Roadmap¶
Brief description of the incoming SymJAX updates.
Short-Term¶
- graph saving and loading (right now one can only save and load variables, this requires the user to redefine a graph prior loading the saved variables in order to keep training/updating/evaluating a graph, saving/loading the graph itself would remove this need
- function based data/output saving in h5 files allowing to (in real time) save and monitor specified quantities evaluated through each function
Long-Term¶
- support of the
vmap
andpmap
Jax functions allowing distributed and optimized computations for more general graphs - general wrapping of Tensorflow modules using Jax as backend