search

lewes6369 / TensorRT-Yolov3

TensorRT for Yolov3
MIT License
489 stars 165 forks source link

Is the data type of the yolo3 model under the darknet or caffe framework is double precision? #65

Closed jianwu585218 closed 4 years ago

jianwu585218 commented 5 years ago

Thanks for your work~ I have a question，that is :
Is the data type of the yolo3 model under the darknet or caffe framework is double precision（FP64）? image

lewes6369 commented 5 years ago

It is also float32 in caffe.

jianwu585218 commented 5 years ago

Thanks Why the second line improved a lot, it is also float32 which as same as caffe? Due to the framework ，or other way(such as Layer & Tensor Fusion)?

lewes6369 commented 5 years ago

It benefits to TensorRT. See here https://developer.nvidia.com/tensorrt

jianwu585218 commented 5 years ago

Thanks for your repley~ Have you test the yolo3-416 's inference time under float16（by TensorRT）？

lewes6369 commented 5 years ago

My graphics cards did not support fast fp16, just take a try on your card.