using TensorFlow Go APIs #1
Description
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.