Geometric Neural Diffusion Processes
Authors: Emile Mathieu, Vincent Dutordoir, Michael Hutchinson, Valentin De Bortoli, Yee Whye Teh, Richard Turner
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experimental results |
| Researcher Affiliation | Collaboration | Emile Mathieu University Of Cambridge Vincent Dutordoir University Of Cambridge Secondmind Labs Michael J. Hutchinson University Of Oxford Valentin De Bortoli Center for Science of Data, ENS Ulm Yee Whye Teh University Of Oxford Richard E. Turner University Of Cambridge, Microsoft Research |
| Pseudocode | Yes | Alg. 1 Conditional sampling with Langevin dynamics. |
| Open Source Code | Yes | The code is publicly available at https://github.com/cambridge-mlg/ neural_diffusion_processes. |
| Open Datasets | Yes | The data is drawn from the International Best Track Archive for Climate Stewardship (IBTr ACS) Project, Version 4 ((Knapp et al., 2010; Knapp et al., 2018)) |
| Dataset Splits | No | The paper specifies training and test data splits, but does not explicitly mention a separate validation set split: 'The training data consists of 2^14 sample paths while the test dataset has 2^12 paths.' |
| Hardware Specification | Yes | Models have been trained on A100-SXM-80GB GPUs. |
| Software Dependencies | No | Models, training and evaluation have been implemented in Jax (Bradbury et al., 2018). We used Python (Van Rossum and Drake Jr, 1995) for all programming, Hydra (Yadan, 2019), Numpy (Harris et al., 2020), Scipy (Virtanen et al., 2020), Matplotlib (Hunter, 2007), and Pandas (Mc Kinney et al., 2010). (Only JAX has an associated version number from its citation (0.2.5), but other key libraries like NumPy, SciPy, Matplotlib, Pandas, and e3nn are not given specific version numbers for reproducibility.) |
| Experiment Setup | Yes | In all experiment, we train the NDP-based models over 300 epochs using a batch size of 256. Furthermore, we use the Adam optimiser for training with the following learning rate schedule: linear warm-up for 10 epochs followed by a cosine decay until the end of training. |