NSAPH-Projects / topological-equivariant-networks

E(n)-Equivariant Topological Neural Networks
MIT License
19 stars 0 forks source link

Change data representation in collate_fn to use lists of integers #57

Closed mauriciogtec closed 2 months ago

mauriciogtec commented 3 months ago

There is no need to use padding for cell membership in the loader representations. Cell membership padding is not used for spatial tasks, which also have huge differences in cell sizes. The data collider supports lists of integers. Padding can be created efficiently on the fly with torch functions for qm9.

This issue should:

  1. Represent internally as a list of integers in the collate fn
  2. For qm9, restore the original padded format for backward compatibility

After this issue is solved, other issues/PRs will integrate the spatial benchmarks.

mauriciogtec commented 3 months ago

Maybe the solution is to have two collators

mauriciogtec commented 2 months ago

Fixed in d736879.