Latent Traversals in Generative Models as Potential Flows
Authors: Yue Song, T. Anderson Keller, Nicu Sebe, Max Welling
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimentally, we demonstrate that our method achieves both more qualitatively and quantitatively disentangled trajectories than state-of-the-art baselines. Further, we demonstrate that our method can be integrated as a regularization term during training, thereby acting as an inductive bias towards the learning of structured representations, ultimately improving model likelihood on similarly structured data. Code is available at https://github.com/ King James Song/PDETraversal. |
| Researcher Affiliation | Collaboration | 1University of Trento, Italy 2University of Amsterdam, the Netherlands. Correspondence to: Yue Song <yue.song@unitn.it>. |
| Pseudocode | No | No pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | Yes | Code is available at https://github.com/ King James Song/PDETraversal. |
| Open Datasets | Yes | For experiments of pre-trained GANs, our method is evaluated on SNGAN (Miyato et al., 2018) with Anime Face (Chao, 2019), Big GAN (Brock et al., 2019) with Image Net (Deng et al., 2009), and Style-GAN2 (Karras et al., 2020) with FFHQ (Karras et al., 2019). [...] For the VAEs experiments, we use the VAE encoder as the auxiliary classifier and evaluate our method on MNIST (Le Cun, 1998) and d Sprites (Matthey et al., 2017) datasets. |
| Dataset Splits | Yes | We randomly select 10% of the images as the training set and the rest as the test set. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory amounts) used for running experiments were provided. |
| Software Dependencies | No | No specific software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9) were provided. |
| Experiment Setup | Yes | We set the total timestep T to 10 for all the datasets and models. In line with Tzelepis et al. (2021), the number of potential functions (traversal paths) K is set as 64 for SNGAN, 120 for Big GAN, and 200 for Style GAN2. The output images are of size 64x64 for SNGAN, of size 256x256 for Big GAN, and of size 1024x1024 for Style GAN2. [...] The learning rate is set to 0.005, and we train the network for 300 epochs with the batch size set as 32. |