changing anchor size in Retinanet can affects freeze layers !

[mzalaki00]

I want to detect some small objects about 200 pixel and I wonder if changing anchor size in RetinaNet config file like
cfg.MODEL.ANCHOR_GENERATOR.SIZES = [[16, 32, 64, 128,512]]
will make the model perform better in transfer learning or the model will require to be trained from scratch? I am using pre-trained models on the Object detection model zoo.