For better understand DeePMD-kit, we implement its partial features using PyTorch and expose interface consuing descriptors.

Technical details of our custom torch.autograd.Function is explained at DeePMD 描述符 se_a 前向和反向.

To ensure result consistency between DeePMD-kit and our PyTorch version, 7 unit tests are implemented.

Execute them one by one.

cd tests
find . -name '*.py' | xargs -I {} python3 -u {}

To verify RMSE decrease during training, a demo INPUT config is prepared.

ln -s tests/water/data data
python3 -u deepmd_pt/main.py tests/water/se_e2_a.json >train.log 2>&1 &
tail -f lcurve.log  # RMSE values of energy and loss are collected

Furthermore, we can draw 2D-line diagram based on the lcurve.out file.

python3 -u visualize.py  # Output is `rmse_over_step.png`

Possible graph could be: