[GitHub] [incubator-mxnet] mk-61 edited a comment on issue #16173: Saving and loading cudNN autotune and graph optimization

[GitBox <gi...@apache.org>]

mk-61 edited a comment on issue #16173: Saving and loading cudNN autotune and graph optimization
URL: https://github.com/apache/incubator-mxnet/issues/16173#issuecomment-598392266
 
 
   How about adding a method to Symbol, to calculate certain aspects of a model? In C API it would look something like:
   
   `int MXCalculateDigest(SymbolHandle handle, /* type TBD * / selector)?`
   
   It would calculate a hash of
   
   - MXNet version
   - model structure
   - attributes, filtered by the selector argument
   - anything else?
   
   Then any library / optimizer would call it, add some other, optimization-specific details (CUDA version, hardware details, it's own library version, etc.) and would make a decision on whether to use some saved optimization result or perform a new optimization step? I can imagine it can save it as a graph attribute, for instance.
   
   It will save us from piping through a centralized API particulars of specific libraries / optimisations, and will give more control to optimisers, who will decide, which aspects are relevant to them and which are not.   
   

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services