Releases: CederGroupHub/chgnet
v0.3.6
What's Changed
🛠 Enhancements
- Fixing VASP parsing with no MAGMOM by @BowenD-UCB in #148
🚀 Performance
- Allowing avoiding sorting cuda memory, enhancing compatibility working on HPC by @BowenD-UCB in dd0dd07
Full Changelog: v0.3.5...v0.3.6
v0.3.5
What's Changed
🛠 Enhancements
🚀 Performance
🏥 Package Health
🤷♂️ Other Changes
Full Changelog: v0.3.4...v0.3.5
v0.3.4
What's Changed
🛠 Enhancements
- Better backward compatibility with
ase
versions - Better backward compatibility with
pymatgen
versions - Allowing
loss_ratio=0
forTrainer
🐛 Bug Fixes
- Add
pip install git+https://gitlab.com/ase/ase
user advice onFrechetCellFilter
ImportError
and allowase_filter
to bestr
by @janosh in #104 - Fix
solve_charge_by_mag()
using wrong key instructure.site_properties.get('final_magmom')
by @janosh in #114
🧹 House-Keeping
🏥 Package Health
🤷♂️ Other Changes
- Add keyword
use_device
toCHGNet.load()
by @tsihyoung in #105 - Update dynamics.py by @zhongpc in #109
cuda_devices_sorted_by_free_mem()
return [] ifnot torch.cuda.is_available()
by @janosh in #115
New Contributors
Full Changelog: v0.3.3...v0.3.4
v0.3.3
What's Changed
🐛 Bug Fixes
If you encounter ase.filters
import issue, please install the newest ase from their source before the next ase release:
pip install git+https://gitlab.com/ase/ase
🛠 Enhancements
Full Changelog: v0.3.2...v0.3.3
v0.3.2
Changes
- Link to Video tutorial @BowenD-UCB
- Allow setting MD start temperature @BowenD-UCB
- Fixed bug triggerred by dtype @janosh in #95
Full Changelog: V0.3.1...v0.3.2
v0.3.1
v0.3.0
0.3.0 version: Improved pretrained weights released
We release the most recent pretrained model: CHGNet 0.3.0 🎉 (see details)
CHGNet.load()
now defaults to '0.3.0' version.
Major changes:
- Increased AtomGraph cutoff to 6A #55
- Resolved discontinuity issue when no BondGraph presents #79
- Added some normalization layers
- Slight improvements on energy, force, stress accuracies
Full Changelog: v0.2.2...v0.3.0
v0.2.2
Important Notes
v0.2.2 has a breaking change in model initialization from scratch i.e. chgnet = CHGNet()
, to address #79.
The developer team is running another round of pretraining and the updated weights will be released in the next release.
For now, chgnet = CHGNet.load()
loads same pretrained model as previous releases.
If user wants to load their own fine-tuned weights trained from previous releases, please do
chgnet = CHGNet.from_file("your_weights_path", mlp_out_bias=True)
What's Changed
- Run
cibuildwheel
on Apple silicon to publish pre-compiled macOS ARM binaries by @janosh in #78 - always report cov by @lbluque in #77
- 1.Dropped torch.det for MPS compatibility 2.better testing for rotation/supercell invariance by @BowenD-UCB in
4cee9e6
- Added function to output site energies by @BowenD-UCB in
77845dc
- Added examples for parsing DFT by @BowenD-UCB in
2ebc57f
- Fixed bug in #79, while not breaking loading pretrained weights by @BowenD-UCB in
1a558ea
- CHGNet now supports Apple MPS:tada:, pytorch-MPS has been tested with torch-2.0.1 by @BowenD-UCB in
516c422
- added num_isolated_atoms property by @BowenD-UCB in
20ebc58
- allow specifying decay fraction in CosineAnnealing LR scheduler by @BowenD-UCB in
dbb0305
- Moved all torch.int64 to torch.int32 for MPS support by @BowenD-UCB in
828478a
- added notes to avoid Colab bug by @BowenD-UCB in
a760ec8
- added installation from source by @BowenD-UCB in
030d897
- removed print message by @BowenD-UCB in
38a5cc4
- minor fix by @BowenD-UCB in
465f833
- fixed test by @BowenD-UCB in
d5ba24b
- loosen numerical tolerance by @BowenD-UCB in
b59d301
- added gMLP_norm option if gMLP is used at read_out by @BowenD-UCB in
195d48c
Full Changelog: v0.2.1...v0.2.2
v0.2.1
More Supports for MD: LAMMPS and ASE-MD ⚡
Main updates in this release:
- Implemented support for Nose-Hoover thermostat with ASE MD, choice of thermostat is now an option. @BowenD-UCB @tsihyoung
- MD and structure optimization now defaults to CUDA with most available memory. Device selection is mostly automatic now. @BowenD-UCB
- MD now has option to generate hidden space features, allowing opportunity for post analysis. @AegisIK @BowenD-UCB
- LAMMPS interface has been implemented, more tests to go. #57 @advancesoftcorp
- NVE MD is implemented with ASE-MD. @BowenD-UCB
- Isolated atom now defaults to warning in MD and relaxer (previously raised an error). Makes MD more stable @BowenD-UCB
- VASP parser now checks electronic convergence at every ionic step. @BowenD-UCB
- Updated paper reference and citation. @janosh
- tests and minor bug-fix. @janosh @BowenD-UCB
- Publish pre-compiled wheels for all platforms to PyPI. @janosh
New Contributors
Full Changelog: v0.2.0...v0.2.1
v0.2.0
CHGNet is much faster now 🚀 🚀 🚀
Main updates in this release:
-
Implement fast Cython graph converter in #40 by @AegisIK, yielding a 2 - 3x speed up in graph conversion depending on system size. Molecular dynamics now runs 1.5 - 2 x faster as well!
-
CHGNet API docs went live at https://chgnet.lbl.gov 🎉 in #39 by @janosh
-
Better tests and docstrings by @BowenD-UCB
All PRs since previous release
- API Docs by @janosh in #39
- Fix
ValueError: ndarray is not C-contiguous
CI errors by @janosh in #41 - 2.4x avg speedup on PMG struct to CHGNET graph conversion by @AegisIK in #40
- More descriptive names for single-letter variables by @janosh in #42
- Fix
ruff
PERF203
+PERF401
violations by @janosh in #43 - FIXED MEMORY LEAK in fast graph converter. Also cleaned up unneeded code. by @AegisIK in #44
- Add
CHGNet.todict
method needed for ASE JSON serialization by @janosh in #49 - Added garbage collection cyclic references turnoff for fast graph converter by @AegisIK in #50
- Solved final memory leak in fast converter by @AegisIK in #51
- Use
MaterialsProject2020Compatibility
to correct final energy infine_tuning.ipynb
by @janosh in #46
New Contributors
Full Changelog: v0.1.04...v0.2.0