error in modulated_deformable_im2col_cuda: no kernel image is available for execution on the device

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

error in modulated_deformable_im2col_cuda: no kernel image is available for execution on the device #160

Open weilunhuang-jhu opened 3 years ago

weilunhuang-jhu commented 3 years ago

Hi, thank you for the impressive work! I failed to test the "test_Vid4_REDS4_with_GT.py" with TITAN RTX but succeeded with GTX1080. When testing with TITAN RTX, I got error as the title. I found that the message came from "EDVR/codes/models/archs/dcn/src/deform_conv_cuda_kernel.cu ". However, I am not familiar with .cu files.

Is it possible to make the dcn module compatible with TITAN RTX? If so, how should I modify it? I am under Ubuntu18.04 and pytorch=1.4.0 (py3.8_cuda10.1.243_cudnn7.6.3_0). Thank you very much in advance!

hjynwa commented 3 years ago

The same error, works on 2080Ti, but fails on TITAN Xp...

xinntao commented 3 years ago

Did you run setup.py successfully?

hjynwa commented 3 years ago

Did you run setup.py successfully?

Successful, but I just found that this error doesn't interrupt the training process... It just prints the error again and again

xinntao commented 3 years ago

Have not encountered this issue before. You may google it to find solutions.

hjynwa commented 3 years ago

Problem solved! Maybe TITAN Xp and 1080Ti are too old for pytorch 1.6. I downgrade cudatoolkit=10.1, pytorch=1.4, torchvision=0.5, remove build/ and rebuild by running python setup.py develop There is no error print anymore.