search

xinntao / EDVR

Winning Solution in NTIRE19 Challenges on Video Restoration and Enhancement (CVPR19 Workshops) - Video Restoration with Enhanced Deformable Convolutional Networks. EDVR has been merged into BasicSR and this repo is a mirror of BasicSR.
https://github.com/xinntao/BasicSR
1.48k stars 320 forks source link

Out of memory #142

Open platten opened 4 years ago

platten commented 4 years ago

Hello,

I have an RTX 2080Ti with 11GB of memory, yet for some reason, I keep running out of memory when trying to do an upscale using EDVR_util.single_forward on any image larger than 250x250 px. The pre-trained model I am using is EDVR_Vimeo90K_SR_L.pth and I am configuring the model in the following way:

EDVR_model = EDVR_arch.EDVR(128, 7, 8, 5, 40, predeblur=False, HR_in=False)

Any ideas? Thanks!

oom

ryul99 commented 4 years ago

I think that an image that is bigger than 250x250 px is too big to load to GPU memory with EDVR model. In my case, I made a script that split image and runs EDVR after split image. This is my script. I hope this helps you. https://gist.github.com/ryul99/a192310cf2cd3ce94f83b928d44a141f

ojasviyadav commented 4 years ago

These folks used 8 Titan GPUs to train their model. I'm not sure if a single 2080Ti would be able to rival that. As @ryul99 pointed out, you would have to use some technique to reduce the data you're loading into memory.