VisCPM icon indicating copy to clipboard operation
VisCPM copied to clipboard

[ICLR'24 spotlight] Chinese and English Multimodal Large Model Series (Chat and Paint) | 基于CPM基础模型的中英双语多模态大模型系列

Results 8 VisCPM issues
Sort by recently updated
recently updated
newest added

我尝试在两张A800 80G 上进行微调,cuda 11.8,报了如下错误,请问要怎么解决呢 ``` Traceback (most recent call last): File "/home/hpc/LAB-data/disk1/VisCPM/./finetune/ft_viscpm_chat/train_viscpm_chat.py", line 210, in main() File "/home/hpc/LAB-data/disk1/VisCPM/./finetune/ft_viscpm_chat/train_viscpm_chat.py", line 206, in main train(model, args) File "/home/hpc/LAB-data/disk1/VisCPM/./finetune/ft_viscpm_chat/train_viscpm_chat.py", line 91, in train...

支持连续不间断图片吗,比如第一轮输入图片和文本,过了几轮再输入图片,看代码好像只支持一次图片?示例也都是只支持一张图片进行多轮对话

求助: 我这样调用的, from VisCPM import VisCPMChat from PIL import Image model_path = './openbmb/VisCPM-Chat/pytorch_model.zhplus.bin' viscpm_chat = VisCPMChat(model_path, image_safety_checker=True) 我模型加载的路径对吗? 加载模型中,GPU到811M左右时就会被自动杀死,显卡够用,这是怎么回事呢?

![image](https://github.com/OpenBMB/VisCPM/assets/15043401/36a51bec-b745-4cf6-83af-3fc3300f3333) ![image](https://github.com/OpenBMB/VisCPM/assets/15043401/268452e8-5657-4aeb-b7d4-2c44311b81b8) why the paint model generate nothing whatever prompt is input?

Dear authors, I meet AttributeError: 'VisCpmBeeTokenizer' object has no attribute 'encoder' when running the following code: ```python #!/usr/bin/env python # encoding: utf-8 from diffusers import DiffusionPipeline from transformers import AutoModel...

run_viscpm_chat_ft.sh中--img_path和initializer.py中--image_path的微调图片路径配置名不一致,要么都是--img_path要么都是--image_path,不然微调会报路径错误

环境: python310 torch 2.2.1 cuda 12.3 运行代码: from VisCPM import VisCPMPaint painter = VisCPMPaint('./models/VisCPM-Paint', image_safety_checker=True, prompt_safety_checker=True, add_ranker=True) # 默认对输入的文本和输出的图片进行安全检查,默认开启重排序 image = painter.generate('人闲桂花落,月静春山空') # 对应上图第一行第二张图片 image.save('./test.png') 报错如下: Traceback (most recent call...