A simple question about loading pretrained model

[ThomaswellY]

I wonder what's the differences between the two methods of paras configuration

1. load_from = '/path/pretrain_model.pth'
   resume = True
2. model=dict(
   ……
   init_cfg=dict(type='Pretrained',prefix='backbone.',
   checkpoint='/path/pretrain_model.pth')
   )
   what's more, how to fine tune specific layers( like full fine-tuning and linear probing )through these two methods

[Tau-J]

load_from will load all weights into the model, while prefix='backbone.' will only load the pretrained weights into the backbone.

[ThomaswellY]

Thanks alot, so the para configurations i listed were all used to train the model(whole or only backbone), then how can i execute fine tuning? I haven't seen any operating guide in doc in mmlab project.

[Tau-J]

emmm.. tbh, currently mmengine did not support freeze layers by config, this PR is under review. I'm afraid that you have to freeze modules yourself in the model.