Why are users able to ignore the different shape of detection head between the pretrained weights and their own model when training custom datasets?

[zqf30]

I found this tip in train_custom_data. I have searched the code for many times, but I do not get a reasonable explanation, which made me more curious about this tirck. Did anyone have ideas about how authors implement this function in code?
ps: if you are also interested in this question, I am very appreciated if you can contact and discuss with me at [email protected].