[ENHANCE] Bring mmdeploy to action recognition model export & Test optimization of action tasks

[jaegukhyun]

This PR includes

1. Bringing mmdeploy to action recognition model export
2. Test optimization of action tasks

Why this should be changed
Unlike other algorithms in otx, action tasks use manual export procedure, not mmdeploy. For unity and convenience this PR brings mmdeploy as a framework to export action algo models.
And also, action tasks cli test train models for 2 epochs, so it requires more time than other algorithms. Therefore this PR reduces it to 1

Changes

1. [otx/algorithms/action/adapters/mmaction/models/detectors/fast_rcnn.py]:
   • Change static pooling method: Unfortunately there is an issue regarding to process 3D pooling in openvino with pytorch 1.13(pytorch 1.12 is fine). Therefore this PR change pooling methods to manual method(avgpool -> torch.mean, maxpool -> torch.max)
   • Change building method 2d person detector: mmdeploy requires to build model using its processor so that the monkey patched function works properly. Therefore method to add 2d person detector is changed
   • Add rewriter flags to forward_infer function and unify forward path to simple_test: From using function rewriter in mmdeploy, the forward_infer function can act like forward function during deployment. And this PR unify the forward route so that AVAFastRCNN use same forward path, which is monkey patched in mmdeploy, with 2d person detector.
2. [otx/algorithms/action/adapters/mmaction/utils/export_utils.py]
   • Create new class Exporter: Previous codes for exporting model is designed as procedural programming. This PR changes as OOP from bringing new class Exporter.
3. [otx/algorithms/action/tasks/inference.py]
   • As the exporting methods changed, the method to call exporting method should be changed
   • Add _initialized_post_hook to bring deploy_cfg of task
4. Others
   • deployment.py are added to action classification and action detection model's template directories.
   • Action integration test run training for 1 epoch and remove skip flag for action detection export test

Known Issues

• In pytorch 1.13 3d pooling do not work in openvino ==> Soloved by temporal manual method

• In pytorch 1.13 Two stage openvino models return empty inference result ==> Will be solved in CVS104657

• Action task do not support dump_features(feature vector, saliency map) ==> Will be treated in another PR

• Bring mmdeploy to mmaction's model

• Test optimization for action tasks

• Change unit tests for this PR

[jaegukhyun]

@cih9088 Could you review overall structure for linking between action tasks and mmdeploy framework, and deployment.py for each action models? Since I'm not familiar with mmdeploy, your review would be helpful. Minor comment will be fixed in this PR and large suggestion such as changing overall architecture will be reflected in another PR after merging this PR

[yunchu]

@jaegukhyun please set the milestone for this PR. I guess it would be included to the next release (v1.1.0), right?

[jaegukhyun]

@jaegukhyun please set the milestone for this PR. I guess it would be included to the next release (v1.1.0), right?

Thank's for the remark, I set it as 1.1.0

[codecov-commenter]

Codecov Report

Patch coverage: 93.24% and project coverage change: -0.03 ⚠️

Comparison is base (0f6616e) 80.54% compared to head (d78b2c4) 80.52%.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #1848      +/-   ##
===========================================
- Coverage    80.54%   80.52%   -0.03%     
===========================================
  Files          481      483       +2     
  Lines        32998    32943      -55     
===========================================
- Hits         26579    26527      -52     
+ Misses        6419     6416       -3     

Impacted Files	Coverage Δ
...n/configs/detection/base/base_detection_dynamic.py	0.00% <0.00%> (ø)
...on/configs/detection/base/base_detection_static.py	0.00% <0.00%> (ø)
...on/adapters/mmaction/models/detectors/fast_rcnn.py	100.00% <100.00%> (ø)
...hms/action/adapters/mmaction/utils/export_utils.py	100.00% <100.00%> (+6.71%)	⬆️
otx/algorithms/action/tasks/inference.py	89.16% <100.00%> (+0.69%)	⬆️
...hms/detection/adapters/mmdet/utils/config_utils.py	93.29% <0.00%> (-1.22%)	⬇️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[cih9088]

I have noticed that task structure is pretty different from cls/det/seg. Would cls/det/seg be refactored in a way where action implemented task?
Additionally, otxdeploy requirement should be added into requirements/action.txt

[jaegukhyun]

@cih9088 Thank you for your detailed review, I reflected almost of your review. Currently action task do not use mpa side code, therefore I create own export codes. After refactoring mpa codes, I'll make another PR for unifying exporting process

[JihwanEom]

LGTM, thank you for hard work!