MoReL

Molecule Representation Learning

TODO:

1. Data download and featurization;
   • ~~Download PubChem data~~;
   • ~~Transform into FP (ECFP4, ECFP6), SMILES, and graphs~~;
   • Target from Dragon 7 (CIC5);
2. Models
   • ~~Transformers~~;
   • ~~GGNN~~;
   • ~~GCN~~;
3. Large scale training system
   • ~~Shared data loader for multiple training threads (mmap or DictProxy)~~;
   • Process communication (HTTP?) for training control?
   • (load and save, run on more GPUs, sync training processes);
4. Training and testing
   • Training and testing functions
   • Learing rate changes (decay, early stop, etc.)
5. Final run

################################################################################ Getting features with computing method ... [Process 0] Training Utilization = 8.74% (31910 msec / 365105 msec) [Process 1] Training Utilization = 9.47% (34551 msec / 364999 msec) ################################################################################ Getting features with mmap method ... [Process 1] Training Utilization = 3.52% (68202 msec / 1940091 msec) [Process 1] Hit 68237 times; Miss 62835 times [Process 1] Feature hit ratio = 52.06% [Process 0] Training Utilization = 3.41% (66087 msec / 1940464 msec) [Process 0] Hit 62835 times; Miss 68237 times [Process 0] Feature hit ratio = 47.94% ################################################################################ Getting features with dict_proxy method ... [Process 0] Training Utilization = 11.12% (53872 msec / 484658 msec) [Process 0] Hit 65126 times; Miss 65946 times [Process 0] Feature hit ratio = 49.69% [Process 1] Training Utilization = 11.15% (54115 msec / 485209 msec) [Process 1] Hit 65946 times; Miss 65126 times [Process 1] Feature hit ratio = 50.31%

################################################################################ Getting features with dict_proxy method ... [Process 1] Training Utilization = 13.98% (33255 msec / 237942 msec) [Process 0] Training Utilization = 30.09% (71477 msec / 237581 msec) ################################################################################ Getting features with mmap method ... [Process 1] Training Utilization = 12.47% (30373 msec / 243535 msec) [Process 0] Training Utilization = 19.92% (48556 msec / 243809 msec) ################################################################################ Getting features with computing method ... [Process 1] Training Utilization = 8.26% (28501 msec / 344994 msec) [Process 0] Training Utilization = 8.20% (28536 msec / 348205 msec)

################################################################################ Getting features with dict_proxy method ... [Process 1] Training Utilization = 14.01% (33520 msec / 239321 msec) [Process 0] Training Utilization = 31.27% (74611 msec / 238637 msec) ################################################################################ Getting features with mmap method ... [Process 1] Training Utilization = 12.48% (29704 msec / 237931 msec) [Process 0] Training Utilization = 22.49% (53514 msec / 237919 msec) ################################################################################ Getting features with computing method ... [Process 1] Training Utilization = 8.28% (28455 msec / 343739 msec) [Process 0] Training Utilization = 8.15% (28068 msec / 344467 msec)

################################################################################ Getting features with computing method ... [Process 0] Training Utilization = 8.04% (28886 msec / 359262 msec) [Process 1] Training Utilization = 8.21% (29549 msec / 359983 msec) ################################################################################ Getting features with mmap method ... [Process 1] Training Utilization = 13.65% (32364 msec / 237054 msec) [Process 0] Training Utilization = 21.12% (50048 msec / 237015 msec) ################################################################################ Getting features with dict_proxy method ... [Process 1] Training Utilization = 14.39% (34152 msec / 237283 msec) [Process 0] Training Utilization = 29.90% (70945 msec / 237245 msec)