Detection head re-implementation for PyNetsPresso compatibility

[illian01]

Description

Please include a summary in English, of the changes in this pull request. If it closes an issue, please mention it here.

Closes: ~

You should link at least one existing issue for PR. Before your create a PR, please check to see if there is an issue for this change.
PRs from forked repository not accepted.

Change(s)

• Separate model backbone and head
• Revert NMS code to receive various batch size
• Fix metric
• Add validation loss computation

Changelog

If you PR to dev branch, please add a brief summary of the change to the Upcoming Release section of the CHANGELOG.md file and include a link to the PR (formatted in markdown) and a link to your github profile.

For example,

- Added a new feature by `@myusername` in [PR 2023](https://github.com/Nota-NetsPresso/netspresso-trainer/pull/2023)

Code Formatting

If you PR to either master or dev branch, you should follow the code linting process. Please check your code with lint_check.sh in ./scripts directory.
For more information, please read the contribution guide in CONTRIBUTING.md.

[deepkyu]

@illian01
Thx for the update.
It seems that the metric calibration is updated as batch-wise calibration.
In later, it would be better not to use for-loop in mini-bach.

Plz check the other comments.

[deepkyu]

@illian01

If we override task-specific save_checkpoint function, it may be better to express as abstractmethod in base pipeline.
Someone may be hard to find where it is defined..

[illian01]

@deepkyu
I overrided save_checkpoint method because we have to separate backbone and head for detection task.
Thus, this overrided method don't need if we serve fx model with integrated one.
I think it would be better to remove this method when we can compress and benchmark the integrated detection task model.
I will fix save_checkpoint to abstractedmethod if we determined to serve detection model as is (backbone and head separately) through internal discussion.

[deepkyu]

@illian01

I updated to call with calc, instead of __call__. Originally, it called with the builtin __call__ as you modified, but in later it is hard to catch whether the metric_factory is method of the class or attribute of the class.

I added calc in later, but leave __call__ just for backward support.

[illian01]

@deepkyu
I got it. I will fix to call self.metric_factory.calc method.