Support the deployment of video supper-resolution algorithm BasicVSR on TensorRT backend 支持视频超分模型basicvsr在TensorRT侧的部署

[OpenMMLab-Assistant-004]

Support the deployment of video supper-resolution algorithm BasicVSR on TensorRT backend

支持视频超分模型basicvsr在TensorRT侧的部署

Support the deployment of BasicVSR algorithm from mmagic onTensorRT.
You can refer to the following resources:

1. [BasicVSR](https://github.com/open-mmlab/mmagic/blob/main/configs/basicvsr/basicvsr_2xb4_reds4.py ）
2. MMagic Deployment
3. Deployment Tutorials

在mmdeploy支持BasicVSR算法在TensorRT侧部署。
相关资源

1. https://github.com/open-mmlab/mmagic/blob/main/configs/basicvsr/basicvsr_2xb4_reds4.py
2. https://mmdeploy.readthedocs.io/en/latest/04-supported-codebases/mmagic.html
3. https://mmdeploy.readthedocs.io/zh_CN/latest/tutorial/01_introduction_to_model_deployment.html

Task Information

Repository：MMDeploy

Difficulty：Medium

Points：45

Estimated Time Consumption：28 days

Task Status

Claimant's Github ID：gachiemchiep

Claim Time：2023年07月28日

PR Link：open-mmlab/mmdeploy#2402

Points to Note

❗️请注意：此任务属于超级视客营任务，不能被领取！请登陆 超级视客营平台 按照流程申请任务，我们会尽快审核～

[gachiemchiep]

Hello guys.

I add the PR here : open-mmlab/mmdeploy@main...gachiemchiep:mmdeploy:codecamp_basicvsr

Currently I run the following command to convert the model.
Unfortunately, the converting process take very long time and eat up all available ram-memory.
So I have to cancel it before it make my server freeze.
I used a server machine which 377G memory and Tesla V100 gpu.
So it is definitely software problem.
Do you guys know how to overcome this problem?

python ./tools/deploy.py     
    configs/mmagic/super-resolution/video-super-resolution_tensorrt_dynamic-32x32-512x512.py     
   $MMDIR/ws/checkpoints/basicvsr_2xb4_vimeo90k-bd.py     
   $MMDIR/ws/checkpoints/basicvsr_vimeo90k_bd_20210409-0154dd64.pth     
   $MMDIR/ws/resources/input/video_restoration/QUuC4vJs_000084_000094_176x144.mp4     
   --work-dir work_dir      --device cuda   --dump-info

[crazysteeaam]

@gachiemchiep You can add this WeChat account (WeChat ID: OpenMMLabwx)

[gachiemchiep]

@crazysteeaam
I did but couldn't understand the message written in Chinese....
my wechat id is : wxid_c76s7u9ce4mf22

[crazysteeaam]

@gachiemchiep I cannot find any account through this wechat id wxid_c76s7u9ce4mf22 🤯

[gachiemchiep]

Here is my qr code

[rolfma19]

I found that the memory usage is related to the number of input video frames (input shape is [b,n,c,h,w]) when using torch.onnx.export() to export BasicVSR to ONNX. The memory usage of exporting is related to n and the exported ONNX graph structure is also related to n.