MS$^3$D: A RG Flow-Based Regularization for GAN Training with Limited Data
Authors: Jian Wang, Xin Lan, Yuxin Tian, Jiancheng Lv
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments utilize four diverse datasets: Oxford Dog (from Oxford-IIIT pet dataset (Parkhi et al., 2012), detailed in Appendix C), Flickr-Faces-HQ (FFHQ) (Karras et al., 2019), Met Faces (Karras et al., 2020a), and Bre Ca HAD (Aksac et al., 2019). Detailed descriptions of these datasets are provided in Appendix C. Metrics. We employ Inception Score (IS) (Salimans et al., 2016), Fr echet Inception Distance (FID) (Heusel et al., 2017), and Kernel Inception Distance (KID) (Binkowski et al., 2018) to evaluate our models. We utilize the official implementations of these metrics as provided by Karras et al. (2020b). In Table 1, we compare our method with multiple GAN losses. In Table 2, we show the comparison with other regularization methods. Furthermore, we compare our method across different GAN architectures. We also conducted experiments on small datasets. |
| Researcher Affiliation | Academia | College of Computer Science, Sichuan University and Engineering Research Center of Machine Learning and Industry Intelligence, Ministry of Education, Chengdu 610065, P.R. China. |
| Pseudocode | Yes | In Appendix B, we present the Py Torch-like pseudo-code of MS3D calculation. Figure 13. Pseudocode of the proposed multi-scale structural self-dissimilarity (MS3D) regularization. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing open-source code or a link to a code repository. It only mentions providing pseudocode and that the method can be implemented using existing deep learning frameworks. |
| Open Datasets | Yes | Our experiments utilize four diverse datasets: Oxford Dog (from Oxford-IIIT pet dataset (Parkhi et al., 2012), detailed in Appendix C), Flickr-Faces-HQ (FFHQ) (Karras et al., 2019), Met Faces (Karras et al., 2020a), and Bre Ca HAD (Aksac et al., 2019). |
| Dataset Splits | No | The paper explicitly mentions training and test set splits ('randomly select 4,492 images for training and use the remaining 498 images as the test set') but does not provide specific details for a validation set split for reproducibility, other than implicitly mentioning 'validation data' in figures. |
| Hardware Specification | Yes | Table 7 summarizes the computational overhead and memory usage during training on Oxford Dog with Nvidia RTX 4090, indicating negligible resource demands. |
| Software Dependencies | No | The paper mentions 'popular deep learning frameworks like Py Torch (Paszke et al., 2019) and Tensor Flow (Abadi et al., 2016)' but does not provide specific version numbers for these software dependencies, which are necessary for full reproducibility. |
| Experiment Setup | Yes | For integrating MS3D with other methods, we set λ to 10 in Style GAN2, while it is set to 100 in other methods due to Style GAN2 s additional constraint terms. In Eq. (8), we chose a coarse-graining factor of 2, a common practice, and varied this factor in numerical experiments on Oxford Dog (Fig. 9(c)). All experiments use the same hyperparameter settings, and model performance is evaluated using the implementation from Karras et al. (2020b). |