Hi,
I’m using TensorRT thankfully, and curious about how to measure the accuracy of half-precision.
Using half-precision(floating point 16), good performance on calculating, but I would like to check does it preserve similar accuracy as using full-precision.
Is it possible to calculate map by using tensorRT on kHALF mode?
If it is not, I am using yolo network which is modified by alexeyab, [url]https://github.com/AlexeyAB/darknet[/url]
and there is the mode called cudnn-half, which says mixed-precision on tensor cores, and it can calculate the map.
Can I think this map value would be similar to tensorRT kHALF mode performance?
Thanks for reading and expect for replies…