Note: AGAT after this version (included) cannot load the well-trained model before. If you need to do so, please use v8.0.5: https://pypi.org/project/agat/8.0.5/
- Fix bugs when traing model with voigt stress tensor.
- Add a node to edge layer: agat/model/model.py.
- Message passing: agat/model/model.py.
- Fix a bug when saving model agat\lib\model_lib.py.
- Fix a bug when training model agat\model\fit.py.
- Modify agat/lib/model_lib.py#L178-L193
- Add default parameter:
vasp_bash_path
high_throughput_dft_calculation.py#L71; default_parameters.py#L242. - Modify
run_vasp()
function: high_throughput_lib.py#L124-L149. - Add transfer learning: default_parameters.py#L97. agat/model/fit.py#L169-L174
- Add split graphs: agat/data/build_dataset.py#L795-L824
- Convert TensorFlow to PyTorch backend.
- Updata docs.
- Add API for controling HP DFT calculation. agat/default_parameters.py
- Add
mask_reversed_magnetic_moments
in agat/default_parameters.py and agat/data/data.py - Modify agat/data/data.py:
- Include stress in the graph: agat/data/data.py#L273-L275, agat/data/data.py#L350-L352.
- Update method of parsing the vasp data: agat/data/data.py#L610, agat/data/data.py#L625-L656, agat/data/data.py#L661-L675.
- Update docs.
- Add agat/app/cata/high_throughput_dft_calculation.py.
- Shift atomic positions before fix bottom atoms: agat/app/cata/high_throughput_predict.py#L225-L227
- Add
default_hp_dft_config
to agat/default_parameters.py#L139-L246. - Add agat/lib/HighThroughputLib.py.
- Add agat/lib/ModifyINCAR.py.
- Upgrade docs.
-
Using self-defined tf-based functions to calculate Pearson r: agat/lib/GatLib.py#L248-L259
This self-defined function can handle
ValueError: array must not contain infs or NaNs
. -
Fix a bug: bug
-
Clip optimizer grads: clipnorm=1.0
-
Fix bugs in high-throughput predict:
-
Deprecate redundant training configurations:
train_energy_model
: agat/model/ModelFit.py and agat/model/ModelFit.pytrain_force_model
: agat/model/ModelFit.py and agat/model/ModelFit.pynew_energy_train
new_force_train
load_graphs_on_gpu
- Fix a bug here: agat/model/ModelFit.py
- Load well-trained models: agat/model/GatEnergyModel.py and agat/model/GatForceModel.py
- Test with best model after training. agat/model/ModelFit.py and agat/model/ModelFit.py.
- Raise exception if error occurs when parsing OUTCAR file. agat/data/data.py
- Remove
os
from the root name space. agat/init.py - Fix a bug when build graphs. See agat/data/data.py and agat/data/data.py. Specifically, one needs to cast
tf.tensor
asnp.array
before building graph properties with a very large tensor. agat/data/data.py. - Debug at these lines of agat/data/data.py: L553 and L585-L588.
- Using relative import. For example: agat/init.py
- Update documentations.
- Import useful objects only. For example: agat/app__init__.py
- Return test MAE after training. agat/model/ModelFit.py and agat/model/ModelFit.py
- Import
pymatgen
module when necessary. See agat/data/AtomicFeatures.py. This feature was changed back. - Specify device when building graphs. See agat/app/GatApp.py, agat/data/data.py
- Add default gpu specification when building database. agat/default_parameters.py
- Attache distributions at dist.
- Release pip wheel.
- Simplify packages. See v1.0.0 for more details of the first release.
First release to reproduce results and support conclusions of Design High-Entropy Electrocatalyst via Interpretable Deep Graph Attention Learning.