Something Wrong for MACE in Four-body chiral counterexample
cenjc opened this issue · 2 comments
Choose "mace": partial(MACEModel, correlation=correlation, hidden_irreps=e3nn.o3.Irreps(f'32x0e + 32x0o + 32x1e + 32x1o + 32x2e + 32x2o')) in Four-body chiral counterexample, then it return that local variable 'last_ir' referenced before assignment.
I find that in './models/mace_modules/cg.py', it seems that the program will enter the brunch if first ir in wigners is not wigners[0][0]?
The environment is
- PyTorch version 2.2.0+cu121
- PyG version 2.4.0
- e3nn version 0.5.1
- Using device: cuda
@cenjc Thanks for your interest. As far as I recall, the version of the MACE modules we are currently using required working with e3nn version 0.4, so I recommend downgrading that.
However, you could also update your MACE modules using the MACE repo: https://github.com/ACEsuit/mace - I believe they have made a new release recently which does not enforce the e3nn version.
Closing due to inactivity.