Deep Dreamer GPU Crashes
Just a note - I've sent in the crash logs - but just in case it helps - I've never had any luck with the GPU support with this app - ever since it was introduced, its crashed every single time I try to use it. NVIDIA GeFORCE GTX 680 with 2MB RAM. Have tried with the latest Nvidia drivers and built in Mac OS X Nvidia drivers on both latest versions of Mac OS 10.10 and 10.11 - no dice, always results in a crash. Really hoping this can be addressed now that a new team is working on this.
Thanks for the feedback. We are looking into these crashes. Currently the Nvidia seems to be the problem here as other cards are fine. We are trying to work out where and what is actually causing the problem which is taking some time as the GPU support comes from Caffe which is part of the Google source code. We are hoping to be able to at least capture these issues rather than things just go pop to be able to better handle them.
Do you get the same problems when running on the CPU. It is a CPU hog but CPU - Fast is working for others who are also having the GPU issues.
Hi - thought I'd take a look in here - doesn't look like much changed in the past 4-5 months - no updates or releases....
I can run this in regular CPU mode, but its very slow and CPU Fast is a CPU hog for sure, as you noted. GPU support literally will accelerate this by many orders of magnitude.
Have there been any advances, updates, etc. or is this thing just dead in the water when it comes to more robust GPU support?
As an FYI: Here are some notes from about 10 months ago when it comes to compiling googles code and caffe to ensure support for CUDA with NVIDIA cards is enabled:
Also for reference, to enable NVIDIA based CUDA support with Pycaffe - which is what I believe Deep Dreamer to be utilizing - there are some explicit instructions that may help (found on the Caffe site for getting things up and running under a Docker container) - specifically see the sections on Pycaffe and cuDNN Caffe with cudnn support - that may help solve the issue we are seeing: http://caffe.berkeleyvision.org/installation.html
There is also a note about CUDA compute capability at the very bottom in hardware where they have basically tested only for cards with a 3.0 or greater score. My GTX 680 scores a 3 easily.
In any case, I hope this app isn't dead - but it sure looks like from the outside it hasnt been getting much love.... Hope I am wrong.
Unfortunately we've not got much out to anyone over the last few months for Deep Dreamer. It's still being worked on, primarily around the stabilising of the GPU mode and we have been experimenting with being able to teach it new images. Unfortunately this stuff just eats cycles and with only two of us that can cause things to slow down significantly.
We are planning on getting an updated release out to address the GPU issues that some people have been having. I appreciate the heads up to the info you provided. We've seen these and been using them while working on stability.
Thanks again for the feedback