search

RuiyangJu / Fracture_Detection_Improved_YOLOv8

ICONIP 2024
MIT License
75 stars 23 forks source link

YOLOv8-RESCBAM implementation #5

Closed haitamrarhai closed 3 months ago

haitamrarhai commented 3 months ago

Hello, I've followed the modification you have done to the ultralytics repo to modify my instance segmentation model adding the RESCBAM in the nn/task.py, nn/modules/conv.py, the init.py and the yaml file in the cfg. But the number of params that this added on the model is more than what you have reported in yoyr paper. In particular : 322.3 GFLOP for m model , 140.3 GFLOP for s model and 42.9 GFLOP for the n model. Do you have any idea why it is so big th enumber of GFLOP, how can i decrease this number to be able to perform faster training.

Thank you in advance.

Here is my git rep where i did these changes, i added even a P2 layer buut even without it i have high values of GFLOP, and I am sure that is the implementation of RESCBAM that does this (https://github.com/haitamrarhai/ultralytics_2.0/tree/haitamrarhai-patch-1/).

RuiyangJu commented 3 months ago

Hello,

I think you can check the following parts first:

  1. Before adding ResCBAM, is your model the same size as YOLOv8?
  2. Is the ResCBAM you designed same as ours (including the number of layers and channels, etc.)?
  3. Are the numbers of ResCBAM you added the same as ours (the yaml file in the cfg)?

Thanks.