Thread replies: 5
Thread images: 3
Thread images: 3
File: resultBA.png (338KB, 448x448px) Image search:
[Google]
338KB, 448x448px
I'm trying to get this thing to work
https://github.com/msracver/Deep-Image-Analogy
Compiled it within 15 minutes no problem, but it always crashes on larger images or if I set the ratio higher than 0.2
On some images it won't even do ratio 0.1, and at that point it looks like pic related
The results look okay-ish with one particular style image with which I can go up to 0.3 ratio
The error always looks like this:some CNN processing happening...
im2col.cu:61] Check failed: error == cudaSuccess (4 vs. 0) unspecified launch failure
I have a gt960m with 4gb gddr5, when I first looked the error up it said that this is due to not enough VRAM
But I fired up GPU-Z and checked that at most, 1.7 gigs of VRAM is being used, and if I fire up a VRAM intensive application caffe crashes at the same place every single time
AND I found out the authors of the paper used an ages old tesla that is actually less powerful than this gt960m (even though in theory it has more cuda cores)
What can I do to make it work? Ideas? Debug builds crash earlier for some reason. I used CUDA 8.0 and compiled without cuDNN
>>
File: resultAB.png (390KB, 448x448px) Image search:
[Google]
390KB, 448x448px
2nd pic
in one of the github issues the guy that maintains the git said that you can get perf improvements from only generating picture A, but looking through the code as someone who hasn't gone past "hello world" in cuda I've no idea how to make that happen
but even then I don't think I have perf issues, maybe the notebook voltage throttling is an issue? I can dig up an old notebook cooler which does bring down the temps considerably (at the cost of a fuckton of dust)
and also the gpu never comes close to thermal throttling, mostly stays around 60-70 degrees even though it's at 96-99% load the whole time, the most I've ever seen it go is 80 degrees and that's with furmark running with prime95 in the background
>>
Another weird thing is that you would think given that I have a maxwell GPU I should enable the "optimise compute performance" toggle in the nvidia control panel
But nope, apparently some miners think this actually makes performance worse, and it makes caffe complain about "all cuda devices" being "busy" and googling around there's no way around it
>>
I also tried the windows graphics driver timeout registry property
I knew it probably wasn't the root cause but I tried anyway
first I tried 8 seconds which is quite a lot
now it's at 60 seconds which is beyond reasonable and even though I suspect it helped a little, I don't think it was significant enough
>>
I even tried doing a debug session in vs2013 and when it crashes it either just flat out refuses to show the line it crashed at (instead some warning screen is displayed) or it displays some totally irrelevant random line (even a comment once)
Is there maybe some way I have to configure vs2013 to properly support debugging something that uses the cuda sdk? Should I try reinstalling it? I just installed it last weekend
I also understand it could be caused by bad code (on the author's part?) but please be merciful I've no idea what I'm doing
Just... if anyone knows what could cause this other than not enough memory, I'll be extremely gratefulCheck failed: error == cudaSuccess (4 vs. 0) unspecified launch failure
Thread posts: 5
Thread images: 3
Thread images: 3