the pretrained model weights about sptnet

[whj363636]

Hi @llj110

Thanks for your interest. The backbone and projector components have been wrapped in a Sequential module called "classifier," while the prompter and classifier are further grouped within another Sequential module called 'model'. Therefore, if your intention is to solely reload the weights for the 'classifier' component, rather than the entire 'model', in both train.py and eval.py, you can maintain the code and simply select the last module from the pretrained model. This module contains the full weights of the backbone and projector.

[llj110]

classifier = nn.Sequential(backbone, projector).cuda()
state_dict = torch.load(args.pretrained_model_path, map_location='cpu')
classifier.load_state_dict(state_dict)
Thanks for your reply. My question is how to load the pretrained model parameters for the classifier using dino_vitbase16_pretrain.pth. The model I downloaded can only load the backbone parameters, but the parameters for the DINOHead cannot be loaded. This issue persists even when I use dino_vitbase16_pretrain_full_checkpoint.pth, which contains the weights for both the backbone and the projector head.

[llj110]

I can just run this code by adjust the code to
state_dict = torch.load(args.pretrained_model_path, map_location='cpu')
backbone.load_state_dict(state_dict)
classifier = nn.Sequential(backbone, projector).cuda(). Therefore, I would like to ask if it is possible to provide the pretrained weights for the classifier.

[whj363636]

I see. Please refer to #4.

[llj110]

Thanks for your valuable reply. I will try to obtain the pretrained SimGCD model