Keras Multi_GPU_Model：像糖蜜一样缓慢

Question

I am investigating Keras for multi gpu modeling. So before I invest time on it, I tried out a simple skipgram model on a 4 gpu instance from lambdalabs.

The one gpu performance, is slightly worse than on Kaggle's kernel (the number of cores are smaller on lambda labs gpus).

But the multi gpu performance supremely sucks. I've read https://github.com/keras-team/keras/issues/9204 but its unclear if there is a lot of weight merging for a simple model like this.

Playing around with batch sizes oscillates between out of memory and converging roughly by the time we have world peace. And the question is : Do we know more about multi gpu issues with keras since last year, and how might I go about trouble shooting it?

Also - I am using fit_generator. Are the arguments "use_multiprocessing" and "workers" for the gpu or for the cpu?(I think the latter).

Answer 1

当use_multiprocessing=True和workers>0 ，keras 将使用 multiprocess 在 cpu 中生成数据并将它们提供给 gpu。