Open junhyoung-ew opened 8 months ago
Hi, thank you for your interest.
The training details, including the training part and frozen part in each stage, as well as the training dataset and optimizer settings, are provided in Section 3 and Section 4.1 of our paper. For Q-former, just set the parameters of the Q-Former trainable and include them in the optimizer. (sorry I didn't get the uniqueness of "Q-former" in your question, maybe a detailed explanation?)
Currently we don't have plans to publish the training code, sorry for that. But personally, I think it's easy to implement based on the provided model code and training details in our paper.
Can you train detail "Style Embedding Extraction"?
Especially, a trainable "Q-Former", how to train this?
Can you publish train code?
Thank you.