Zero face detected after trt conversion(Retinaface)

[dubeamit]

trafficstars

Env

• GPU: NVIDIA A10
• OS: Ubuntu 20.04.4 LTS
• CUDA Version: 11.6
• Driver Version: 510.47.03
• TensorRT version: 8.0.1.6
• gcc version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0

I am running this inside a container from nvidia: container docker pull nvcr.io/nvidia/tensorrt:21.07-py3 link

About this repo

• repo: tensorrt/retinaface
• branch: master (TensorRT 7 API)
• retinaface

Your problem

1. build opencv from source opencv version 4.6.0

2. pytorch version: torch==1.12.0+cu116, torchaudio==0.12.0+cu116, torchvision==0.13.0+cu116

3. Followed the instructions in the repo cloned the Pytorch_Retinaface.git and generated the retinaface.wts

4. here test.jpg file was generated with all faces detected

5. cloned https://github.com/wang-xinyu/tensorrtx.git repo

6. after doing cmake, make and serialize, when running inference zero face are detected in worlds_largest_selfie

Expectation: to detect faces after trt conversion

also tried with different trt versions and cuda but still not getting and detections after trt conversion different combination:

1. container: docker pull nvcr.io/nvidia/tensorrt:21.06-py3, TensorRT-7.2.3.4, torch==1.12.0+cu113, torchaudio==0.12.0+cu113, torchvision==0.13.0+cu113
2. container: docker pull nvcr.io/nvidia/deepstream:6.1-triton, tensorrt==8.2.5.1, torch==1.12.0+cu116, torchaudio==0.12.0+cu116, torchvision==0.13.0+cu116

even after trying different versions of trt and cuda and pytorch still not getting any detections after trt conversion is done.

[wang-xinyu]

Which pytorch trained model were you using?

[dubeamit]

Which pytorch trained model were you using?

I used the pretrained model of retinaface Resnet50_Final.pth mentioned in the repo here

[wang-xinyu]

@dubeamit Maybe your pytorch version it too new, at that repo, they were using 1.1. Can you try 1.1 or 1.3?

[dubeamit]

@wang-xinyu so this worked with pytorch 1.4 and cuda-11.1 on Tesla T4 GPU. But with pytorch 1.4 and cuda 11.3 on A10 GPU it gets stuck in loading model and I can't generate the weight.

So I used the same container on Tesla T4 gpu having TensorRT container image version 20.09 is based on TensorRT 7.1.3 and cuda 11.0.

After getting the retinaface.wts file and running ./retina_r50 -s I get the following error

root@nvmbdprp023198:~/tensorrtx/retinaface/build# ./retina_r50 -s
Loading weights: ../retinaface.wts
Building engine, please wait for a while...
[07/07/2022-11:32:15] [W] [TRT] Half2 support requested on hardware without native FP16 support, performance will be negatively affected.
[07/07/2022-11:32:23] [E] [TRT] ../rtSafe/cuda/caskUtils.cpp (98) - Assertion Error in trtSmToCask: 0 (Unsupported SM.)
Build engine successfully!
retina_r50: /root/tensorrtx/retinaface/retina_r50.cpp:251: void APIToModel(unsigned int, nvinfer1::IHostMemory**): Assertion `engine != nullptr' failed.
Aborted (core dumped)

[wang-xinyu]

Since it worked in one of your environment, then it's not the code issue. You need to check the gpu driver, cuda, tensorrt installation, etc.

[dubeamit]

Which is max version of cuda and tensorrt that this code can support? since I've already tried with docker containers having

1. TensorRT-8.2.5.1, cuda11.6
2. TensorRT-7.2.3.4, cuda11.3
3. TensorRT-8.0.1.6, cuda11.6
4. TensorRT-7.1.3.4, cuda11.0

in the 4th one I get unsupported error. in others I simply don't get any detection.

for the other combinations by changing the conf_thres to 0.5 I get detections all over the place see the attachment

Also host machine(A10) has latest cuda and nvidia drivers as mentioned it the first post.

[snlpatel001213]

use https://github.com/wang-xinyu/tensorrtx/tree/master/docker/trt8 to build the docker, which may solve the problem

[dubeamit]

@snlpatel001213

use https://github.com/wang-xinyu/tensorrtx/tree/master/docker/trt8 to build the docker, which may solve the problem

I tried the above docker but still not getting any detections. cuda version inside container 11.2, in host machine 11.6

while converting to trt I get the following prints

[07/18/2022-17:01:57] [W] [TRT] TensorRT was linked against cuBLAS/cuBLASLt 11.6.5 but loaded cuBLAS/cuBLASLt 11.4.1
[07/18/2022-17:02:00] [W] [TRT] TensorRT was linked against cuDNN 8.2.1 but loaded cuDNN 8.1.1
[07/18/2022-17:02:00] [W] [TRT] TensorRT was linked against cuBLAS/cuBLASLt 11.6.5 but loaded cuBLAS/cuBLASLt 11.4.1
[07/18/2022-17:02:00] [W] [TRT] TensorRT was linked against cuDNN 8.2.1 but loaded cuDNN 8.1.1

Pytorch version:

torch==1.12.0+cu113
torchaudio==0.12.0+cu113
torchvision==0.13.0+cu113

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.