Floating/Double precision

lamoureux-lab / TorchProteinLibrary

PyTorch library of layers acting on protein representations

https://lamoureux-lab.github.io/TorchProteinLibrary/

MIT License

116 stars 23 forks source link

Floating/Double precision #18

Closed lupoglaz closed 4 years ago

lupoglaz commented 5 years ago

Many layers work with double precision right now: basically all the CPU based layers do + RMSD layers. We need to unify c++ code/cuda kernel calls using ATen dispatch templates.

lupoglaz commented 5 years ago

Templated basic CPU functions implemented in commit 0938d8d10487eb934c55ecacc708573b76d3f236

lupoglaz commented 5 years ago

FullAtomModel: floating types dispatch: d36bc3404db28948348fa91bdeb7c2a6bc9bed69

lupoglaz commented 5 years ago

The layers that still work with only one type: documented:

VolumeConvolution

VolumeSelect

should stay double only:

PDB2CoordsUnordered

getRandomTranslation

getRandomRotations

not documented:

VolumeRotate

VolumeRMSD

volume2Xplor

lupoglaz commented 4 years ago

VolumeConvolution and VolumeSelect work with diff. types of precision now. Closing this issue