您好，实在抱歉，打扰您一下，refineMask代码我跑了好几天啦，但是总是出现一个问题：训练着训练着loss就为Nan啦，我也不知道哪里出了问题，想请教一下。我的环境是：3090的卡，cuda11.1，python3.8，torch1.8，gcc/g++7.5， mmdet2.3.0，mmcv-full1.0.5，COCO2017数据集，sample_per_gpu=1，lr=0.001
DongHappyyy opened this issue · comments
No description provided.
Please upload the running log here.
No description provided.
Please upload the running log here.
已经上传log啦，请多麻烦您给看一看，十分感谢
No description provided.
Please upload the running log here.
已经上传log啦，请多麻烦您给看一看，十分感谢
Only using one gpu cannot reproduce the result. If you need to use one gpu to train the model, decrease the lr to 0.0025 .
No description provided.
Please upload the running log here.
已经上传log啦，请多麻烦您给看一看，十分感谢
Only using one gpu cannot reproduce the result. If you need to use one gpu to train the model, decrease the lr to 0.0025 .
Thank you very much! I will try it right now!
No description provided.
Please upload the running log here.
已经上传log啦，请多麻烦您给看一看，十分感谢
Only using one gpu cannot reproduce the result. If you need to use one gpu to train the model, decrease the lr to 0.0025 .
Thank you very much! I will try it right now!
你好，我的环境跟你一样的，请问你复现代码成功了。我的环境在mmdet2.3.0一直训练不起来。然后我迁移代码到mm2.12.0版本训练了12轮，但是测试模型后没有效果，AP为0