Installation issues: I adapted the TSN Docker container – feel free to use it

[v-iashin]

Hi everyone!

I was trying to make the code run on my machine and I faced significant trouble installing the correct versions of packages because they became obsolete. Then I tried docker. The experience with it wasn't smooth either because dense_flow installation wasn't compatible with my GPU. So, I tried the docker container provided for the TSN network. I tested it with 1080Ti (worked) and 2080Ti (didn't work w/ CUDNN_STATUS_EXECUTION_FAILED) with 64GB RAM.

docker pull iashin/anet2016-cuhk:latest

or build on top of it by specifying the header of your Dokerfile:

FROM iashin/anet2016-cuhk:latest

The output from the examples:

• python examples/classify_video.py --use_flow data/plastering.avi

frame sample 53: 0.322223 second
frame sample 54: 0.322393 second
total time: 20.200887 second
----------------Classification Results----------------------
Plastering 0.9559713
Rock climbing 0.020194435
Painting fence 0.011123132
Hanging wallpaper 0.010004199
Painting furniture 0.0007557427
Hand washing clothes 0.00065550825
Laying tile 0.00040866304
Throwing darts 0.00033438313
Paintball 0.00019233384
Hitting a pinata 0.000125195

• python examples/classify_video.py --use_flow https://www.youtube.com/watch?v=QkuC0lvMAX0

total time: 91.161525 second
----------------Classification Results----------------------
Windsurfing 1.0
Sailing 6.5195724e-16
Surfing 4.3629453e-17
Kite flying 2.1080195e-18
Wakeboarding 5.306861e-19
Snowboarding 6.466279e-21
Plataform diving 3.3573223e-22
Doing motocross 1.9524087e-22
Scuba diving 1.3819956e-22
Fixing bicycle 1.2276588e-22

Here is the Dockerfile I used to build it (also might be useful if you are not familiar with docker and would like to install it on your machine or as a reference for the package version)

FROM bitxiong/tsn:cuda9_cudnn7

RUN git clone --recursive https://github.com/yjxiong/anet2016-cuhk
WORKDIR /app/anet2016-cuhk
RUN bash models/get_reference_models.sh

RUN cp -r ../lib/ .
RUN cp ../cv2.so .

RUN pip install easydict==1.6
# YouTube API changes often so you may want to install a more recent version of youtube_dl:
RUN pip install youtube_dl==2020.9.20

RUN echo "export ANET_HOME="/app/anet2016-cuhk"" >> ~/.bashrc

[hgminh95]

Thanks, that is really helpful (:

[yagmurshn]

Hi, i had the same trouble cas of dense_flow package which caused me not to run cuhk project on locally. There's no opencv version that it works compatible with dense_flow. Btw cuhk project does not work with python 3 version. My question is after i pull your docker image to my locally and enter inside to container, when i try to run following command prompt in order to see the classification results for plastering.avi, i'm getting the error below, so i cann't run the classification.

python examples/classify_video.py --use_flow data/plastering.avi

root@488aab6d3886:/app/anet2016-cuhk# python examples/classify_video.py --use_flow data/plastering.avi 
WARNING: Logging before InitGoogleLogging() is written to STDERR
E0805 18:02:37.780357    16 common.cpp:138] Cannot create Cublas handle. Cublas won't be available.
E0805 18:02:37.780797    16 common.cpp:145] Cannot create Curand generator. Curand won't be available.
Setting device 0
F0805 18:02:37.781041    16 common.cpp:190] Check failed: error == cudaSuccess (35 vs. 0)  CUDA driver version is insufficient for CUDA runtime version
*** Check failure stack trace: ***
Aborted (core dumped)

Could the cuda driver version inside the docker image be outdated or incompatible?

[yagmurshn]

The error i still keep getting, when i run following script. python examples/classify_video.py data/plastering.avi

F0807 20:06:21.927471  1948 cudnn_conv_layer.cu:34] Check failed: status == CUDNN_STATUS_SUCCESS (8 vs. 0)  CUDNN_STATUS_EXECUTION_FAILED
*** Check failure stack trace: ***
Aborted (core dumped)

[wssxjtu]

Have you solved this problem yet? I'm having the same problem. @yagmurshn

[wssxjtu]

The error i still keep getting, when i run following script. python examples/classify_video.py data/plastering.avi

F0807 20:06:21.927471  1948 cudnn_conv_layer.cu:34] Check failed: status == CUDNN_STATUS_SUCCESS (8 vs. 0)  CUDNN_STATUS_EXECUTION_FAILED
*** Check failure stack trace: ***
Aborted (core dumped)

I'm having the same problem. @yagmurshn

[yagmurshn]

Regrettably, I've exhausted all local solutions, including using nvidia-docker with the --gpus all parameter, but I continue to encounter this error.

[yagmurshn]

It's quite disappointing that the CUHK paper received such widespread recognition in various articles and even secured the winner rank, yet its code failed to work at all, the code did not function as expected.

[wssxjtu]

It's quite disappointing that the CUHK paper received such widespread recognition in various articles and even secured the winner rank, yet its code failed to work at all, the code did not function as expected.

Regrettably, I've exhausted all local solutions, including using nvidia-docker with the --gpus all parameter, but I continue to encounter this error.

Thank you very much for your reply!