[Sprint] Detection head re-implementation for PyNetsPresso compatibility

[illian01]

Description

• Present NetsPresso Trainer detection model can't pass train -> compression -> retrain -> benchmark process because PyNetsPresso can't compress present detection head.
• In this sprint, the detection head will be re-implemented, and tested by passing train -> compression -> retrain -> benchmark process.

Checklists

Common

• [x] I would create a corresponding branch for this issue from the designated (mostly dev) branch.
• [x] This issue only contains a preceding agreement between project owners.

TODO

• [x] Detection training with various backbone should succeed.
• [x] NP Compressor with trained fx checkpoint should succeed.
• [ ] NP Launcher with trained onnx checkpoint should succeed.

[illian01]

184

There are two issues to solve.

• Compression of fx model with PyNetsPresso is not working. This will be solved after internal discussion.
• Retrain the fx model is not working in #184.

[illian01]

194

To complete this issue, we have to solve these issues.

• Present model has separated backbone and head. To integrate them, compressor should support torch.meshgrid or trainer find work around.
• Present model can't be converted to TensorRT. NonZero operation of onnx not supported by converter's TensorRT version. May need TensorRT update or work around.

[illian01]

We decided followings through internal discussion:

• Leave 2-stage detector to support only training, compression (backbone), and fx retraining.
• We don't support TensorRT for 2-stage detector now, because this has limitation by TensorRT version.

Thus, this issue will be closed after refactoring 2-stage detector, and re-open when we need TensorRT of the model.