search

lmb-freiburg / flownet2

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
https://lmb.informatik.uni-freiburg.de/Publications/2017/IMKDB17/
Other
1k stars 318 forks source link

FlowNet2 KITTI memory requirements #180

Closed nikolausmayer closed 5 years ago

nikolausmayer commented 5 years ago

I have tried your network, and it gives me useful informations. My gpu is gtx980 which only have 4gb. I think it is not enough for Flownet2 adn kitti. The weight is over 600mb it seems. I can only use light weight which is just about 100 mb.

Do I need a gpu with big memory?

Originally posted by @WANGLXUE in https://github.com/lmb-freiburg/flownet2/issues/164#issuecomment-448589049

nikolausmayer commented 5 years ago

My GTX 970 with 4GB runs FlowNet2-KITTI just fine on a KITTI sample with 1226x370 pixels. My maximum memory consumption is less than 3400MB. if you have multiple GPUs, make sure the network is using the correct one.

nikolausmayer commented 5 years ago

@nikolausmayer Thanks. I used your flying chairs and it works. I also understand it is because my input image size is too big. So how do you think if I want to input the original size 1920x1080. How big memory do I need for my gpu? Perhaps my professor can buy me a good one. Thanks again.

Originally posted by @WANGLXUE in #164 (comment)

nikolausmayer commented 5 years ago

@WANGLXUE I just tested on a GTX 1080, and I can process 1920x730 images with the full FlowNet2 architecture. I don't have ready access to FlowNet2 on a larger GPU, so I can't tell whether 11GB VRAM are enough for 1920x1080. If a newer FlowNet is ok too, you could try https://github.com/lmb-freiburg/netdef-docker which runs in TensorFlow and does not need as much memory.

nikolausmayer commented 5 years ago

(closed due to inactivity)