summary

• Python version should be added as version requirement
  • Mine worked with 3.7.6 (below 3.8 should be fine?)
• Will updating numpy from 1.14.2 to 1.14.5 in ernie-vil requirements work with the modules?

details

to reproduce

• python version
  • Running py-bottom-up-attention/requirements.txt, No matching distribution found for torch shows
  • Following this , it worked after changing python version of virtual environment to 3.7.6
• Following SCORE_REPRO.md, version conflict shows as follows
  • Installation succeeded, not sure if it's compatible with all the modules

The conflict is caused by:
    The user requested numpy==1.14.3
    scipy 1.2.1 depends on numpy>=1.8.2
    opencv-python 3.4.2.17 depends on numpy>=1.14.5

notes

• Other than the issues above, #6 helped
• Other than the two above, messages related to dependencies pip showed me are all ignored
• Am trying to follow SCORE_REPRO.md. Will try to keep the issue updated

In addition, according to the official document, looks like detectron2==0.1 is incompatible with torch==0.6? The former is installed by cd vilio/py-bottom-up-attention; python setup.py build develop via cloned detectron2 source, while the latter is required by top-directory requirements.txt. If this version requirement is not strict, upgrading detectron2 to later version or downgrading torch to 1.4 would be helpful I guess. FYI error shows something like "ImportError: cannot import name '_C'". together with Undefined C++ symbols (e.g. "GLIBCXX..") or C++ symbols not found. - see the official q/a

Also, for detectron2 import error I needed to copy detectron2_mscoco_proposal_maxnms.py to different location and change the argument of hm_feats.bash pointing that python script to new path. Refer to aforementioned official detectron2 q/a

FastRCNNOutputs was removed in detectron2 0.6. In my case downgraded to 0.5. Also, transform_gen should be replaced with aug to use with 0.5/0.6 (reference)

Decided to use module-friendly environment