results

Results

• get-results-from-wandb.py: downloads the results from W&B, the output is saved in results.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 from results.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.

  • auto-generated-tables.dat
  • auto-generated-data-for-extrapolation-plot/

• 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 of evaluation.py. A similar script could be used to extract the learned representation fro the output of representation.py.
  • by_reaction/: predictions for each text reaction: reaction index (in the dataset), target, prediction (in the dataset stdev units).

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.