Implementation of EquivariantPowerSpectrum and EquivariantPowerSpectrumByPair convenience calculators
#317
Conversation
…rder_nu"` dimension in keys.
|
Here is a pre-built version of the code in this pull request: wheels.zip, you can install it locally by unzipping |
| # ====================================================================== | ||
|
|
||
|
|
||
| class LambdaSoap(TorchModule): |
There was a problem hiding this comment.
I think I would prefer something like LambdaPowerSpectrum/LambdaSoapPowerSpectrum/EquivariantPowerSpectrum/EquivariantSoapPowerSpectrum
I'm fighting against the SOAP == PowerSpectrum thinking, and making this explicit would help IMO
| self, | ||
| spherical_expansion_hypers: dict, | ||
| atom_types: Optional[List[int]] = None, | ||
| angular_cutoff: Optional[int] = None, |
There was a problem hiding this comment.
do we really want this as a separate hyper for angular cutoff? Should we use the one from spx hypers instead?
| spherical_expansion_hypers: dict, | ||
| atom_types: Optional[List[int]] = None, | ||
| angular_cutoff: Optional[int] = None, | ||
| selected_keys: Optional[Labels] = None, |
There was a problem hiding this comment.
this should be part of compute/forward: different calls to the same calculator might request different keys
| return lsoap | ||
|
|
||
|
|
||
| class LambdaSoapXSphExByPair(TorchModule): |
There was a problem hiding this comment.
This is \rho^2 x g_ij, right? since g_ij is SphericalExpansionByPair, this should be BispectrumByPair IMO (with potential Lambda, Soap, Equivariant in there as well).
jwa7
changed the title
LambdaSoap and LambdaSoapXSphExByPair convenience calculatorsEquivariantPowerSpectrum and EquivariantPowerSpectrumByPair convenience calculators
Implements:
EquivariantPowerSpectrum) calculator that wrapsDensityCorrelationsCorrelateTensorWithDensitycalculator that performs a single CG tensor product between a tensor of arbitrary body order and a density (body-order 2) tensor, where the samples between the tensor and density are different.EquivariantPowerSpectrumByPaircalculator, which generates aSphericalExpansionandSphericalExpansionByPair, manipulates the metadata appropriately, then usesCorrelateTensorWithDensitycalculator to performrho_i x g_ijto generate an Equivariant power spectrum by pair📚 Documentation preview 📚: https://rascaline--317.org.readthedocs.build/en/317/