, Laine, and Aila 2019). Style GAN (Karras, Laine, and Aila 2019) was trained on Celeb A-HQ dataset (Karras et al. 2018). |
| Dataset Splits | No | The paper mentions using FFHQ and Celeb A-HQ datasets for training but does not explicitly specify the training, validation, and test dataset splits needed for reproduction. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models used for running its experiments. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions or other libraries). |
| Experiment Setup | Yes | We set hyperparameters as C = 24, λ0 = 1 and λ1 = 2/3. Specifically, we set λ2 = 1 for updating the energybased model parameters W. For updating the generators of Big GAN and Style GAN, we set λ2 = 0.1 and λ2 = 0.05 respectively. For Style GAN, λ3 was set to be 3e-2 at first and exponentially decayed to 3e-6 during 1000 batches. For Big GAN, λ3 was set the same but exponentially decayed during 500 batches. We set T = 50 for Big GAN and T = 100 for Style GAN. We initialized each dimension of parameters W to be zero. We used the learning rate of 10-4 for the generator and discriminator in Big GAN and 10-3 for Style GAN. We used Adam optimizer with β1 = 0 and β2 = 0.9 for the generator and discriminator. We used SGD optimizer for parameters W. |