Model freezing

[Hanminghao]

A very good job you mentioned that you froze the Image encoder and Text encoder for the pre-trained model, but I don't seem to find the corresponding model freeze code in the code. Looking forward to your reply.

[Jiangbo-Shi]

Thanks for your question. The image encoder is utilized to extract the patch features in an offline way; The parameter of the text encoder is not updated. Only the defined prompt embeddings are learnable. For more specific details, please refer to this work CoOp.