Derek Hsu

我知道为什么在作cluster的时候会有nan了，是因为有些类别我没有写在config裡面，因为他们的选取框太少。我把这些资料都清除了，现在看到的是正确的 h:[4.45, 5.18, 6.81, 10.43, 11.8, 14.21, 17.51, 29.85, 54.53] w:[12.65, 5.12, 110.84, 12.71, 6.89, 21.45, 11.08, 14.0, 14.84] 但是训练的时候rcann_class_loss依然是nan，我应该从什麽地方查起会比较好？

感谢你的回覆，那我知道原因了，因为我原本的图片是非常大的长方形图片，约20000~40000x1833，所以把图形切到只有1833x1833的多张图片，然後再丢入网路，但是1833x1833对网路还是太大，所以我在作的时候只好把图再缩成1024x1024，结果就是你看到的这个样子。我的图形里面，目标很小，但是背景很大，所以一大片背景里面，东西只有小小一个。按照dh,dw的计算方式，log的结果就会变很大。 如果我还要用frcnn的话，我是不是应该再把大图在切小，小到让GT能大於32x32的区域？因为我很难直接透过直接放大把gt都弄到32*32的大小，这样塞不进我伺服器的GPU内。

或者，使用比较低的base size，我看到有论文里面用到更小的basic size，像是2, 4, 8之类的 https://researchportal.port.ac.uk/portal/files/8862349/CCCV_Final.compressed.pdf

https://python.langchain.com/en/latest/modules/chains/index_examples/chat_vector_db.html Actually, I guess this is the answer that mix chat history and knowledge base. When did this come up?

Thanks for your effort. I think this should be enough for my request.

> > @mohitsrivastavazzz @freecoderwaifu @Vynsyc > > All true statements... > > Except [THIS](https://github.com/blue-pen5805/Stable-Diffusion-Webui-Civitai-Helper) fork has working buttons. Haven't had time to look into diff, just know that its working....

You've used a wrong model file. Download again and check you model file.

You just need to download the original model again. Remember to download the "RAW" model, not the html. Just open your download model or check the size, and you will...

I think this issue is done and should be closed?

I created a simple [FastAPI Server](https://github.com/derekhsu/minicpmv-unofficial-openai-server) which is compatible with OpenAI API. Check this and welcome any comment.