Encoder weights from local directory

[Foundsheep]

Hi,

It might have been dealt with somewhere, but I couldn't find any post about it.
Bascially, I'm trying to create a model instance without depending on the external network, which means when I instantiate a pre-trained model, it seems like downloading weights from somewhere else.

Can I have its weights in my local machine and instantiate using it?
By the way, I'm using SegFormer just to make things clear

[qubvel]

Hi @Foundsheep, to prevent loading weights from remove you can specify encoder_weights=None. Then you can load from a local directory with the standard torch approach as torch.load and model.load_state_dict

[Foundsheep]

Hi @qubvel,

Thanks for the quick reply.
Then where could I find the model's weights file? If it's somewhere in the documentation, sorry about bothering I couldn't find it.

[qubvel]

You can download it while you have an internet into the model, then save with torch.save + model.encoder.state_dict()