adding 'with_cp' option to beit

[ydhongHIT]

The memory cost of beit backbone is huge and you didn't implement the 'with_cp' option to save the memory. However, I add it by myself and the gpu memory doesn't get smaller. Maybe my implement has some problem or there is any reason for this phenomenon?

[MengzhangLI]

Juse because the memory on readme is set cudnn_benchmark = False. You could modify it and try again.

By the way, I really want to support DDRNet in MMSegmentation but do not have too much time to do it, would you mind helping us making a PR to support it? Thanks in advance.

[ydhongHIT]

Juse because the memory on readme is set cudnn_benchmark = False. You could modify it and try again.

By the way, I really want to support DDRNet in MMSegmentation but do not have too much time to do it, would you mind helping us making a PR to support it? Thanks in advance.

Thanks for your interest. I am working on another program currently and will try to PR after finishing this. The 'cudnn_benchmark' has been set to 'True'. And setting 'with_cp' to 'True' indeed reduces the memory of some models under the same settings.

[ydhongHIT]

Setting 'cuda_benchmark' to 'False' helps, thank you! But setting 'with_cp' to 'True' indeed reduces the memory of some models even though 'cuda_benchmark' is 'True'.

[MengzhangLI]

OK, thats cool! Do you have clear time schedule about your current program?

[ydhongHIT]

OK, thats cool! Do you have clear time schedule about your current program?

Maybe two or three months but not sure.

[MeowZheng]

Juse because the memory on readme is set cudnn_benchmark = False. You could modify it and try again.
By the way, I really want to support DDRNet in MMSegmentation but do not have too much time to do it, would you mind helping us making a PR to support it? Thanks in advance.

Thanks for your interest. I am working on another program currently and will try to PR after finishing this. The 'cudnn_benchmark' has been set to 'True'. And setting 'with_cp' to 'True' indeed reduces the memory of some models under the same settings.

Thanks for suggestion, and we will add with_cp in BEiT, or would you like to create a nice pr to add it. If you have any implement problem, you might refer this pr #1431