search

rese1f / MovieChat

[CVPR 2024] MovieChat: From Dense Token to Sparse Memory for Long Video Understanding
https://rese1f.github.io/MovieChat/
BSD 3-Clause "New" or "Revised" License
534 stars 41 forks source link

Ambiguous vicuna version #3

Closed wanghao-cst closed 1 year ago

wanghao-cst commented 1 year ago

Awesome work.

May I make sure that the vicuna weight is v0 not v1? There may be a typo as shown below. image

wanghao-cst commented 1 year ago

Additionally, I see that you mentioned training free; can I use vicuna V1 weight? Is there any impact if I switch to the v1 version?

Espere-1119-Song commented 1 year ago

The vicuna weight is v0 not v1.1. The v1.1 you see is a naming error, I will fix it. MovieChat is pretrained on llama-7b and vicuna-v0. If you switch to the v1 version, it may cause parameter mismatch and the model effect is not good.

wanghao-cst commented 1 year ago

The vicuna weight is v0 not v1.1. The v1.1 you see is a naming error, I will fix it. MovieChat is pretrained on llama-7b and vicuna-v0. If you switch to the v1 version, it may cause parameter mismatch and the model effect is not good.

Thank you for your reply.