部署成功了，但是识别图片结果为什么看起来很离谱

[kimiller]

识别结果为什么这么离谱呢，是一个小白按着教程部署的，不知道哪里有问题

[SnakeHacker]

[silent780]

Same issue, Web demo did a good job, but my local model is just look like a shit, my card is 4090 and model is 13B. Is this just normal or something wrong with my project?

[Korner83]

You might loaded in the wrong weights, I would advise to double check the settings in the yml files to the proper models. Have you followed the steps in the description?

[kimiller]

I didn't follow steps as in the description. I have used below cmd to deploy it into Google Colab, actually I don't know what happened in the cmd,I think I need redeploy it and use another method cmd like this ：

!git clone -b dev https://github.com/camenduru/minigpt4 !wget https://huggingface.co/ckpt/minigpt4/resolve/main/minigpt4.pth -O /content/minigpt4/checkpoint.pth !wget https://huggingface.co/ckpt/minigpt4/resolve/main/blip2_pretrained_flant5xxl.pth -O /content/minigpt4/blip2_pretrained_flant5xxl.pth

!pip install -q salesforce-lavis !pip install -q bitsandbytes !pip install -q accelerate !pip install -q gradio==3.27.0 !pip install -q git+https://github.com/huggingface/transformers.git -U

%cd /content/minigpt4 !python app.py

[kimiller]

@SnakeHacker 哈哈哈笑死我了

[kimiller]

@silent780 not sure about that if it is normal , did you deploy it step by step as toturial providing? I used another method to deployed it,I am going to deploy it step by step as toturial providing

[Korner83]

@kimiller follow the steps and you should call demo.py and not the app.py I think with these paramteres: --cfg-path eval_configs/minigpt4_eval.yaml --gpu-id 0

on windows: python demo.py --cfg-path eval_configs/minigpt4_eval.yaml --gpu-id 0

[kimiller]

@Korner83 Thank you!!!Let me try it

[SnakeHacker]

模型骂人了，学到了国粹。。

[thiner]

我的demo第一次也出现了这种胡说八道的情况。后来发现是我在修改minigpt4/configs/models/minigpt4.yaml文件时，误将llama_model指向了vicuna的delta文件。你不一定是跟我一样的情况，但是建议仔细检查一下配置文件是否正确。

[kimiller]

@thiner 我测了一下minigpt4 在线demo的情况，也是有乱说的时候，识图的功能应该还没这么完善吧，不过这已经做的不错了

[chenchuntan]

有没有最低要求的配置 我电脑不行 CPU是I5 显卡是1060的

[youyuanrsq]

有没有最低要求的配置 我电脑不行 CPU是I5 显卡是1060的

跑7B的模型最低要12G的显存

[YiyangZhou]

I have tried many models and found that mPLUG-Owl https://github.com/X-PLUG/mPLUG-Owl seems to have much stronger visual capabilities than the others, with more accurate image recognition

[kimiller]

@YiyangZhou thank you, tested this one, but it seems to be a copy of MiniGPT-4.

[vateye]

@YiyangZhou thank you, tested this one, but it seems to be a copy of MiniGPT-4.

No, I have tested it. It seems to perform better than miniGPT4 with the smaller model (ViT-L and LLaMA 7B only).

[kimiller]

@YiyangZhou thank you, tested this one, but it seems to be a copy of MiniGPT-4.

No, I have tested it. It seems to perform better than miniGPT4 with the smaller model (ViT-L and LLaMA 7B only).

@vateye yes.I agree with that it's performance better than miniGPT4 . looks like just made some optimizations , no essential changes.

[YuzhouPeng]

https://github.com/X-PLUG/mPLUG-Owl

Could you please tell me what models have you tried? I may want to try these models also. Thank you so much!

[YuzhouPeng]

I have tried many models and found that mPLUG-Owl https://github.com/X-PLUG/mPLUG-Owl seems to have much stronger visual capabilities than the others, with more accurate image recognition

Could you please tell me what models have you tried? I may want to try these models also. Thank you so much!

[YiyangZhou]

I have tried many models and found that mPLUG-Owl https://github.com/X-PLUG/mPLUG-Owl seems to have much stronger visual capabilities than the others, with more accurate image recognition

Could you please tell me what models have you tried? I may want to try these models also. Thank you so much!

minigpt4:https://github.com/Vision-CAIR/MiniGPT-4 llava:https://github.com/haotian-liu/LLaVA open_flamingo:https://github.com/mlfoundations/open_flamingo mPLUG-Owl:https://github.com/X-PLUG/mPLUG-Owl BLIP2: (1)Try it with the transformers package：https://huggingface.co/docs/transformers/index (2)lavis:https://github.com/salesforce/LAVIS

[sssssshf]

有没有终端命令的demo阿 必须要用浏览器吗

[coding-alt]

@thiner 我测了一下minigpt4 在线demo的情况，也是有乱说的时候，识图的功能应该还没这么完善吧，不过这已经做的不错了

部署在 linux 还是 windows 上？我部署ubuntu 上乱码，mathplotlib的问题，字体库都装全了也无果。

[xueweiprince]

我也试了下在线demo，识别的结果实在是太烂了

[coding-alt]

图生文可以自己更换实现方案。

[631068264]

SnakeHacker

你是怎样让他说中文的。。。。

[zhouyuezhou1]

@thiner 请问那应该指向哪个文件呢