search

ctongfei / nexus

Experimental tensor-typed deep learning
https://tongfei.me/nexus/
MIT License
257 stars 15 forks source link

Remove Half floating point type #6

Closed fvlankvelt closed 6 years ago

fvlankvelt commented 6 years ago

The Half floating point type, aka Float16, is an interesting target for numerical computation as it trades accuracy for scaling to large numbers of parameters or data points. However, implementation is not complete and it does not add to expressiveness of the API.

ctongfei commented 6 years ago

Torch C core has HalfTensors. It is supposed to be there for the support of HalfTensors. Once I got time its implementation would be complete.

fvlankvelt commented 6 years ago

yeah, time is very often the limiting factor. But a partial implementation is better parked in a branch. Or just as a feature request, to be picked up when the time is available.

So maybe the cleanup PR should be accompanied by a branch with the change re-applied. So that the feature can be merged again when it is complete?

ctongfei commented 6 years ago

that's ideal. We could move Half to a branch.