ProtpardelleWrapper.init_from_noise currently uses the unfiltered structure to derive the number of atoms. By convention, we should remove zero-occupancy atoms and any with nan coordinates, so that it agrees with what happens in sampleworks.eval.structure_utils.process_structure_to_trajectory_input. In other models, this happens internally in the model-specific codebases, generally in their data loaders. Given the comparative simplicity of the Protpardelle inputs, we don't use their data loader, so need to do this filtering.
ProtpardelleWrapper.init_from_noise currently uses the unfiltered structure to derive the number of atoms. By convention, we should remove zero-occupancy atoms and any with nan coordinates, so that it agrees with what happens in
sampleworks.eval.structure_utils.process_structure_to_trajectory_input. In other models, this happens internally in the model-specific codebases, generally in their data loaders. Given the comparative simplicity of the Protpardelle inputs, we don't use their data loader, so need to do this filtering.