Training of Q-former last layer and CLORI module

[xuliwalker]

Thanks for providing such a great work! I have a question about your training process: did you train the last linear projection layer of Q-former and the proposed CLORI module separately or together? From your paper, it seems like these two parts are trained separately.