Restoration based Generative Models
Authors: Jaemoo Choi, Yesom Park, Myungjoo Kang
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our model improves the quality and efficiency of both training and inference. Furthermore, we show the applicability of our model to inverse problems. Our comprehensive empirical studies on image generation and inverse problems demonstrate that RGMs generate samples rivaling the quality of DDMs with several orders of magnitude faster inference. |
| Researcher Affiliation | Academia | 1Department of Mathematical Sciences, Seoul National University, Seoul, South Korea. |
| Pseudocode | Yes | Algorithm 1 Training of RGMs with Posterior sampling, Algorithm 2 Relaxed training algorithm of RGMs, Algorithm 3 Sampling Procedure of RGMs |
| Open Source Code | Yes | The code is available at https://github.com/Jae-Moo/RGM/. |
| Open Datasets | Yes | We focus on the widely used CIFAR10 unconditional image generation benchmark (Krizhevsky et al., 2009) and also validate the performance of RGMs on largescale (256 256) images: Celeb A-HQ (Liu et al., 2015) and LSUN Church (Yu et al., 2015). |
| Dataset Splits | No | The paper refers to widely used benchmark datasets like CIFAR10, Celeb A-HQ, and LSUN Church, for which splits are standard. However, it does not explicitly state the training, validation, or test split percentages or sample counts within the text. |
| Hardware Specification | Yes | We train our models on CIFAR-10 using 4 V100 GPUs. |
| Software Dependencies | No | The paper mentions using Adam optimizer and building code on DDGAN, but it does not provide specific version numbers for software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | To optimize our RGMs, we mostly followed the previous literature (Xiao et al., 2021a), including network architectures, R1 regularization, and optimizer settings. We use a learning rate of 2 10 4 for generator update in all experiments and a learning rate of 10 4 for discriminator update. We use λ 1 = 10 3 for image size of 32, and λ 1 = 5 10 5 for image size of 256. The models are trained with Adam (Kingma & Ba, 2014) in all experiments. In CIFAR10 experiments, we train RGM-KLD-D and RGM-KLD-SR (naive) for 200K iterations and RGM-KLD-SR for 230K iterations. |