Skip to content

v3.0.0 - YOLOv5 in PyTorch
Compare
Choose a tag to compare
@zhang-dut zhang-dut released this 25 May 14:46
· 22 commits to main since this release
v3.0.0
ee12283

The main body of this project comes from v7.0 - YOLOv5 SOTA Realtime Instance Segmentation of ultralytics/yolov5
This version uses GhostNet as the backbone network, and adds CA(Coordinate attention) attention mechanism in v2.0.0, placing it in 4 different positions.

Important Updates

  • Replacing the backbone network: CSPDarknet53 -> GhostNet, to make it lightweight🎉NEW
  • Add CA attention mechanisms and place them in 4 different positions🎉NEW
    • After three Concats near the Head, 17/21/25Layer.
    • After the first Concat, 13Layer.
    • Before SPPF, 9Layer.
    • Head last, 24Layer.
  • Using VOC2007 dataset for training.

Specific Changes

  • Modify "models/parts_yolov5s-ghostnet_ca(or ca2, ca3, ca4).yaml"🎉NEW
    • After three Concats near the Head, 17/21/25Layer in "models/parts_yolov5s-ghostnet_ca.yaml".
    • After the first Concat, 13Layer in "models/parts_yolov5s-ghostnet_ca2.yaml".
    • Before SPPF, 9Layer in "models/parts_yolov5s-ghostnet_ca3.yaml".
    • Head last, 24Layer in "models/parts_yolov5s-ghostnet_ca4.yaml".
  • Modify "train.py"
    • line 441: initial weights path -> weights/yolov5s.pt
    • line 442: model.yaml path -> models/parts_yolov5s-ghostnet_ca(or ca2, ca3, ca4)🎉NEW
    • line 443: dataset.yaml path -> data/parts_voc2007.yaml
Assets 2
Loading