Neural network for predicting template relevance a la Segler and Waller's Neural Symbolic paper.

• RDKit (most versions should be fine)
• numpy

• RDKit (most versions should be fine)
• tensorflow (r0.12.0)
• h5py
• numpy

Learn to predict template relevance.

1. Grab reaction precedents from templates stored in MongoDB python scripts/get_reaxys_data.py

2. Calculate fingerprints and store in .h5 file python scripts/make_data_file.py data/reaxys_limit1000000000_reaxys_v2_transforms_retro_v9_10_5.txt 2048

3. Train model python retrotemp/nntrain_fingerprint.py -t data/reaxys_limit1000000000_reaxys_v2_transforms_retro_v9_10_5.txt -o 163723 -m models/6d3M_Reaxys_10_5 --fp_len 2048

4. Find best validation performance

   regex="model\.(.*)\.meta"
   for f in `ls -tr models/6d3M_Reaxys_10_5/*.meta`
   do
       if [[ $f =~ $regex ]]
       then
           ckpt="${BASH_REMATCH[1]}"
           echo $ckpt
           python retrotemp/nntrain_fingerprint.py  -o 163723 -m models/6d3M_Reaxys_10_5 --fp_len 2048 -c "$ckpt" -t data/reaxys_limit1000000000_reaxys_v2_transforms_retro_v9_10_5.txt --test valid
       fi
   done
   

5. Retrain on whole dataset (?) for same number of epochs. Note: this is because we want a high-performing deployed model and no longer need to hold out any data. python retrotemp/nntrain_fingerprint.py -t data/reaxys_limit1000000000_reaxys_v2_transforms_retro_v9_10_5.txt -o 163723 -m models/6d3M_Reaxys_10_5 --fp_len 2048 --fixed_epochs_train_all 15

6. Run standalone tensorflow version to dump to numpy arrays