Veith Roethlingshoefer
Feb 14, 2017, 3:37:09 PM2/14/17
to Caffe Users
Hello,
I have successfully installed and run inference on PSPNet (https://github.com/hszhao/PSPNet).
This network uses matcaffe. I use CUDA 7.0 CuDNN 4.0 on a GTX 1080.
If you compile matcaffe without Cudnn, i.e. USE_CUDNN := 0, everything works as expected, just a tad slow and more memory intensive.
I get some nicely segmented images.
BUT, if I compile with USE_CUDNN:=1, the whole process breaks. The result of every image I segment is just a uniform image now, even when using the same images as without Cudnn.
For example using cityscapes, every pixel of the image is set to 'purple', which corresponds to 'road'.
Has anyone ever experienced this or something similar and knows a fix? I'll also gladly take your best guess, since I really have no idea how this might happen.
Best regards and thank you!
I have successfully installed and run inference on PSPNet (https://github.com/hszhao/PSPNet).
This network uses matcaffe. I use CUDA 7.0 CuDNN 4.0 on a GTX 1080.
If you compile matcaffe without Cudnn, i.e. USE_CUDNN := 0, everything works as expected, just a tad slow and more memory intensive.
I get some nicely segmented images.
BUT, if I compile with USE_CUDNN:=1, the whole process breaks. The result of every image I segment is just a uniform image now, even when using the same images as without Cudnn.
For example using cityscapes, every pixel of the image is set to 'purple', which corresponds to 'road'.
Has anyone ever experienced this or something similar and knows a fix? I'll also gladly take your best guess, since I really have no idea how this might happen.
Best regards and thank you!
Reply all
Reply to author
Forward
0 new messages