Since hydrogens are treated separately, every molecule has n potential training points for hydrogens.
Usually the training set is split per molecule which consequentially leads to almost a factor of 10 more training points for hydrogens (depends on molecule size of course).
If someone is not aware and just builds the kernel this can easily fill up 500gb of memory with only 20k molecules.
Since the hydrogen representation has limited uniqueness anyway it is safe to subsample and build a separate learning curve.
This info is not explicitly documented in the tutorial though and has to be added
Since hydrogens are treated separately, every molecule has n potential training points for hydrogens.
Usually the training set is split per molecule which consequentially leads to almost a factor of 10 more training points for hydrogens (depends on molecule size of course).
If someone is not aware and just builds the kernel this can easily fill up 500gb of memory with only 20k molecules.
Since the hydrogen representation has limited uniqueness anyway it is safe to subsample and build a separate learning curve.
This info is not explicitly documented in the tutorial though and has to be added