search

openmm / spice-models

Models trained on the SPICE dataset
MIT License
10 stars 0 forks source link

Cannot try due large forces #2

Open raimis opened 2 years ago

raimis commented 2 years ago

According to the description, SPICE.hdf5 is downloaded from https://github.com/openmm/spice-dataset/releases: https://github.com/openmm/spice-models/blob/b9023c7955c6e8f3f3fa9b73f6a4d5e9b3280990/five-et/README.md?plain=1#L3-L4

The file contains samples with large force (https://github.com/openmm/spice-dataset/issues/32) and createSpiceDataset.py doesn't filter them.

When I try to train the loss just fluctuates ~1e7 without any improvement.

raimis commented 2 years ago

@peastman could you clarify the force cutoff? And how did you apply it?

peastman commented 2 years ago

The filtering was implemented in https://github.com/openmm/spice-dataset/pull/41, which was merged after that release was created. Once we're agreed on all remaining changes (especially anything for https://github.com/openmm/spice-dataset/issues/42), I'll create a 1.1 release.

raimis commented 2 years ago

@peastman the cutoff in https://github.com/openmm/spice-dataset/pull/41 is 1 Hartree/Bohr (≈ 49132 kJ/mol/nm), while in the email you mentioned 20000 kJ/mol/nm. Which ones was used?

peastman commented 2 years ago

My original model described in the first draft of the paper used 20,000 kJ/mol/nm. The five newer models described in the current draft used 1 hartree/bohr.