getnamo / TensorFlow-Unreal

TensorFlow plugin for the Unreal Engine.
Other
1.16k stars 211 forks source link

Refactor to support Native and Remote variants #53

Open getnamo opened 4 years ago

getnamo commented 4 years ago

Motivation It should be possible to switch the machine learning backend while keeping the unreal frontend api the same, allowing for a very flexible dev environment. By specifying an abstract UMachineLearningBaseComponent (https://github.com/getnamo/machine-learning-remote-ue4/blob/master/Source/MachineLearningBase/Public/MachineLearningBaseComponent.h) we can sub-class these components into remote, unreal python, and native variants. Depending on which backend is needed these should be swappable without any dev code changes.

From the server side we can also specify a base MLPluginAPI which wouldn't be tensorflow specific. This would open up Pytorch backends without any unreal frontend code change.

Having a remote server component would also enable linux/mac builds without restricting python/tensorflow versions to compatible ones with unrealenginepython. It would also enable remote ML on e.g. phone devices (native may support TFLite at some point).

Remote work

Native work Intended to be inference focused initially

Tensorflow-ue4 Refactor

getnamo commented 4 years ago

remote variant working, native pending

getnamo commented 4 years ago

Early refactor release with auto-server launch support: https://github.com/getnamo/tensorflow-ue4/releases/tag/1.0.0alpha2