trainning speed

[AsiamGit]

using multi-scale mode slows down the trainning speed Hi, thanks for sharing your excent work, about the trainning speed,I test several implement and this repo is much better ,but there is one problem:using multi-scale mode slows down the trainning speed when 416416 , 8V100, 64 batchsize, 0.45s/per epoh, cpu occupy 90% when multi-scale , 8*V100, 64 batchsize, 1.8s/per epoh, cpu occupy 20% have you ever met this problems?