search

bermanmaxim / LovaszSoftmax

Code for the Lovász-Softmax loss (CVPR 2018)
http://bmax.im/LovaszSoftmax
MIT License
1.38k stars 269 forks source link

About the slow speed on tensorflow #9

Closed MarkYangjiayi closed 6 years ago

MarkYangjiayi commented 6 years ago

Hello, I loved your work on Lovasz softmax very much and implemented it on a modified version of Deeplabv3+ in Tensorflow. However, I experienced significant speed drop, the time used per step increased from 0.4s ( using cross entropy) to now almost 3.8s. Is this normal or did I do something wrong? Thankyou!

MarkYangjiayi commented 6 years ago

@bermanmaxim

bermanmaxim commented 6 years ago

Hi @MarkYangjiayi, thanks for the interest.

This is understood to be because of Tensorflow's implementation of cumsum. See this issue for reference (I reopened it for visibility): [issues/6].

A solution would be to interface with nvidia cub's implementation of cumsum but I haven't had the time to do it, and we have been working with pytorch where cumsum is fast.

bermanmaxim commented 6 years ago

Closing as duplicate.