… to 79 character lines

oops, my IDE had been on 120 character lines this whole time -- switching to 79 character lines so future black passes don't turn things into quadruply nested messes

…with a graph-net in the loop

it's fishy that the initial hydration free energy prediction is so poor

i suspect i may have made a unit mistake in my numpy/jax --> pytorch port

… PyTorch GBSA

…another function that converts to espaloma unit system

…ssertions

Thanks to @yuanqing-wang for carefully stepping through this with me

Co-Authored-By: Yuanqing Wang <[email protected]>

Obtained from https://raw.githubusercontent.com/MobleyLab/FreeSolv/ff0961a3177863c8002e8487ff9718c513974138/database.txt

…et using parsley 1.2

the line to compute the reduced work was written as if "solv_energies" was "valence_energies + gbsa_energies" but of course it was just "gbsa_energies"...

in column `quick_xyz` -- will shortly replace this with a column `xyz` with more thorough parsley 1.2 vacuum sampling

… negative delta g predictions

thanks to @jchodera for spotting this! #54 (comment)

Co-Authored-By: John Chodera <[email protected]>

* confirmed that it can overfit to a small subset of FreeSolv! 🎉
* RMSE on whole of freesolv hasn't yet matched the quality of OBC2 🙏

…ze: validation set RMSE 1.8kcal.mol

* increased stepsize 1e-3 rather than 1e-4
* decreased layer and node dimensions from 128 to 32

+ a few formatting and documentation enhancements

* define random seed once at top of file rather than before each step
* remove verbose flag
* use same learning rate, n_iterations, n_mols_per_batch, n_snapshots_per_mol for both trajectories

unfortunatlely, at least one prediction on the validation-set appears to be nan (so the overall validation-set rmse is reported as nan)

element    # of molecules containing it
C          639
H          629
O          344
N          169
Cl         114
S          40
F          35
Br         25
P          14
I          12

* add training / validation curves for early-stopping
* add bootstrapped rmses to final scatterplots

run 10ns of md per molecule (rather than the measly 0.01ns per molecule in 5866029 )

validation-set rmse ~1.2-1.5 kcal/mol

using more thorough vacuum md, specified here 8e50eec

anecdotally, this appears to increase the training-set vs. validation-set error gap, suggesting that insufficient equilibrium sampling might have made the validation-set performance reported in #54 (comment) look more favorable than it should!

…alculations vs. experiment on the {C, H, O} subset

Observations:
* n=10 overfitting seems to achieve a lower error than previously
* 50/50 train/validate seems to initialize and optimize at a higher error than in first version