Some questions about BiFPN in FCOS

[Chao Yin]

If I want to reproduce the fcos model using bifpn, how should I modify the configuration file?

1. Use the default bifpn configuration as follows:
   
   The model performance after training is very low, as follows:
   
2. Note that the default configuration is bifpn, and the input feature layer defaults to_ C.MODEL. BiFPN. IN_ Features = ["res2", "res3", "res4", "res5"], should it be changed to ["res3", "res4", "res5"]? I'm testing, but I'm in a hurry. Can you give me some suggestions? thank you