PandaGPT部署演示

github：GitHub - yxuansu/PandaGPT: PandaGPT: One Model To Instruction-Follow Them All

1、搭建环境

conda create -n pandagpt python=3.9
 
conda activate pandagpt

创建虚拟环境pandagpt，python3.9，激活环境

git clone https://github.com/yxuansu/PandaGPT.git
 
cd /content/drive/MyDrive/PandaGPT
 
pip install -r requirements.txt
 
pip install torch==1.13.1+cu117 -f https://download.pytorch.org/whl/torch/
 
pip install gradio mdtex2html

下载PandaGPT源码，安装依赖包。

2、下载模型

curl -L https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth -o /content/drive/MyDrive/imagebind_ckpt/_huge.pth

准备ImageBind Checkpoint，保存为/content/drive/MyDrive/imagebind_ckpt/_huge.pth

准备Vicuna Checkpoint，保存为/content/drive/MyDrive/vicuna_ckpt/7b_v0/vicuna-7b-v0，详见

 Llama-7b-hf和vicuna-7b-delta-v0合并成vicuna-7b-v0_Spielberg_1的博客-CSDN博客

pip install huggingface_hub
 
from huggingface_hub import snapshot_download
 
save_dir="/content/drive/MyDrive/pandagpt_ckpt/7b"
repo_id="openllmplayground/pandagpt_7b_max_len_1024"
 
snapshot_download(repo_id=repo_id,local_dir=save_dir, local_dir_use_symlinks=False)

下载PandaGPT权重，保存到/content/drive/MyDrive/pandagpt_ckpt/7b

还可以从huggingface网站下载模型等。

三、部署演示

cd /content/drive/MyDrive/PandaGPT/code

进入PandaGPT目录，打开/PandaGPT/code/web_demo.py文件，

传入的imagebind_ckpt_path、vicuna_ckpt_path模型路径修改成你的目录

delta_ckpt_path修改为pandagpt权重路径，最后是.pt文件

最后一行，share参数改为True，这样就可以分享链接，在网页上打开了

终于可以运行啦

cd /content/drive/MyDrive/PandaGPT/code
 
 
CUDA_VISIBLE_DEVICES=0 python web_demo.py

我们看一下效果，打开链接，看到页面有四个框，依次上传图片、音频、视频、热力图，

文件上传到模型受网络影响，图片可能要等待几秒，音频和视频要更长一点，有文件上传完成提示就更好了(期待大神更新)。

在[input]框中输入想要的提问，点击[submit]，等到推理出结果，在[chatbot]中对给出回答

传入一张猫咪，提问：“这是什么颜色”    回答：灰色和白色

上传一段雷声，提问：这是什么声音  

上传汉堡图片，提问：举两个生产商

总结：总的来看，pandagpt使用文本和图像对进行训练，能够实现跨模态的输入的，生成文本，还是很让人欣喜的，感谢分享，期待大神继续完善。
 

遇到的问题：

ImportError: LlamaTokenizer requires the SentencePiece library but it was not found in your environment. Checkout the instructions on the installation page of its repo: https://github.com/google/sentencepiece# installation and follow the ones
that match your environment. Please note that you may need to restart your runtime after installation.

pip install Sentencepiece

/mnt/PandaGPT/code/web_demo.py:154: GradioUnusedKwargWarning: You have unused kwarg parameters in Blocks, please remove them: {'scale': 4}
  with gr.Blocks(scale=4) as demo:
/mnt/PandaGPT/code/web_demo.py:157: GradioUnusedKwargWarning: You have unused kwarg parameters in Row, please remove them: {'scale': 4}
  with gr.Row(scale=4):
/mnt/PandaGPT/code/web_demo.py:163: GradioUnusedKwargWarning: You have unused kwarg parameters in Video, please remove them: {'type': 'file'}
  video_path = gr.Video(type='file', label="Video")
/mnt/PandaGPT/code/web_demo.py:167: GradioDeprecationWarning: The `style` method is deprecated. Please set these arguments in the constructor instead.
  chatbot = gr.Chatbot().style(height=300)
/mnt/PandaGPT/code/web_demo.py:171: GradioDeprecationWarning: The `style` method is deprecated. Please set these arguments in the constructor instead.
  user_input = gr.Textbox(show_label=False, placeholder="Input...", lines=10).style(container=False)
/mnt/PandaGPT/code/web_demo.py:173: GradioUnusedKwargWarning: You have unused kwarg parameters in Row, please remove them: {'scale': 1}
  with gr.Row(scale=1):
/mnt/PandaGPT/code/web_demo.py:175: GradioUnusedKwargWarning: You have unused kwarg parameters in Row, please remove them: {'scale': 1}
  with gr.Row(scale=1):
Running on local URL:  http://0.0.0.0:24000

To create a public link, set `share=True` in `launch()`.
Error: no DISPLAY environment variable specified

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 32.00 MiB (GPU 0; 23.70 GiB total capacity; 7.16 GiB already allocated; 12.69 MiB free; 7.22 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

查看显卡显存被占用，杀掉占用进程

Ubuntu下跑Aplaca报错：torch.cuda.0utofMemoryError: CUDA out of memory.解决办法（查看CUDA占用情况&清除GPU缓存）_李卓璐的博客-CSDN博客

FileNotFoundError: [Errno 2] No such file or directory: 'ffprobe'

pip install ffmpeg

感谢：
FileNotFoundError: [Errno 2] 没有这样的文件或目录: 'ffprobe': 'ffprobe' | 那些遇到过的问题

python - FileNotFoundError: No such file or directory: 'ffprobe' FileNotFoundError: No such file or directory: 'ffmpeg' - Stack Overflow

python - FileNotFoundError: [Errno 2] No such file or directory: 'ffprobe': 'ffprobe' - Stack Overflow

Our pilot experiments show that         我们的实验显示

reason over   推理         reasons over the user's input    对用户的输入进行推理

a set of   一套

understanding/reasoning    理解/推理

knowledeg-grouding description   基于知识的描述

multi-turn conversation 多轮对话

general-purpose  多用途，多功能

pilot experiments   中试试验，发布之前的实验

perform complex tasks 执行复杂实验

image description generation  生成图像描述

writing stories inspired by videos   创作受视频启发故事

answering questions about audios   回答关于音频的问题

take multimodal input   simultaneously   同时接收多种模式的输入