2024-07-08

Added

• Dataset generation script now in data_gen/: (a) gns_data moved in there and (b) lagrangebench_data was added #30
• Better assertions and comments #31
• Note on missing determinism of GPU operations - see README.md #31

Fixed

• Docs #30
• model.init inputs are now also JAX arrays #31
• Fix working with external datasets #31

Changed

• Update JAX to latest 0.4.29 #30
• Remove JAX-MD from dependencies. Now neighbor search happens through JAX-SPH #30
• [BREAKING] cfg.dataset_path became cfg.dataset.src #31

---

Full Changelog: v0.2.0a1...v0.2.0

2024-03-04 - Neural SPH

Added

• Dataset generation notebook #24
• runner_test.py, check_cfg, LAGRANGEBENCH_DEFAULTS #28
• Model checkpoints #29

Fixed

• Rollout extrapolation steps #25

Changed

• Zenodo URL to datasets with force.py file #26
• Move experiments/* content to lagrangebench/runner.py and lagrangebench/defaults.py #28
• Gonfigs from argparse to omegaconf #28
• Trainer from function to class #28

---

Full Changelog: v0.1.2...v0.2.0a1

2024-01-11

Added

• Extended docs with (#17):
  • Reference to notebooks.
  • Baseline results from the NeurIPS 2023 paper.
• README, mainly #22:
  • LagrangeBench logo.
  • Clickable Badges with URLs to the paper, RTD, PyPI, Colab, and some git workflows.
  • Contribution guidelines.
  • Notes on MacOS and jax-metal, see #18.
• Tests, see #21
  • Our tests are written using unittest, but we run them with pytest. For now we keep that standard.
  • Currently, the tests cover roughly 70% of the codebase, namely:
    • the case_setup including preprocessing and integration modules,
    • whether the equivariant models are equivariant,
    • whether all 3 neighbor search backends give correct results on small edge cases,
    • the fushforward utils, and
    • the rollout loop by introducing a dummy 3D Lennard-Jones dataset of 3 particles for 2k steps.
• Github Workflows, mainly in #21:
  • Linting checks with ruff. Ruff now replaces black.
  • pytest under Python 3.9, 3.10, 3.11 including codecov.
  • Automatic publishing of tagged versions to PyPI.
• Batched rollout loop using vmap. Promises significant speedups, as validation during training used to take around 15%-30% of the time. And of course, batching during inference is nice to have. I noticed that there is an optimal batch size without changing the inference speed much, but there is a regime for larger batches where we don't get OOM, but validation becomes significantly slower. Tuning this batch_size_infer parameter with a few test runs is my current best advice. See #20 and #21.
• pkl2vtk to convert a pickle rollout to a series of .vtk files for visualization.
• Metadata and configs in pyproject.toml and other config files, see #21.

Fixed

• Multiple neighbor list reallocations during training, see #15.
• When using both random noise and pushforward, the noise seed is now independent of the max number of pushfoward steps, see #16.

Changed

• Remove explicit force functions from the codebase and put them in force.py Python files in the dataset directory of the datasets with forces (2D DAM, 2D RPF, 3D RPF). This comes along with a new version of the datasets on Zenodo here https://doi.org/10.5281/zenodo.10491868, see #23.
• Rename some variables and improve docstrings, see #17.
• Swap the order of sender and receiver to align with jax-md, see #17.
• Upgrade dependencies and fix jax==0.4.20, jax-md==0.2.8, and e3nn-jax==0.20.3.

---

Full Changelog: v0.0.2...v0.1.2

NeurIPS 2023 D&B

Code used to generate the results for the NeurIPS 2023 Datasets & Benchmarks paper.

Extensively tested functionalities (on Ubuntu 22.04 with Python 3.10.12 and Poetry 1.6.0):

• training/inference using config files - as described in README
• running the 3 notebooks

2023-08-20

First release of lagrangebench.