Refactor darknet backbone

[zhiqwang]

🚀 Feature

Refactor darknet backbone as separate modules

Motivation

To make the model structure more controllable, I decided to abandon the way of loading model with a yaml config files as Ultralytics and refactor the darknet backbone as separate modules.

Especially, the parse_model() as below containing the DarkNet, BackboneWithPAN, and PathAggregationNetwork modules, so first I will make this modules more visible and clear.

https://github.com/zhiqwang/yolov5-rt-stack/blob/9104887c4cf10061585eb5b8982270790437013d/models/backbone.py#L79

Then, this modifications would change the keys of checkpoint in state_dict, it's also necessary to revise the checkpiont update scripts.

Pitch

• Refactor DarkNet
• Refactor BackboneWithPAN
• Refactor PathAggregationNetwork
• Modify checkpoint update scripts

Alternatives

I found several works on this, like:

• https://github.com/sporadicCactus/pytorch_headless_YOLOv4
• https://github.com/RangiLyu/nanodet/blob/main/nanodet/model/fpn/pan.py

Additional context

Also follow the philosophy of torchvision especially for:

• fpn
• BackboneWithFPN
• mobilenet_v2
• mobilenet_v3

[zhiqwang]

The model structure can be viewed as following:

Copyright statement ©: This article is the original article of the blogger and follows the CC 4.0 BY-SA copyright agreement. Please attach the original source link and this statement for reprinting.
Link to this article: https://blog.csdn.net/Q1u1NG/article/details/107511465

[zhiqwang]

More architecture visualization:

• Overview of model structure about YOLOv5 ultralytics/yolov5#280 (comment)

• YOLOv5 version 6.1 architecture ultralytics/yolov5#7160