If I want to change the backbone of the model, do I need to use "load_from = None"

[DishantMewada]

Hi,

I want to use faster_rcnn with swin backbone. I have changed the backbone in the config file. When I change 'load_from = None', I get very low results like 0.04 mAP.
But when I use 'load_from = faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth', I do get comparable results to r50 like 0.3 mAP. Doing this, throws many warnings like unexpected keys, size mismatch etc.

My question is what is the right way to change the backbone and train the model?
Does 'load_from = None' mean we are training the model from scratch?
My dataset is small with 3k images and 5 classes. Do I need to train the model first on the entire coco dataset using 'load_from = None' and then use that model on my dataset (transfer learning)?

Thank you so much.