I converted the models tofloat32 using this script:https://gist.github.com/pcuenca/23cd08443460bc90854e2a6f0f575084, but found precision problems when targetingfloat16. It'd be interesting to see what the performance is forfloat16, but we need to determine what layers/ops need to be kept infloat32. Anyone interested please let us know and we can work on it or test together :)