The way we have the LigandOnlyModel/GAT model set up is kind of a weird special case of the regular Model. We could probably rework the GAT._get_representation and GAT._get_energy_func methods to work the same way as they do in the structure-based models, and then have the strategy be a ComplexOnlyStrategy, so we only process the one representation. Either way, we should probably move the self.readout call into the representation, since the current representation for GAT outputs an mxn array, where n is the feature dimensionality and m is the number of nodes. This is not the way we have things setup for everything else, where the output of the representation should just be an n-dim vector.
The way we have the
LigandOnlyModel/GATmodel set up is kind of a weird special case of the regularModel. We could probably rework theGAT._get_representationandGAT._get_energy_funcmethods to work the same way as they do in the structure-based models, and then have the strategy be aComplexOnlyStrategy, so we only process the one representation. Either way, we should probably move theself.readoutcall into the representation, since the current representation for GAT outputs an mxn array, where n is the feature dimensionality and m is the number of nodes. This is not the way we have things setup for everything else, where the output of the representation should just be an n-dim vector.