search

WXinlong / SOLO

SOLO and SOLOv2 for instance segmentation, ECCV 2020 & NeurIPS 2020.
Other
1.68k stars 307 forks source link

Tensorrt implementation #129

Open batrlatom opened 3 years ago

batrlatom commented 3 years ago

Tensorrt would increase the speed of the model 3x. It would help to make it usable on devices like jetson nano and jetson xavier nx.

ghost commented 3 years ago

Are you using Tensorrt to deploy on Jetson xavier nx?

batrlatom commented 3 years ago

Normally yes, but not yet for your model. I do not know about any instance segmentation network which could be reasonably exported to tensorrt ( maskrcnn could, but the speedup is not huge ). Semantic segmentations like FCN and Deeplab works quite well. This is more feature request, if you would be able to take a look onto it. Usually, It could be done via pytorch -> onnx -> tensorrt. I saw that few people asked about onnx export, this is first step to go onto tensorrt.

zzzz737 commented 3 years ago

Tensorrt would increase the speed of the model 3x. It would help to make it usable on devices like jetson nano and jetson xavier nx.

Hello, do you export solov2 to tensorrt sucessful? I'm failed!

lucasjinreal commented 3 years ago

@zzzz737 There was someone sucessfully convert a solov2-liked instance segmentation model to tensorrt:

https://zhuanlan.zhihu.com/p/341813709

request them maybe they can also convert solov2

talebolano commented 3 years ago

see my repository. maybe helpful