CODEWITHSUNDEEP
onnxtensorrttensorrt-python
Faisal Hejary
Faisal Hejary

Reputation: 13

infer using mixed precision in tensorrt

I'm currently using DETR for object detection. I want to convert it as follows: pytorch -> onnx -> tensorrt I have the code to do so and tested the model achieving the same performance in all formats. the thing is, the model is in fp32 and when I convert it to fp16 I lose a lot of performance. My idea is to convert some layers to fp16 and leave the rest as fp32 to keep as much accuracy.
my question is. how to convert specific layers of the tensorrt model into fp16? I couldn't find any documentation on this. any and all help is appreciated.

Upvotes: 1

Views: 408

Answers (0)

Related Questions