search

jcjohnson / neural-style

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

Caching style image calculation #280

Open kirilledelman opened 8 years ago

kirilledelman commented 8 years ago

Just curious - is there a way this algorithm can cache, or store intermediate result to a file, namely the style image, and reuse this cached data when invoked again with the same style image but different content?

htoyryla commented 8 years ago

The style image is run through the net once in the beginning to get a "measure" of the style to which the intermediate results will then be compared. This initial measuring of the style is quite fast and it would not make any noticeable difference, I think, to store it for re-use.

The situation is completely different in the recent developments which, instead of iterating to get the image, generate the image in one go using a feed-forward network which has been pretrained using a specific style. The downside using this approach is that it takes a long time to train a model. See https://arxiv.org/abs/1603.08155 and https://github.com/yusuketomoto/chainer-fast-neuralstyle.

kirilledelman commented 8 years ago

@htoyryla Yeah, I tried chainer-fast-neuralstyle, and it's prohibitively slow - takes days to train for one style, results are very grainy, and you don't know what you're going to get. Thanks for your response though.