A Hybrid of Imperative and Symbolic Programming

https://en.diveintodeeplearning.org/chapter_computational-performance/hybridize.html

I actually changed the GoogLeNet or ResNet code into HybridBlock, used HybridSequential, and then net.hybridize(). Running times on a GPU were hardly different…
This hybrid concept is not well documented. Maybe one has to use the arguments of hybridize to get any speed-ups, or change the order of net.initialize() and net.hybridize(). All a bit confusing to me, warrants a proper tutorial that explains what is going on.