Prisma generate neural style offline now

[markzhong88]

So they can do it within a iPhone processor and it takes around 5 seconds without using GPU. The result is also very good, any thoughts?

[jcjohnson]

Prisma uses a method similar to that described in my recent paper:

http://arxiv.org/abs/1603.08155

[3DTOPO]

@jcjohnson Do you still have any plans to release your implementation? :)

[jcjohnson]

@3DTOPO Yes, I've just been extremely busy with other projects this summer.

[markzhong88]

@jcjohnson Yeah, I believe they use the feed forward method, but they leverage the technology so good, generating the best image so far, and run very fast locally on iPhone without using GPU.

[3DTOPO]

@jcjohnson I understand! Not to press, but might you have any time frame in mind?

@markz-nyc they are almost certainly utilizing the GPU on the iPhone.

[jcjohnson]

@3DTOPO I'm presenting the paper at a conference in October, and I'd really like to have code released by then. In the meantime you can check out Texture Nets (https://github.com/DmitryUlyanov/texture_nets); their feedforward results are as good or better than mine.

[3DTOPO]

@jcjohnson thanks for the info and tip! I had seen texture_nets, but I have spent all my time experimenting with https://github.com/yusuketomoto/chainer-fast-neuralstyle

I'll check it out, thanks!

Do you think this project has the frameworks needed for a port of your implementation?

https://github.com/collinhundley/Swift-AI

[jcjohnson]

@3DTOPO Looks like Swift-AI doesn't have convolutional layers; those are essential.

Apple's BNNS framework is closer (https://developer.apple.com/reference/accelerate/1912851-bnns) but doesn't seem to have convolution transpose (also called deconvolution, full convolution, or upconvolution).

[carloser200]

great Project

[nguyenvo09]

I like this project.

On Mon, Aug 29, 2016 at 8:22 PM, Ben Janik notifications@github.com wrote:

great Project

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/jcjohnson/neural-style/issues/313#issuecomment-243314554, or mute the thread https://github.com/notifications/unsubscribe-auth/AA9yzOz65UOAKnUuqaWINcBPK5cipJL0ks5qk5PagaJpZM4JutzL .

Nguyen Vo

[julien-c]

@jcjohnson For large convolutional nets, Metal performance shaders are the way to go, not BNNS (which run on the CPU after all). See this for instance : https://developer.apple.com/library/content/samplecode/MetalImageRecognition/Introduction/Intro.html

[3DTOPO]

BNNS is hardware accelerated and built on top of Metal.

[julien-c]

@3DTOPO Hmm... Source please?

[austingg]

@3DTOPO @julien-c BNNS for cpu and Metal use gpu

[julien-c]

@3DTOPO You should probably delete your comment

[3DTOPO]

I thought it was based on the WWDC video I watched and @jcjohnson recommendation.

But apparently it is part of the Accelerate Framework that is CPU optimized not GPU optimized, so I was wrong.

Link about BNNS and Metal Performance Shaders (MPS) CNN: https://www.bignerdranch.com/blog/neural-networks-in-ios-10-and-macos/