search

sniklaus / pytorch-pwc

a reimplementation of PWC-Net in PyTorch that matches the official Caffe version
GNU General Public License v3.0
608 stars 122 forks source link

Performance issue #18

Closed donnydonny123 closed 5 years ago

donnydonny123 commented 5 years ago

Hi @sniklaus ,

In the original PWC-net paper, PWC-net's processing time is about 28ms. image

But when I measured the time passed in run.py's estimate function and moduleNetwork, they cost about 330ms for estimate function and 173ms for moduleNetwork, which is too large to ignore.

I want to know that is there a reason that causes this issue. Thanks!

sniklaus commented 5 years ago

What is the input resolution for the table that you quoted? And which graphics card are you using?

donnydonny123 commented 5 years ago

Hi, In the original paper's conclusion, 1024*436 images is used for benchmark. The command I am using is the same as provided in https://github.com/sniklaus/pytorch-pwc#usage , Which is python run.py --model default --first ./images/first.png --second ./images/second.png --out ./out.flo GPU is GTX 1080 ti. Thank you for your reply.

sniklaus commented 5 years ago

Interesting find, thank you for bringing this up! Calling moduleNetwork in a for loop using the provided images clocks in at around 130 ms per estimate on my 1080 Ti, still much too slow. Note to everyone reading this and who might not be aware of the asynchronous computation nature of GPU computing, making sure to use torch.cuda.synchronize() to get the correct timing. Have you tried benchmarking the official release? If so, what is the timing you get? Thanks!

sniklaus commented 5 years ago

Closing the issue due to inactivity, I would still love to hear more about this though!