repr

Script to generate t-SNE maps for the GDB7-22-TS set

for REPR in MPN.npy slatm_gdb.npy 3dreact-gdb-inv-random-noH-dft-true-ns64-nv64-d32-l2-vector-diff-both.123.dat; do
    for COLOR in bonds targets; do
        ./dim-reduction.gdb.equireact.py --method tsne --dump --repr_path ${REPR} \
                                         --tsne_perp 64 --tsne_ex 12 --how_to_color ${COLOR};
    done
done

The plots are produced with tsne.gp.

Depending on arguments, ./dim-reduction.gdb.equireact.py can also use other dimensionality reduction methods or produce interactive html maps.

How to get the representations:

$\mathrm{SLATM}_d$ representation

../../baseline_slatm/slatm_gdb.npy is computed with the tools of /baseline_slatm/ (not in the repo because it's too big)

3DReact representation

3dreact-gdb-inv-random-noH-dft-true-ns64-nv64-d32-l2-vector-diff-both.123.dat is obtained with

(cd ../../;
python representation.py --checkpoint results/checkpoints/cv10-gdb-inv-random-noH-dft-true-ns64-nv64-d32-l2-vector-diff-both-*.log;
sed -n -e "/>>>/s/>>>//p" `ls logs/evaluation/*.log -rt | tail -n1` > results/repr/3dreact-gdb-inv-random-noH-dft-true-ns64-nv64-d32-l2-vector-diff-both.123.dat
)

ChemProp representation:

MPN.npy is obtained with

python ../../baseline_chemprop/src/cgr-repr.py --fingerprint_type MPN \
       --checkpoint ../../baseline_chemprop/results/gdb-true/fold_0/fold_0/model_0/model.pt \
       --data_path ../../data/gdb7-22-ts/gdb.csv