Trained model weights only for backbone ?

[LedaguenelArthur]

Hi,

I am doing some experiments using 3D-RetinaNet and wanted to use the trained weights on Road that you share with the link : https://drive.google.com/drive/folders/1tOwQtQD3HWiTTp_ZgPCEWd4W-UKiglbt?usp=sharing

However these weights do not load properly on the whole network (missing keys and unexpected keys) using:
net.load_state_dict(torch.load(path))
I get for instance "anchors.cell_anchors.0" in missing keys and "module.anchors.cell_anchors.0" in unexpected keys.

The only way I could manage to make them work is by doing:
net.backbone.load_my_state_dict(torch.load(path))

But this means that regression and classification heads are not initialized with trained weights and the model is not directly usable at test time.

Anything I am missing or doing wrong ?

Best regards,
Arthur