1.1.2
版本发布时间: 2022-08-10 09:58:08
mir-group/flare最新发布版本:1.3.3(2022-12-20 05:00:27)
What's Changed
Full Changelog: https://github.com/mir-group/flare/compare/1.0.0...1.1.2
Offline training
- Compute and print mean absolute error at each step (frame), not only at steps where a frame is added to SGP.
- Add data distribution statistics at the end of the training, summarizing how many frames are picked up from each dataset
Bug fixing
- Fix a lammps calculator bug of prism cell
- Fix the on-the-fly MAE calculation and logging when any of energy/forces/stress is excluded for training.
- Allow DFT calculator to be saved as json in the checkpoint, in case some can be non-pickable
- Add atom_indices into the dict of dumped SGP
Other features and tutorials
- Add timer for each part in OTF training log
- Add mapped uncertainty to the
build_map
method inSGP_Calculator
- Add python interface to support customized descriptors. Check the tutorial here
- Add a tutorial for computing thermal conductivity from flare + Phoebe. Check the tutorial here
Backward compatibility notices
- In the yaml file for offline training, the parameters in
FakeDFT
can be removed since they are redundant
dft_calc:
name: FakeDFT
kwargs:
filename: fake_dft.xyz
format: extxyz
index: ":"
io_kwargs: {}
change to
dft_calc:
name: FakeDFT
kwargs: {}
params: {}
- In the yaml file for offline training, the
filename
changes tofilenames
, allowing user to list multiple files as datasets.
otf:
md_engine: Fake
md_kwargs:
filename: fake_dft.xyz
change to
otf:
md_engine: Fake
md_kwargs:
filenames: [fake_dft_1.xyz, fake_dft_2.xyz]
Then at the end of the training, the log file will report how many frames are selected to add to SGP from fake_dft_1.xyz
, fake_dft_2.xyz
, respectively.
- In the yaml file for online or offline training, set up a lower bound of DFT calls for training hyperparameters
otf:
freeze_hyps: 10
change to
otf:
train_hyps: [1, 10] # [a, b] represents training hyps when the number of DFT frames SGP has collected is between a and b