cudaSuccess (4 vs. 0) unspecified launch failure about deep-image-analogy HOT 5 OPEN

deep_image_analogy.exe ../models/ ../demo/content.png ../demo/style.png ../demo/output/ 0 0.2 2 0

changing ratio from 0.5 to 0.2 gets the program to run on GTX 1070 with 8GB.

from deep-image-analogy.

Can you fire up some gpu analysis tool and check memory usage? Mine never goes above 1.8Gb and I still get this error

Did you compile with or without cuDNN? Did you use CUDA 7.5 or 8.0?

Have you tried making only one-way style transfers like suggested in another issue? If so, how did you make that happen?

from deep-image-analogy.

I also can't run the application due to memory problems (presumably) on my GeForce GTX 1050, which has 4 GB of GPU memory. Lowering the ratio works, since it will scale down to images before passing them through the network, but the results are pretty bad.

On a Tesla K40c with 12 GB, everything works fine. However, the memory consumption there seems to never exceed 1 GB, so I wonder why I can't run it on my GPU.

from deep-image-analogy.

@Callidior Yep I "solved" my problem the same way, a Tesla K80 can handle maximum resolution at 1.0 ratio

I've heard about memory spikes, maybe it does run out of memory but it does so quickly the sensor app doesn't pick it up?

Well I'm glad actually, at least this confirms that this is not a temp throttling issue

And yeah the results are "pretty bad" alright

from deep-image-analogy.

See if disable the time limit (TDR) would help #25

from deep-image-analogy.