CUDA out of memory when training using resneSt200

[zhouzhengguang]

Hi, thanks for your work. I use 8GPUs with each 12G memory to train the deeplab using resneSt200 on ADE20K with base-size=512/crop-size=480, however it raises "CUDA out of memory", how do you make it work?

[zhanghang1989]

I used Pascal V100 (32 GB) memory GPU to train the model, but I think 16GB should be enough.

For 12GB GPU, you can try using pytorch checkpoint, which is very helpful.