search

jwyang / faster-rcnn.pytorch

A faster pytorch implementation of faster r-cnn
MIT License
7.65k stars 2.33k forks source link

Data loader speed #574

Open xialeiliu opened 5 years ago

xialeiliu commented 5 years ago

I found that the gpus are waiting for cpus most time, and cpu usage is very low even with num_workers = 8, that's might slow down the training process. Do you have any idea how to improve the speed of data loader?

xialeiliu commented 5 years ago

Is there a parallel data loader available?

kentaroy47 commented 5 years ago

you might want to try increasing the batchsize? this will relax the data loading issue quite much.

xialeiliu commented 5 years ago

I already used the largest batch size I can in a 8-gpu machine. that's might not be the solution.