Hello,
Currently working on writing a plugin, I am taking a look at grappler and what I can do with it/what it can provides for my plugin.
My entry point for understanding grappler is the community tutorial on plugins, which i can’t thank enough for existing, but is also far from complete and also contains a few code errors.
Then there is all the info from the sources, as in grappler.h.
When implementing TF_InitGraph, the plugin author sets a TP_OptimizerConfigs field and also functions for creating an optimiser, optimising and destroying an optimiser.
So I have a few questions:
- Does activating options in
TP_OptimizerConfigsenable some existing graph optimisation techniques from tensorflow? - If 1. is true, is it compatible with providing a custom optimisation function?
- If 2. is true, when does the custom optimisation function is executed in comparison with the tensorflow ones (before, after, or some different logic)?
- The tutorial mentions the C API as well as using the graph proto, are they interchangeable? The part about the C API isn’t very clear to me unfortunately.
Please let me know if any of these needs clarification.
Thank you