Learning Disentangled Representation by Exploiting Pretrained Generative Models: A Contrastive Learning View
Authors: Xuanchi Ren, Tao Yang, Yuwang Wang, Wenjun Zeng
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Dis Co on three major generative models (GAN, VAE, and Flow) on three popular disentanglement datasets. Dis Co achieves the state-of-the-art (SOTA) disentanglement performance compared to all the previous discovering-based methods and typical (VAE/Info GAN-based) methods. |
| Researcher Affiliation | Collaboration | Xuanchi Ren1 , Yang Tao2 , Yuwang Wang3 , Wenjun Zeng4 1HKUST, 2Xi an Jiaotong University, 3Microsoft Research Asia, 4EIT |
| Pseudocode | No | The paper describes the Dis Co framework and its workflow using text and a flowchart in Figure 2, but it does not include explicit pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | Source code is at https://github.com/xrenaa/Dis Co. |
| Open Datasets | Yes | We consider the following popular datasets in the disentanglement areas: Shapes3D (Kim & Mnih, 2018) with 6 ground truth factors, MPI3D (Gondal et al., 2019) with 7 ground truth factors, and Cars3D (Reed et al., 2015) with 3 ground truth factors. |
| Dataset Splits | No | The paper mentions using popular datasets and notes 'images are resized to the 64x64 resolution' and 'for all methods, we ensure there are 25 runs', along with 'batch size B to 32' in Appendix A.1. However, it does not explicitly specify the training/validation/test split percentages, sample counts for each split, or provide citations for predefined splits for reproducibility. |
| Hardware Specification | No | The paper mentions 'limited by GPU resources' in the context of Flow models, but it does not provide specific details about the hardware used for experiments, such as GPU models (e.g., NVIDIA A100), CPU types, or other computing specifications. |
| Software Dependencies | No | The paper mentions using an 'Adam optimizer' and 'open-source implementation' for Glow, but it does not specify version numbers for any software components, libraries, or programming languages (e.g., 'PyTorch 1.9', 'Python 3.8') that would enable reproducible software setup. |
| Experiment Setup | Yes | For the hyperparameters, we empirically set the temperature τ to 1, threshold T to 0.95, batch size B to 32, the number of positives N to 32, the number of negatives K to 64, the loss weight λ for Led to 1, the number of directions D to 64 and the dimension of the representation J to 32. We use an Adam optimizer (Kingma & Ba, 2015) in the training process, as shown in Table 5. Adam: beta1 0.9 Adam: beta2 0.999 Adam: epsilon 1.00e-08 Adam: learning rate 0.00001 Iteration: 100,000 |