Yolov3: Cannot find --transfer in train.py and how to freeze layers except darknet backbone?

[yoga-0125]

May I ask where is --transfer in train.py for transfer learning? And how to set requires_grad = False for all layers except darknet backbone?

[glenn-jocher]

@yoga-0125 to train from an existing model or backbone simply use --weights model, for example:

python3 train.py --weights darknet53.conv.74

or for better results:

python3 train.py --weights ultralytics68.pt

We've eliminated the transfer flag and do not recommend freezing any layers, as this nearly always results in worse performance.

[yoga-0125]

Thank you very much for your reply!

[adrianosantospb]

python3 train.py --weights ultralytics68.pt

Imagine I have new data and I'd like to improve my production model. When I use python3 train.py --weights myProductionModel.pt command, only is the darknet backbone updated?

[glenn-jocher]

@adrianosantospb I don't understand your question. The --weights command loads the weights you specify into the --cfg you specify. If there is a mismatch it will attempt to load all weights into layers that have matching names and shapes.