apex version about pcme

[SanghyukChun]

Hello, you can use the official docker from dockerhub for reproducing my environment

docker pull sanghyukchun/pcme:torch1.2-apex-dali

https://hub.docker.com/layers/sanghyukchun/pcme/torch1.2-apex-dali/images/sha256-cd64cd5dcb06f8036e7d531bb9549d00a67005384fa7aa2b2e55ca773a7af2f0?context=explore

Also, when I checked pip freeze | grep apex
I got apex==0.1

Hope this answer can help you

[SanghyukChun]

BTW, you can omit apex if you do not use the FP16 option. Just set use_fp16: False from the configuration
https://github.com/naver-ai/pcme/blob/main/config/coco/pcme_coco.yaml#L61

or, simply

python train.py <default options> --train__use_fp16 0

[hellerguyh]

Appreciate your assistance!
After some experimentation, I found that using torch==1.8.0+cu111, torchvision=0.9.0+cu111, and torchtext==0.5.0 resolves the issue. Upgrading to this version of torch enabled integration with apex, although using the --train__use_fp16 option also provided a viable solution. The specific apex commit id that worked for me was ae75763.

On another note, I recommend adding pyyaml==5.4.1 to the requirements file. Newer versions of pyyaml alter the API, making them incompatible with the project code.