Skip to content

Instantly share code, notes, and snippets.

@farizrahman4u
Last active March 27, 2018 18:31
Show Gist options
  • Save farizrahman4u/48ed368ed51b61138292069eac3a2035 to your computer and use it in GitHub Desktop.
Save farizrahman4u/48ed368ed51b61138292069eac3a2035 to your computer and use it in GitHub Desktop.
  • On the Java end, exposing an op from libnd4j involves adding classes and methods to 3 different places. We should consider having a universal function/op factory.

  • Blur the lines between Nd4j/samediff. All the ops from the universal function factory mentioned above should be available under Nd4j namespace. In other words, Nd4j.some_op should work on both INDArray as well as SDVariable inputs.

  • More concrete shape inference. Shape inference in Samediff seems to be flaky, and dependent on op execution. Shape inference should be greedy with provision for unknown dimensions (this would be more involved and require changes in libnd4j). Should also consider symbolic shapes, i.e, SDVariable.shape() would return a 1-d SDVariable.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment