Turning a pretrained model into feature extractor backbone for efficientDet

[AnhMinhTran]

Hi, after training a model, I would like to to manipulate it into a backbone for the EfficientDet using timm. I do know that you can use features_only = True, but how do I load my previous trained weight into the model? Do I use checkpoint_path with model_best.pth.tar or do I need to clean check point then model.load_state_dict? I have looked at the documentation but haven't seen info that talked about this

(Edit)
These lines work:

    model1 = timm.create_model('efficientnet_v2s',
                               #checkpoint_path='models/feature_extract.tar',
                               num_classes=7,
                               features_only=True)

    model1 = timm.create_model('efficientnet_v2s',
                               checkpoint_path='models/feature_extract.tar',
                               num_classes=7)
                               #features_only=True)

But not this one:

    model1 = timm.create_model('efficientnet_v2s',
                               checkpoint_path='models/feature_extract.tar',
                               num_classes=7,
                               features_only=True)

[rwightman]

@AnhMinhTran it's a known limitation, features_only modifies the backbone and checkpoint loads after the modification (I have a TODO to clean that up but not a priority). Please see the last comment in this for a workaround #488