Error while training question generation about unilm HOT 10 OPEN

Could you provide more information to verify whether the apex package was successfully installed?

from unilm.

You asked to not to add certain flags did the same but still getting errors while running decoder script for question generation which looks for apex, (running on CPU),
got segmentation fault

from unilm.

HI @DivyaWadehra,

Have you successfully installed the NVIDIA apex? You can install it following the instructions in README.md. Specifically:

cd $(mktemp -d)
git clone -q https://github.com/NVIDIA/apex.git
cd apex
git reset --hard 1603407bf49c7fc3da74fceb6a6c7b47fece2ef8
python setup.py install --user --cuda_ext --cpp_ext

from unilm.

Yes I did the same, but the error remains

from unilm.

Hi @DivyaWadehra ,

I am wondering whether you are running the code using CPUs? If so, you need to remove the dependency on apex.*.

The instructions in https://github.com/microsoft/unilm#docker could help. The docker makes the env setup way easier.

from unilm.

No I am not running it on CPU, I am using GPU

from unilm.

It seems like the apex was not correctly installed:

PWD_DIR=$(pwd)
cd $(mktemp -d)
git clone -q https://github.com/NVIDIA/apex.git
cd apex
git reset --hard 1603407bf49c7fc3da74fceb6a6c7b47fece2ef8
python setup.py install --user --cuda_ext --cpp_ext
cd $PWD_DIR

from unilm.

I did the same, but didn't get the error resolved

from unilm.

@DivyaWadehra strangley I had forked the version you had here in Jan, and I didnt get the segmentation fault: https://github.com/elnazsn1988/unilm/tree/master/layoutlm
however I used latest commit and am now facing the segmentation fault error, were you able to overcome this?

from unilm.

@DivyaWadehra Update - Using pytorch 1.6 removed all issues and allowed installation of latest commit.

from unilm.