search

VITA-Group / FasterSeg

[ICLR 2020] "FasterSeg: Searching for Faster Real-time Semantic Segmentation" by Wuyang Chen, Xinyu Gong, Xianming Liu, Qian Zhang, Yuan Li, Zhangyang Wang
MIT License
525 stars 107 forks source link

Parse Issue of TensorRT #7

Closed yu-changqian closed 4 years ago

yu-changqian commented 4 years ago

When I use the computation code of the tensorrt in your repo, there is a problem.

When the TensorRT ONNXParser parses the onnx model, the returend value is false. And the TensorRT gives an error: Network must have at leaset one output.

Do you have meet the similar problem? And how to solve it?

chenwydj commented 4 years ago

Did you change anything in the provided latency measurement script? i.e. code under FasterSeg/latency.

I met this problem before. Usually, I have to go through the error log from the TensorRT to find the reason, it will say which layer the error comes from -- mostly it is because some layers in Pytorch is not supported by TensorRT.

yu-changqian commented 4 years ago

No. I didn't do any change.

What's the version of your TensorRT? Maybe the different versions cause the problems.

chenwydj commented 4 years ago

Mine is 5.1.5.

yu-changqian commented 4 years ago

Wired. Mine is also 5.1.5.

Meanwhile, the versions of PyTorch, CUDA are 1.3 and 10.1.

yu-changqian commented 4 years ago

Did you do some specific change on the "interpolate"?

I find that the TensorRT 5.0 is not supportted the "Resize".

chenwydj commented 4 years ago

Could you try Pytorch 1.1.0? And also my onnx version is 1.5.0

chenwydj commented 4 years ago

Did you do some specific change on the "interpolate"?

I find that the TensorRT 5.0 is not supportted the "Resize".

Nothing special. TensorRT 5.1 supports the interpolation.

yu-changqian commented 4 years ago

Now I use Pytorch 1.2 and TensorRT 5.1. All is Fine. Thanks for your kindly reply.