Skip to main content

Using mmclassification; CSPDarkNet, YOLO series

· 2 min read
SeulGi Hong

Episode 3 for my VIPrior challenge journey link!

Situation

Let's improve the detection performance through transfer learning of the backbone. So, I need to switch to mmclassification and train the backbone, but it should be compatible with YOLOX in the existing mmdetection, so I need to train CSPDarkNet.

CSPNet (Cross Stage Partial Network)​

  • CSPNet enables efficient learning by removing gradient computation redundancy, without modifying the existing architecture.
  • Paper Link

image

image

Let's train!​

Code modification: Since CSPNet exists but the DarkNet version is missing, I copied some code from mmdetection.

  • Just the basic version
  • Label Smooth
  • Multiple losses are not commonly used in classification. Refer to mmcls issue reply. link
  • Training results: my mmcls wandb training logs. log
  • However, there was an small mistake when converting the data to coco format, so retraining is required...

Miscellaneous​

Joined the detection team in the challenge in a rush, lacking basic skills (this is my first time doing detection). After the rush is over, I feel that I need to take time to read about the YOLO series...

Materials to read:

  • A helpful article explaining CSPNet: Link
  • DarkNet, YOLO: Link
  • Comparison of the YOLO series: Link