search

NVIDIA / FastPhotoStyle

Style transfer, deep learning, feature transform
Other
11.15k stars 1.2k forks source link

Much Slower Than the Reported Time #19

Closed onlywuyiwuyi closed 6 years ago

onlywuyiwuyi commented 6 years ago

Hi, I tested your code by running demo.sh with a K40m GPU, but my CUDA version is 8.0 (not 9.1). The total time is about 145s, more than 10 times slower than the reported time in the paper (11.39s for 1K image size). Besides a better GPU (Titan XP), I wonder whether the new CUDA is the key for the high performance. Thanks.

mingyuliutw commented 6 years ago

Could your report the times for the 3 processing steps individually?

onlywuyiwuyi commented 6 years ago

Elapsed time in stylization: 2.512861 Elapsed time in propagation: 141.474549 Elapsed time in post processing: 1.093370

Yijunmaverick commented 6 years ago

@onlywuyiwuyi Did you install scikit-umfpack as suggested in USGAE.md?

pip install scikit-umfpack

onlywuyiwuyi commented 6 years ago

Thanks. By installing scikit-umfpack, the second stage is accelerated to 25s. Is this time reasonable?

Yijunmaverick commented 6 years ago

No idea about your image size and GPU specs, but I think it looks okay.