notes.md

experiments

does size matter?

unsupervised zinc, 250k vs 2mil

supervised or unsupervised? more labels help?

zinc 250k vs chembl_filtered (more labels)

what training objective works best?

superised dgi masking unsupervised followed by supervised unsupervised + supervised (together)

how to pad

is there a correlation between the features distribution of downstream task and the of the pretriain dataset

use Bhattacharyya distance or Hellinger distance to measure

supervised chembl with calculated targets

# logp only
python pretrain_gt_supervised.py --config config/pretrain_supervised_chembl_syn_logp.json --exp_name pretrain_supervised_chembl_cal_logp --gpu_id 1

# all calculated targets TODO
python pretrain_gt_supervised.py --config config/pretrain_supervised_chembl_syn_all.json --exp_name pretrain_supervised_chembl_cal_all --gpu_id 1

# selected targets
python pretrain_gt_supervised.py --config config/pretrain_supervised_chembl_syn_selected.json --exp_name pretrain_supervised_chembl_cal_selected --gpu_id 2

# all natural and calculated targets TODO: bce loss + L1 loss??? ######
python pretrain_gt_supervised.py --config config/pretrain_supervised_chembl_natural_labels.json --exp_name pretrain_supervised_chembl_all --dataset chembl_filtered --gpu_id 3

# all natural targets TODO
python pretrain_gt_supervised.py --config config/pretrain_supervised_chembl_natural_labels.json --exp_name pretrain_supervised_chembl_natural_labels --dataset chembl_filtered --gpu_id 3

finetune chembl

TODO

pretrain zinc-full pretrain chembl-selected