├── cache
│ ├── ckpt
│ │ ├── bert-base-uncased
│ │ ├── blip2
│ │ │ ├── blip2_pretrained_flant5xxl.pth
│ │ ├── eva
│ │ │ ├── eva_vit_g.pth
│ │ ├── Qwen7B-chat
│ │ │ ├── config.json
│ │ │ ├── ...
├── ...
├── lavis
│ ├── output
│ │ ├── pp_7b_video
│ │ │ ├── pretrain
│ │ │ | ├── global_step2181
│ │ │ | | ├── model.pth
│ │ │ ├── sft_video
│ │ │ | ├── global_step2005
│ │ │ | | ├── unfreeze_llm_model.pth
请将模型权重下载后都放在
cache/ckpt
下
mkdir cache
cd cache
mkdir ckpt
mkdir dataset
(a) eva vit-g
wget https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/eva_vit_g.pth
(b) bert-base-uncased
huggingface,下载如下的文件即可
(c) blip2_pretrained_flant5xxl
blip2_pretrained_flant5xxl.pth
wget https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth
(建议放入 lavis/output/pp_7b_video/pretrain/global_step2181
)
在本仓库的release里放有checkpoint,可以直接下载
wget https://github.com/Coobiw/MiniGPT4Qwen/releases/download/MPP-Qwen-Next_ckpt-and-data/ckpt-and-data.zip
unzip ckpt-and-data.zip