Hey everyone. I’m new to the forum, so please don’t be too hard on me if I missed a rule.
I’m trying to call tensorflow from a jitted numba function. You don’t really need to know anything about numba to know what I need, but in case you are interested, here is a link to what I need from tensorflow: Types and signatures — Numba 0.56.4+0.g288a38bbd.dirty-py3.7-linux-x86_64.egg documentation
Essentially, I need a way to get back a true function pointer for a tensorflow function (probably compiled with tf.function and get_concrete_function). With this, I can call tensorflow from my numba function. The trouble is, as I started looking through the TF code, there are so many layers that I’m having trouble finding the right place to hook into and I can’t find any documented API for the python TF API which provides access to a function pointer.
The other possible way to implement it that I’ve thought of is that there may be some way to use the TF C++ API to gain access to a function I can call. This is less optimal and ultimately the code needs to be able to run from python because that is the requirement for the project, but there may be some way to do it from here.
Anyway, if anyone can help point me in the right direction, it would be appreciated.