Which layers in the Efficientdet are used for classification of class scores?

[santhoshnumberone]

I am trying to use transfer learning on pre trained model of zylo117 Yet-Another-EfficientDet-Pytorch well documented repo, to train my custom data.

I find that index 77/Cell Phone is the class that my custom data is being detected as in most of the cases.

I though why not freeze the feature extraction layers using requires_grad=False and train only the prediction of the class final layer with requires_grad=True while replacing the final layer 90 classes weight with only one class 77 weight so that it can be trained to learn the new data.

I can't seem to figure out which is the exact layer where the class scores are present, could any help me out here?

I have raised the same question here as well you can have a look for better understanding Which layer is used for classification of classes? #720

Hope someone can guide me here