Stability and Generalizability in SDE Diffusion Models with Measure-Preserving Dynamics
Authors: Weitong Zhang, Chengqi Zang, Liu Li, Sarah Cechnicka, Cheng Ouyang, Bernhard Kainz
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive experimental results corroborate the effectiveness of D3GM across multiple benchmarks including a prominent application for inverse problems, magnetic resonance imaging. |
| Researcher Affiliation | Academia | 1 Imperial College London, UK, 2 University of Tokyo, JP, 3 University of Oxford, UK 4Friedrich-Alexander University Erlangen-Nürnberg, GER |
| Pseudocode | No | The paper describes methods and processes in text and with equations but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The NeurIPS checklist answer for 'Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?' is '[No] Justification: We have provide all the details and instructions along with the paper and the appendix for the reproducible results. The datasets involved in the paper are public.' |
| Open Datasets | Yes | We utilized the fast MRI dataset [69], containing single-channel, complex-valued MRI samples. Implementation details can be found in Appx. G. IXI3 dataset is the largest benchmark considered in our MRI SR evaluation. |
| Dataset Splits | Yes | We applied 584 proton-density weighted knee MRI scans without fat suppression. These were subsequently partitioned into a training set (420 scans), a validation set (64 scans), and a testing set (100 scans). |
| Hardware Specification | Yes | Our models are trained on three RTX 6000 GPUs for about four days, each with 40GB of memory. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer', 'Squeeze-and-Excitation [38]', 'NAF [8]', 'spectral normalization (SN) [40]', and 'weight decay (WD) [35]', but it does not provide specific version numbers for any of these software dependencies. |
| Experiment Setup | Yes | For most of the experiments, the training patch-size is set to 128x128 with a batch size of 16. We utilize the Adam optimizer with β1 = 0.9, β2 = 0.999, a learning rate of 10^-4 with a decay strategy. Our models are trained on three RTX 6000 GPUs for about four days, each with 40GB of memory. Random seed is 42. All mathematical variants follow the cosine schedule as per [42]. The variance λ is set to 10 for the OU and stationary processes. |