v0.10.3

[0.10.3]

Matbench
Our new model SevenNet-l3i5 is now available.

Added

• SevenNet-l3i5, checkpoint, preset. (keywords: 7net-l3i5, sevennet-l3i5)

Changed

• Fast --help (do not import torch)
• README refined

What's Changed

• Add: sevennet-l3i5 by @YutackPark in #136
• Upload l3i5 & update README.md by @JisuJung928 in #137
• Fast --help by @YutackPark in #139
• Sevennet-l3i5 unit test by @YutackPark in #141
• Fix: add l3i5 cp as data by @YutackPark in #142

New Contributors

• @JisuJung928 made their first contribution in #137

Full Changelog: v0.10.2...v0.10.3

v0.10.2

[0.10.2]

Added

• Accelerated graph build routine if matscipy is installed @hexagonerose
• matscipy vs. ase neighborlist unit test
• If valid set is not given but data_divide_ratio is given, validaset is created using random split. (shift, scale, and conv_denoiminator uses original whole statistics)

Changed

• matscipy is included as a dependency
• data_divide_ration defaults to 0.0 (meaning not used)

Fixed

• For torch version >= 2.4.0, Loading graph dataset no more raises warnings.
• Raise error when unknown element is found (SevenNetCalculator)

What's Changed

• [Feat] change ase.neighborlist to from matscipy.neighbours by @hexagonrose in #124
• [refactor] make _remove_self_edges be skipped when doing matscipy graph build by @hexagonrose in #129
• docs: update sevenn_graph_build of README.md by @YutackPark in #128
• [Feat] add graph build test of ase and matscipy in pytest by @hexagonrose in #131
• Add: data split by @YutackPark in #134

New Contributors

• @hexagonrose made their first contribution in #124

Full Changelog: v0.10.1...v0.10.2

v0.10.1

[0.10.1]

Added

• experimental SevenNetAtomsDataset which is memory efficient, can be enabled with dataset_type='atoms'
• Save meta data & statistics when the SevenNetGraphDataset saves its data.

Changed

• Save checkpoint_0.pth (model before any training)
• SevenNetGraphDataset._file_to_graph_list -> SevenNetGraphDataset.file_to_graph_list
• Refactoring SevenNetGraphDataset, skips computing statistics if it is loaded, more detailed logging
• Prefer use .get when accessing config dict

Fixed

• Fix error when loading SevenNetGraphDataset with other types of data (ex: extxyz) in one dataset

Full Changelog: v0.10.0...v0.10.1

What's Changed

• chore: update dependencies (PyG >= 2.5.0) by @YutackPark in #110
• v0.10.1 by @YutackPark in #112
• docs: update torch version by @YutackPark in #116
• refactor: file not found error + warnings for graph dataset by @YutackPark in #119
• merge main by @YutackPark in #122
• refactor: please use get by @YutackPark in #121

v0.10.0

[0.10.0]

SevenNet now has CI workflows and its python coverage is 78%!
Substantial changes in command-line apps and their outputs.
Previous functionalities are still accessible with correct flags (sevenn -m train_v1 ... , sevennet_graph_build --legacy ...)

Added

• sevenn_preset for below changes
• [train_v2]: train_v2, with lots of refactoring, support load_testset_path.
• [train_v2]: SevenNetGraphDataset replaces old AtomGrpahDataset, which extends InMemoryDataset of PyG.
• [train_v2]: sevenn_graph_build for SevenNetGraphDataset.
• [train_v2]: Any additional datasets will be evaluated and recorded if it is given as 'load_{NAME}set_path' key.
• Univ keyword for chemical_species
• More options: energy_key, force_key, stress_key for sevenn_graph_build, than ase.calculator @thangckt
• OpenMPI distributed training @thangckt

Changed

• Some cumbersome things(type_map, requires_grad ), that hurt readability, are now hidden inside AtomGraphSequential.
• log.sevenn and lc.csv automatically find a safe filename (log0.sevenn, log1.sevenn, ...) to avoid overwriting.
• [train_v2]: train_v2 loads its training set via load_trainset_path, rather than previous load_dataset_path.
• [train_v2]: log.csv -> lc.csv, and columns have no units, (easier to postprocess with it) but still on log.sevenn.
• [train_v2]: Train valid split by ratio is temporarily removed
• Previous presets are renamed with _v1 suffix

Fixed

• [e3gnn_serial]: can continue simulation even when atom tag becomes not consecutive (removing atom dynamically), @gasplant64
• [e3gnn_parallel]: undefined behavior when there is no atoms to send/recv (for non pbc system)
• [e3gnn_parallel]: incorrect force/stress in some edge cases (too small simulation cell & 2 process)
• [e3gnn_parallel]: revert commit 14851ef, now e3gnn_parallel is sane.
• [e3gnn_*]: += instead of = when saving virial stress and forces @gasplant64
• Now Logger correctly closes a file.
• ... and lots of small bugs I found during writing pytest.

New Contributors

• @thangckt made their first contribution in #89
• @gasplant64 made their first contribution in #81
• @pre-commit-ci made their first contribution in #101

v0.9.3

Warning: This release has a serious bug in LAMMPS parallel. Please use v0.9.3.post1 in the 'tags', or use pip install.

SevenNet is released to PyPI, project is managed using pyproject.toml, not setup.py

• PyPI release + docs
• torch_scatter dependency removed
• torch_geometric dependency added to the pyproject.toml

Features:

• sevenn_preset for basic training yaml files
• sevenn_patch_lammps for easier installation
• keyword input for pre-trained models: continue, sevenn_get_model, sevennet_calculator

Bugfix:

• C++17 for torch version > 2.0

Known bug:

• In training yaml, scale: 'elemwise_force_rms' is broken. Fixed in github, but not in pip.

The last snapshot of XPLOR cutoff bug

The last snapshot of SevenNet has XPLOR cutoff bug and wrong SevenNet-0 potential.

v0.9.1

This is checkpoint before updating pair_e3gnn_parallel.cpp

I have fixed bugs for small simulation cell but nswap == 6

This version could gives wrong results in the special case where

1. cell is very small but still nswap == 6 (6 communication for each cubic faces).
2. There exists a small overlap of cutoffs between process 2 and its PBC replicated cell.

The error is fixed but the version is not thoroughly debugged by others.

v0.9.0

Temporal release after merge from dev.

Stress calculation on serial version implemented.

sevenn_get_model
sevenn_inference
sevenn_graph_build

commend updated, README.md updated.

v0.8.1

Average num neighbor fixed.

This is backup for following merge from branch dev

First release v0.8.0

Minimal documentation for usage and installation is prepared. But still bugs are present.

Know bugs:

• When parsing VASP OUTCARs with structure_list, if the folder contains a POSCAR with selective dynamics, it does not read the OUTCAR correctly.
• When parsing VASP OUTCARs with structure_list, spin polarized calculations are not yet supported.
• Models with a large number of parameters (>=5 message passing layers, >= 64 channels) show initial RMSEs of energy and force significantly larger than the original nequip. Since the difference should be marginal, we are actively investigating this issue.
• The calculated stress on LAMMPS is incorrect.
• Functionality 'continue' is unstable.