QwenLM / Qwen2-VL

Qwen2-VL is the multimodal large language model series developed by Qwen team, Alibaba Cloud.
Apache License 2.0
2.18k stars 125 forks source link

7B模型本地效果和modelscope效果不一致 #177

Open an1018 opened 1 week ago

an1018 commented 1 week ago

modelscope测试效果比本地效果更好,请问可能是什么原因导致的: 1)model scope版本:千问2多模态视觉模型-7B体验空间 2)本地使用的github首页推理代码,配置如下:

kq-chen commented 1 week ago

试试看是不是和min_pixels = 256*28*28,max_pixels = 1280*28*28 有关?

image
an1018 commented 1 week ago

@kq-chen 用的默认的,没有设置min和max processor = AutoProcessor.from_pretrained(model_dir) 而且尝试使用web_demo_mm.py测试,结果和本地一样,效果不好。请教一下modelscope哪些参数设置和github代码不一样

kq-chen commented 1 week ago

https://modelscope.cn/studios/qwen/Qwen2-7B-VL-demo/file/view/master?fileName=app.py&status=1

可以看看这个文件,modelscope的设置了min_pixels = 2562828,max_pixels = 12802828