-
get-results-from-wandb.py
: downloads the results from W&B, the output is saved inresults.txt
.-
results.txt
contains 3DReact's results per run (run id, run name, seed, MAE, RMSE) and averaged across 10 folds (run name, mean MAE, MAE stdev, mean RMSE, RMSE stdev).
-
-
get-tables.py
takes the errors fromresults.txt
,../baseline_chemprop/results/*/fold_?/test_scores.csv
,../baseline_slatm/results/slatm_10_fold_*.npy
and produces the$\mathrm{\LaTeX}$ tables for the paper and data for gnuplot scripts. -
checkpoints/
: several checkpoint files of 3DReact used to generate the figures in the paper.- use
../evaluation.py
and../representation.py
to compute the errors by reaction and learned representation, respectively. -
log2dat.bash
: script to extract the errors by reaction from the output ofevaluation.py
. A similar script could be used to extract the learned representation fro the output ofrepresentation.py
. -
by_reaction/
: predictions for each text reaction: reaction index (in the dataset), target, prediction (in the dataset stdev units).
- use
Scripts used to generate data for some of the Figures for the paper (see the READMEs inside):
repr/
— for the t-SNE maps.gdb-reaction-classes/
— for the error distributions per reaction class.cyclo-rmsd/
— for the geometry sensitivity in the case of the Cyclo-23-TS set.