search

OpenGVLab / InternImage

[CVPR 2023 Highlight] InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions
https://arxiv.org/abs/2211.05778
MIT License
2.48k stars 231 forks source link

more info for FPS #95

Open gongchengli opened 1 year ago

gongchengli commented 1 year ago

Hi all, recently I successfully export segmentation model from pytorch to tensorrt using the provided script. Then I run this script (from mmdeploy) to evaluate the speed. python tools/test.py ../InternImage/segmentation/deploy/configs/mmseg/segmentation_tensorrt_static-512x512.py ../InternImage/segmentation/configs/ade20k/upernet_internimage_t_512_160k_ade20k.py --model ../InternImage/segmentation/deploy/outputs/end2end.engine --speed-test --device cuda

The FPS is about 10 in tensorrt, about 5 in pytorch. The model is UperNet-InternImage-T-512 . I done this on nvidia Jetson AGX Orin 64GB, not sure if it's correct compared with your performance.

SamratThapa120 commented 1 year ago

Hello @gongchengli , what is your inference speed using tensorrt without using GPU?