Do we really need java2cpp ?

[yupbank]

Since tensorflow.tensor also support protobuf

https://github.com/tensorflow/tensorflow/blob/754048a0453a04a761e112ae5d99c149eb9910dd/tensorflow/core/framework/tensor.proto

I saw there would be a performance issue here

about 80% of this time is spent on slow data conversions between Spark and TensorFlow

would it worth it to convert data into proto and cache it if needed

[thunterdb]

I thought that we had removed all the dependencies to javacpp? Under the hood, the official java-tensorflow is used for transfer.

…

On Wed, Mar 21, 2018 at 4:19 PM, Peng Yu ***@***.***> wrote: Since tensorflow.tensor also support protobuf https://github.com/tensorflow/tensorflow/blob/ 754048a0453a04a761e112ae5d99c149eb9910dd/tensorflow/core/ framework/tensor.proto I saw there would be a performance issue here <https://github.com/databricks/tensorframes/wiki/Example:-K-Means> about 80% of this time is spent on slow data conversions between Spark and TensorFlow would it worth it to convert data into proto and cache it if needed — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#142>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHPjARI5rwS-17v0hXot2FkfAhc4M67sks5tgn2dgaJpZM4S0k2q> .

[yupbank]

oh.. nice nice, maybe we can update this https://github.com/databricks/tensorframes#acknowledgements

and... can we cache the tensors? for the k means example would do multiple passes. it would save the convert time