search

DLYuanGod / TinyGPT-V

TinyGPT-V: Efficient Multimodal Large Language Model via Small Backbones
BSD 3-Clause "New" or "Revised" License
1.24k stars 76 forks source link

On the stage 4，the Q-Former is trained, however the Q-Former should not been trained in the paper #23

Closed TangYuan96 closed 9 months ago

TangYuan96 commented 9 months ago

the freeze is false, https://github.com/DLYuanGod/TinyGPT-V/blob/12b036a34090fb1f06d81f25c388b13db4c21fe3/minigpt4/models/minigpt_v2.py#L68

llizhaoxu commented 9 months ago

Hi

Thanks for your reminder. Stage 4 is actually a test version. In the version we will release in the future, the Q-Former is not trained.

TangYuan96 commented 9 months ago

Thanks for your reply. I get it.