TF_annotation does not use a GPU with the CUDA component installed.

[azhavoro]

How to reproduce:
build docker-compose -f docker-compose.yml -f components/tf_annotation/docker-compose.tf_annotation.yml -f components/cuda/docker-compose.cuda.yml up -d --build and run tf_annotation. Inference will be runned on CPU.

Seems the tensorflow_gpu package is overridden by tensorflow from datumaro requirements.

[itsharsh]

Is this the reason for slow processing of automatic annotation?

[tealeeseng]

I changed tensorflow version to 1.12 and GPU runs. tensorflow 1.13.1 seems to be only running fine with cuda 10 as per
https://stackoverflow.com/questions/50622525/which-tensorflow-and-cuda-version-combinations-are-compatible

I also encountered error message "ImportError: libcublas.so.10.0: cannot open shared object file: No such file or directory" under tensorflow-gpu 1.13.1", therefore, i change the tensorflow version to 1.12.

For your quick change,
grep -R "=1.13" *

and update files accordingly. I also commented away the line at datumaro/requirements.txt, as it is installing CPU version too.

datumaro/requirements.txt:# tensorflow==1.13.1

[itsharsh]

@nmanovic
I updated the cvat repo and built it again using:
docker-compose -f docker-compose.yml -f components/tf_annotation/docker-compose.tf_annotation.yml -f components/cuda/docker-compose.cuda.yml up -d --build

and getting the following error:

[azhavoro]

@itsharsh Hi,
at the moment tf_annotation with gpu support is broken.Please see #1133 for details.

[itsharsh]

Okay, got it,
will wait for the fix, thanks...

[tealeeseng]

#1138 works for me.