Some weights of ViTModel were not initialized from the model

[guanjunwu]

Some weights of ViTModel were not initialized from the model checkpoint at facebook/dino-vitb16 and are newly initialized: ['encoder.layer.0.adaLN_modulation.1.bias', 'encoder.layer.0.adaLN_modulation.1.weight', 'encoder.layer.1.adaLN_modulation.1.bias', 'encoder.layer.1.adaLN_modulation.1.weight', 'encoder.layer.10.adaLN_modulation.1.bias', 'encoder.layer.10.adaLN_modulation.1.weight', 'encoder.layer.11.adaLN_modulation.1.bias', 'encoder.layer.11.adaLN_modulation.1.weight', 'encoder.layer.2.adaLN_modulation.1.bias', 'encoder.layer.2.adaLN_modulation.1.weight', 'encoder.layer.3.adaLN_modulation.1.bias', 'encoder.layer.3.adaLN_modulation.1.weight', 'encoder.layer.4.adaLN_modulation.1.bias', 'encoder.layer.4.adaLN_modulation.1.weight', 'encoder.layer.5.adaLN_modulation.1.bias', 'encoder.layer.5.adaLN_modulation.1.weight', 'encoder.layer.6.adaLN_modulation.1.bias', 'encoder.layer.6.adaLN_modulation.1.weight', 'encoder.layer.7.adaLN_modulation.1.bias', 'encoder.layer.7.adaLN_modulation.1.weight', 'encoder.layer.8.adaLN_modulation.1.bias', 'encoder.layer.8.adaLN_modulation.1.weight', 'encoder.layer.9.adaLN_modulation.1.bias', 'encoder.layer.9.adaLN_modulation.1.weight'] You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.

Hi, is this problem normal and we can ignore it?

As referred in this, the authors said it does not influence model's performance. Does that mean no more checkpoints I should download and I can only follow the guidance in the readme.md?

[bluestyle97]

Yes, you can simply ignore it.