Truncated Diffusion Probabilistic Models and Diffusion-based Adversarial Auto-Encoders

Authors: Huangjie Zheng, Pengcheng He, Weizhu Chen, Mingyuan Zhou

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 EXPERIMENTS We aim to demonstrate that TDPM can generate good samples faster by using fewer steps of reverse diffusion. We use different image datasets to test our method and follow the same setting as other diffusion models (Ho et al., 2020; Nichol & Dhariwal, 2021; Dhariwal & Nichol, 2021; Rombach et al., 2022) for our backbones. We also have two ways to set up the implicit generator that starts the reverse diffusion.
Researcher Affiliation Collaboration Huangjie Zheng1,2 & Pengcheng He2 & Weizhu Chen2 & Mingyuan Zhou1 The University of Texas at Austin1, Microsoft Azure AI2
Pseudocode Yes The process of training and sampling of these configurations are summarized in Algorithm 1 and 2.
Open Source Code Yes For the implementation details, please refer to Appendix D.6 and our code at https://github.com/Jeg Zheng/ truncated-diffusion-probabilistic-models.
Open Datasets Yes We use CIFAR-10 (Krizhevsky et al., 2009), LSUN-bedroom, and LSUN-Church (Yu et al., 2015) datasets in unconditional experiments, and CUB-200 (Welinder et al., 2010) and MS-COCO (Lin et al., 2014) for text-to-image experiments.
Dataset Splits No The paper mentions using well-known datasets like CIFAR-10 and MS-COCO, but does not explicitly state the specific training, validation, and test dataset splits (e.g., percentages or sample counts) used for these experiments.
Hardware Specification Yes We report both FID and the sampling time (s/image) on one NVIDIA V100 GPU in Figure 4. We train our models using V100 GPUs... The GPU time of sampling (s/image) is measured on one NVIDIA A100.
Software Dependencies Yes We train our models using V100 GPUs, with CUDA 10.1, Py Torch 1.7.1.
Experiment Setup Yes Optimization: We train our models using the Adam optimizer (Kingma & Ba, 2015), where most of the hyperparameters match the setting in Xiao et al. (2022), and we slightly modify the generator learning rate to match the setting in Ho et al. (2020), as shown in Table 8. Table 8: Optimization hyper-parameters. (lists specific hyperparameters like learning rates, batch size, iterations)