search

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).
MIT License
1.73k stars 784 forks source link

warmup epochs #293

Open mohamedabdallah1996 opened 6 years ago

mohamedabdallah1996 commented 6 years ago

please any one can clarify:

  1. what is the warmup epochs ?
  2. when should I use it ?

thanks

sugi-chan commented 6 years ago

Hi,

As I understand for the projects I have done using this repository I use "warmup" initially to let a model adjust to a new dataset. It gets talked about in some other issues. Basic pipeline has 2 steps, first is to run training with warmup for 3-5 epochs initially to let it adjust (you can save those weights), then once it finishes you can use those weights to do standard training. Hope that helps a bit.