OTX export for visual prompting task

[sungchul2]

Summary

This PR includes:

• enable otx export
  • export dividing SAM into image encoder and decoder (prompt encoder + mask decoder)
• enable otx eval with OpenVINO IR
• refactoring & bug fix

I followed a flow of OpenVINOSegmentationTask for OpenVINOVisualPromptingTask, so they look so similar.
But, for visual prompting task, there are two models, image encoder and decoder.
I think this point is just difference between both tasks.

TODOs

• otx export
  • ONNX
    
  • IR
• OpenVINO inference
  • sync inference (as default)

    $ otx eval
    2023-07-06 10:52:00,838 | INFO : Avg time per image: 4.619890505930892 secs
    2023-07-06 10:52:00,838 | INFO : Total time: 124.73704366013408 secs
    2023-07-06 10:52:00,838 | INFO : Visual Prompting OpenVINO inference completed
    2023-07-06 10:52:00,839 | INFO : Computing mDice
    2023-07-06 10:52:03,808 | INFO : mDice after evaluation: 0.9487857719406433

  • async inference
    • It showed performance degradation, need to check
• Unit test
• Integration test
• rebase

How to test

Unit test (local)

$ tox -vv -e unittest-all-py310 -- tests/unit/algorithms/visual_prompting
====================== 165 passed, 110 warnings in 33.36s ======================

Integration test (local)

$ tox -vv -e tests-all-py310 -- tests/integration/cli/visual_prompting/test_visual_prompting.py
=================== 8 passed, 1 warning in 122.65s (0:02:02) ===================

Checklist

• I have added unit tests to cover my changes.​
• I have added integration tests to cover my changes.​
• I have added e2e tests for validation.
• I have added the description of my changes into CHANGELOG in my target branch (e.g., CHANGELOG in develop).​
• I have updated the documentation in my target branch accordingly (e.g., documentation in develop).
• I have linked related issues.

License

• I submit my code changes under the same Apache License that covers the project.
  Feel free to contact the maintainers if that's a concern.
• I have updated the license header for each file (see an example below).

# Copyright (C) 2023 Intel Corporation
# SPDX-License-Identifier: Apache-2.0

[jaegukhyun]

I left some comments. Frankly there are many codes that is not fully understood.
I think we should revisit this after FF. Before merging this PR I want to ensure there's no big accuracy drop before and after export. Do you have some data for this? I want to make sure of this since preprocessing and postprocessing of OV model is implemented in this PR, not from ModelAPI.

[sungchul2]

I left some comments. Frankly there are many codes that is not fully understood. I think we should revisit this after FF. Before merging this PR I want to ensure there's no big accuracy drop before and after export. Do you have some data for this? I want to make sure of this since preprocessing and postprocessing of OV model is implemented in this PR, not from ModelAPI.

From this version, exported model showed slight mDice performance drop from 0.9503 to 0.9488.
As you said, it may be due to not using modules from modelapi.
I'm proceeding to change preprocessing and postprocessing to using modelapi in the next PR (#2311).