search

warmshao / FasterLivePortrait

Bring portraits to life in Real Time！onnx/tensorrt support！实时肖像驱动！
521 stars 47 forks source link

About performance with onnxruntime #100

Open Ryuukeisyou opened 1 week ago

Ryuukeisyou commented 1 week ago

My env: gpu nvidia 4090 system windows cuda 12.4 cudnn 9.1

I migrated onnxruntime code for grid_sample 5D from liqun/imageDecoder_cuda branch to the main branch and compiled. code is here https://github.com/Ryuukeisyou/onnxruntime Now the onnxruntime is running at about 20fps (0.05s for each step) for inference only, not counting the pre or post processing.

I used the following to run the test, same with what in the readme: python run.py --src_image assets/examples/source/s10.jpg --dri_video assets/examples/driving/d14.mp4 --cfg configs/onnx_infer.yaml

Does anybody know if this is the normal speed, or maybe I missed something when migrating? Any help would be appreciated, thanks!

warmshao commented 1 week ago

python run.py --src_image assets/examples/source/s10.jpg --dri_video assets/examples/driving/d14.mp4 --cfg configs/onnx_infer.yaml

The speed with onnxruntime is normal, but I suggest you use TensorRT to achieve faster inference speed. You need to follow the tutorial to install TensorRT, perform model conversion, and finally run： python run.py --src_image assets/examples/source/s10.jpg --dri_video assets/examples/driving/d14.mp4 --cfg configs/trt_infer.yaml