search

jcjohnson / neural-style

Torch implementation of neural style algorithm
MIT License
18.31k stars 2.7k forks source link

high resolution #196

Open JeffCrusey opened 8 years ago

JeffCrusey commented 8 years ago

my setup is crashing anything beyond about 600 pixels. id like to get as high of a resolution as possible. any tips for make something closer to 2k?

xenon2 commented 8 years ago

buy more ram

charlesfg commented 8 years ago

A LOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOT more ram!

janisozaur commented 8 years ago

https://geizhals.eu/hp-lrdimm-64gb-m4z04aa-a1338532.html

xenon2 commented 8 years ago

I tried to render 1920x1080 image and it takes about 50GB of RAM (using cpu backend).

jcjohnson commented 8 years ago

with cuDNN and Adam you should be able to fit 1920x1080 in 12GB of GPU memory.

On Fri, Apr 1, 2016 at 2:41 PM, Jakub notifications@github.com wrote:

I tried to render 1920x1080 image and it takes about 50GB of RAM (using cpu backend).

— You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub https://github.com/jcjohnson/neural-style/issues/196#issuecomment-204575090

charlesfg commented 8 years ago

@jcjohnson do you think that is feasible to implement a trade off between time vs RAM. Using something like a memory map using SSD ?

jcjohnson commented 8 years ago

That is not easy to implement; it would also seriously degrade performance. I've written more details in other issues.

On Fri, Apr 1, 2016 at 3:11 PM, Charles Ferreira Gonçalves < notifications@github.com> wrote:

@jcjohnson https://github.com/jcjohnson do you think that is feasible to implement a trade off between time vs RAM. Using something like a memory map using SSD ?

— You are receiving this because you were mentioned. Reply to this email directly or view it on GitHub https://github.com/jcjohnson/neural-style/issues/196#issuecomment-204587771

htoyryla commented 8 years ago

Justin notifications@github.com kirjoitti 2.4.2016 kello 0.43:

with cuDNN and Adam you should be able to fit 1920x1080 in 12GB of GPU memory.

Likewise, if one prefers L-BFGS, using nin-imagenet-conv instead of VGG19 produces 1920x1440 in 10GB (measured using CPU here).

Hannu

hexus commented 8 years ago

@htoyryla How long does that take for 1000 iterations? :grin:

htoyryla commented 8 years ago

Chris Andrew notifications@github.com kirjoitti 10.4.2016 kello 4.57:

@htoyryla How long does that take for 1000 iterations?

1 hour 40 minutes, on an i7 3.6 GHz 8 threads.

Hannu

tommybananas commented 8 years ago

Does increasing RAM help in gpu mode too or only cpu?

jcjohnson commented 8 years ago

More system memory will not help in GPU mode.

0wnrepo commented 8 years ago

It would be good to centralize these user reported benchmarks

monik3r commented 8 years ago

@htoyryla how do the results of nin-imagenet-conv look vs VGG19?

htoyryla commented 8 years ago

I find them useful enough. Difficult to compare when a) I cannot get images this large with VGG19 b) results with identical parameters on different models are not directly comparable.

Have a look here http://liipetti.net/erratic/tests/1920px-tests/

Since I made those tests, @ProGamerGov has reported that a lower tv_weight can produce sharper images with NIN (without adding too much noise that is). And I would not recommend using the highest conv layers with NIN, at least not without understanding what they contribute, see https://github.com/jcjohnson/neural-style/issues/226 .