brain icon indicating copy to clipboard operation
brain copied to clipboard
verified publisher icon ReconfigureIO

using TensorFlow Go APIs

Open Mahdi89 opened this issue 8 years ago • 0 comments

On using TF Go bindings I can think of two ways forward:

  • Mapping the entire constructed TF graph to our Teak dataflow (which is also closer to row stationary form).
  • Add custom new op to TF and have it registered as a new kernel. Our compiler takes care of mapping the kernel to Teak dataflow form.

Imo the former may require a clear understanding of the TF graphs but the latter sounds more viable as TF treats the new kernels as independent entities.

Mahdi89 avatar Oct 08 '17 18:10 Mahdi89