VAEBM: A Symbiosis between Variational Autoencoders and Energy-based Models
Authors: Zhisheng Xiao, Karsten Kreis, Jan Kautz, Arash Vahdat
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate our proposed VAEBM through comprehensive experiments. Specifically, we benchmark sample quality in Sec. 5.1, provide detailed ablation studies on training techniques in Sec. 5.2, and study mode coverage of our model and test for spurious modes in Sec. 5.3. We choose NVAE (Vahdat & Kautz, 2020) as our VAE, which we pre-train, and use a simple Res Net as energy function Eψ, similar to Du & Mordatch (2019). |
| Researcher Affiliation | Collaboration | Zhisheng Xiao Computational and Applied Mathematics The University of Chicago zxiao@uchicago.edu Karsten Kreis, Jan Kautz, Arash Vahdat NVIDIA {kkreis,jkautz,avahdat}@nvidia.com |
| Pseudocode | No | The paper describes methods using mathematical formulations and text, but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions using the official implementation of NVAE at 'https://github.com/NVlabs/NVAE' and official implementations of FID and IS, but does not provide concrete access to the source code for their proposed VAEBM methodology. |
| Open Datasets | Yes | Experimental results show that our model outperforms previous EBMs and state-of-the-art VAEs on image generation benchmarks including CIFAR-10, Celeb A 64, LSUN Church 64, and Celeb A HQ 256 by a large margin, reducing the gap with GANs. |
| Dataset Splits | No | The paper mentions using 'CIFAR-10 train set and test set images' and 'training images' for FID computation, implying standard splits for publicly available datasets, but does not explicitly detail the training/validation/test split percentages or sample counts for all datasets used, nor does it explicitly mention a dedicated 'validation' set. |
| Hardware Specification | Yes | We run the experiments on a computer with a Titan RTX GPU. |
| Software Dependencies | Yes | We use Py Torch 1.5.0 and CUDA 10.2. |
| Experiment Setup | Yes | We summarize some key hyper-parameters we used to train VAEBM in Table 8. On all datasets, we train VAEBM using the Adam optimizer (Kingma & Ba, 2015) and weight decay 3e 5. We use constant learning rates, shown in Table 8. Following Du & Mordatch (2019), we clip training gradients that are more than 3 standard deviations from the 2nd-order Adam parameters. |