exalearn / ExaMol / 6266609784

98%

push

github

Introduce multi fidelity learning (#105)

* Use parallelism to get correct number of blocks

See parsl/parsl#1647

* Compute the solvation energy too

Also save to gzipped files. The size is starting to be notable

* Implement a multi-fidelity MPNN

Uses delta learning to predict the properties at intermediate levels

* No longer test for errors being thrown

* Only output the highest level of fidelity

Also some flake8 fixes

* Document how multi-fidelity trainign works

* Minor changes to the documentation

* Use a more robust relaxation technique (#108)

* Use MDMin to reduce to 0.1 eV/Ang, then BFGS

Still want to test this before we merge to main,
but fixes #106

* Use FIRE and a higher threshold for switching

* Use molecule which takes longer to optimize in test

* Use isnan and not isinf for detecting placeholders

* Switch to one scale layer per network

* Compute diff between adjance level, not from first

* Also fix how we compute inference

Delta between adjacent levels, not the beginning

Changed our test routine to ensure we do this right

* Initial training runs for multi-fidelity learning

* Update data loader test to handle new fixtures

* Train using subset of available data, test on all fidelities

* Minor bug: start decaying LR immediately

84 of 84 new or added lines in 4 files covered. (100.0%)

1785 of 1819 relevant lines covered (98.13%)

0.98 hits per line