SingL3

@mfiro http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/support/eval_script/eval_tools.zip or http://shuoyang1213.me/WIDERFACE/support/eval_script/eval_tools.zip

I think the code of wider face is correct. The only different is that you should read all the detection results with confidence and normalize them using (confidence - min)...

Nope. My loss just increased to something like 10 million and never decrease. And this loss is not reasonable. Actually, I am trying to train a model that can accept...

@13070151771 意思是只要把 roidb = filter_for_training(roidb) 去掉就可以吗？ 我看代码，大多数都对无目标图可以处理，但是有少部分似乎就无法处理，改完之后loss经常突然增大

Actually, I have modified it to accept these images and now I am testing it.

@akshitac8 you will have to modify the data loading and the label assignment part of anchor and roi.

不好意思，对这个是新手，在VS里用调试器运行，过了很久很久，会弹出一个窗口说： 调试器的工作进程(msvsmon.exe)意外退出，调试将中止。 从VS自带的诊断工具里看，进程内存直接到了6个多G，然后就卡住了，我电脑内存只有8G。

It may raise error in different process: ``` Compiling lazy_static v1.4.0 Compiling openssl v0.10.63 Compiling mime v0.3.17 Compiling base64 v0.21.7 Compiling bitflags v2.4.2 Compiling version_check v0.9.4 Compiling strsim v0.10.0 Compiling...

@andreaskoepf Yes, at least latest version last week and deepspeed. I am using 8xA100(80G) with cuda 11.7. I have tried reducing pretrain datasets [here](https://github.com/LAION-AI/Open-Assistant/blob/main/model/model_training/configs/config.yaml#L114)(only `alpaca_gpt4` is reserved) and it can...

I try this PR #4910 and I think it fix this.