Inference with fp16 precision

[mdcaio]

Hi, I see that it is possible to do training with fp16 precision by adding something like fp16 = dict(loss_scale=512) in the config file (I haven't actually tried).

I would like to do inference at fp16 precision. I'm currently doing inference using init_detector and inference_detector from the inference API. Is there a config parameter to use to run inference at fp16 precision? (to speed up inference?)