User Experience Issues with Auto Annotation Run Model Script

[benhoff]

If an auto annotation script fails for someone, it is currently difficult to debug with a casual user.

Ref: #896

Ref: There's a conversation in gitter where Boris walked through with someone how to copy relevant files from their local machine to docker so they could use the run_model.py script.

This is a current deficiency with the auto annotation run_model.py script. The script is setup to locally debug, but it would be better to be a more general user-facing debugging script.

Need to implement a second interface where users can identify:

1. The relevant model
2. The relevant task

and get a traceback that they can share for easier debugging.

This would solve the issue of explaining to users how to copy files from their local machines to docker. I also had a user (with a functioning CVAT docker installation), try to run run_model.py script but did not correctly activate OpenVINO's setupvars.sh on their local machine. This lead to importing OpenVINO errors that were unrelated to the problem that the user was trying to solve (poor user experience).

In order to solve this:

Lines 21-24 of the auto annotation script will need to be changed so that the py, json, xml, and bin files are no longer required. https://github.com/opencv/cvat/blob/32027ce884c0584015874c4ae99ba7f53ffb46c0/utils/auto_annotation/run_model.py#L21

Instead the script would enforce either the passing of the aforementioned 4 variables, or a "model name"/"task reference" number.

Would need to explore what implementation (Docker based, REST based, etc) would make the most sense.

[benhoff]

Alternatively, could setup Auto Annotation's Load Model UI, so that it could use an existing task to test/debug the model with. This would catch significantly more errors presented back to the user on upload, as testing currently uses an all black image.