Towards Building A Group-based Unsupervised Representation Disentanglement Framework

Authors: Tao Yang, Xuanchi Ren, Yuwang Wang, Wenjun Zeng, Nanning Zheng

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, we train 1800 models covering the most prominent VAE-based methods on five datasets to verify the effectiveness of our theoretical framework. Compared to the original VAE-based methods, these Groupified VAEs consistently achieve better mean performance with smaller variances.
Researcher Affiliation Collaboration Yang Tao1 , Xuanchi Ren2 , Yuwang Wang3 , Wenjun Zeng4 , Nanning Zheng1 1Xi’an Jiaotong University, 2HKUST, , 3Microsoft Research Asia, 4EIT
Pseudocode No No pseudocode or algorithm block was explicitly labeled or formatted as such.
Open Source Code No The paper references official implementations of *other* methods (e.g., Control VAE and RGr VAE) with GitHub links, but does not state that its *own* proposed method's source code is publicly available.
Open Datasets Yes To evaluate our method, we consider several datasets: d Sprites (Higgins et al., 2017), Shapes3D (Kim & Mnih, 2018), Cars3D (Reed et al., 2015), and the variants of d Sprites introduced by Locatello et al. (Locatello et al., 2019b): Color-d Sprites and Noisy-d Sprites.
Dataset Splits No The paper states that 'In all the experiments, we resize the images to 64x64 resolution' and lists datasets, but does not provide specific train/validation/test split percentages or counts.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or cloud instance specifications used for running experiments.
Software Dependencies No The paper mentions 'implemented by Pytorch Paszke et al. (2017)' but does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes We run using different hyperparameters and random seeds for every VAE-based model implemented by Pytorch Paszke et al. (2017). As shown in Table 4, for β-VAE, we assign 3 choices for β and 10 random seeds for both the Original and Groupified VAEs: 3x10x2 = 60 settings for each dataset. Similarly, we also assign 60 settings for Factor VAE and β-TCVAE. For Anneal VAE, we assign three choices for C and 3 choices for the start and end pair, also assign 10 random seeds. In summary, for all 5 datasets, we run (((3x10x2)x3) + 3x3x10x2)x5 = 1800 models. For other hyperparameters, please refer to Table 5 (b).