Multi-task MPNN target sizes do not match about molpal HOT 4 CLOSED

Hi @miquelduranfrigola,

We currently don’t support multitask optimization in molpal, so setting the number of tasks to anything other than 1 will break the code. How are you currently using the code?

from molpal.

Hi @davidegraff

Thanks for the fast reply! I was interested in the fact that you are using PyTorch Lightning, so my plan was to use the MPNN class of MolPal as a drop-in-replacement for my ChemProp multi-task regression models, which are typically slow. I hope this makes sense?

Thanks!
M

from molpal.

yeah so there is a bug right now in molpal.models.mpnmodels.py#L191.
This block:

def make_datasets(
    self, xs: Iterable[str], ys: Sequence[float]
) -> Tuple[MoleculeDataset, MoleculeDataset]:
    """Split xs and ys into train and validation datasets"""

    data = MoleculeDataset([
        MoleculeDatapoint(smiles=[x], targets=[y])
        for x, y in zip(xs, ys)
    ])
    ...

assumes that ys is an array of single-task target values, so it's a vector of length n rather than an array of shape n x 1. If you iterate through an n x m array and then wrap it in a list (as in targets=[y]), then the target of each point is a list of m floats rather than m separate targets.

I just fixed this by checking the target shape in MPNN.train(#10). let me know if this problem persists after the latest commit

from molpal.

Hi @davidegraff it works nicely now with the latest commit! Many thanks for your help!

from molpal.