IDEA-Research/DWPose

Training time so long

Opened this issue · 1 comments

Good job! How long does it take if I would like train on s1_dis on whole-body and UBody on 1 A100 device, if I take the default config(270 epoch) the estimated time will be 7 days? I don't think that's normal, what kind of data ratio should I maintain if I want to add a subset like Ubody to train s1_dis?

Good job! How long does it take if I would like train on s1_dis on whole-body and UBody on 1 A100 device, if I take the default config(270 epoch) the estimated time will be 7 days? I don't think that's normal, what kind of data ratio should I maintain if I want to add a subset like Ubody to train s1_dis?

Is there a solution to this? The paper says that 8 GPUs were used. Need 8*A100 GPUs?