Hamiltonian Score Matching and Generative Flows
Authors: Peter Holderrieth, Yilun Xu, Tommi Jaakkola
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 9 Experiments |
| Researcher Affiliation | Collaboration | Peter Holderrieth MIT CSAIL phold@mit.edu Yilun Xu NVIDIA yilunx@nvidia.com Tommi Jaakkola MIT CSAIL tommi@csail.mit.edu |
| Pseudocode | No | No explicit pseudocode or algorithm block found. Methods are described in prose and mathematical equations. |
| Open Source Code | No | Code can be provided upon request. |
| Open Datasets | Yes | Specifically, we train a Oscillation HGF on CIFAR-10 unconditional and conditional. FFHQ (unconditional)-64x64 |
| Dataset Splits | No | Appendix L lists training details like "We set the reference batch size to 516 on CIFAR-10 and 256 on FFHQ. We train for 200 million images in total", but does not explicitly state the train/validation/test data splits. |
| Hardware Specification | Yes | All the experiments are run on 8 NVIDIA A100 GPUs. |
| Software Dependencies | No | We used Py Torch as a library for automatic differentiation [38]. |
| Experiment Setup | Yes | We set the reference batch size to 516 on CIFAR-10 and 256 on FFHQ. We train for 200 million images in total, corresponding to approximately 3000 epochs and 48 hours of training time for CIFAR-10 and 96 hours for FFHQ. As outlined in the experiments section, the hyperparameters and training procedure are the same as [26]: namely, we used the Adam optimizer with learning rate 0.001, exponential moving average (EMA) with momentum 0.5, data augmentation pipeline adapted from [28], dropout probability of 0.13, and FP32 precision. For sampling, we use the 2nd order Heun s sampler [26]. |