New Model: Fast-Style-Transfer

[backnotprop]

License

My Code is MIT, though the model requires an agreement for commercial use

Summary

I implemented Fast Style Transfer in CoreML. This is similar to Fast Neural Style, but is TensorFlow.
My article explains the process.

Model URL

https://github.com/lengstrom/fast-style-transfer

Demo URL

https://medium.com/@rambossa/diy-prisma-fast-style-transfer-app-with-coreml-and-tensorflow-817c3b90dacd
https://github.com/mdramos/fast-style-transfer-coreml

Samples

Here are some CoreML models I created in the process: https://drive.google.com/drive/folders/1CBSanBHbXC5-bJNTTk3-r1WSq56z0eKG?usp=sharing
-- just include these in the ios app before running

Checklist

• Only one item is in this issue
• The model info contains all the required fields
• The demo project is compilable (AFTER INCLUDE THE MLMODELS)
• [?] Has proper reference
• If this model takes image as input, an image type is selected instead of multiarray
  (See https://developer.apple.com/wwdc17/710 if you don't know how to do this)

[likedan]

Yo. The demo looks very similar to https://github.com/prisma-ai/torch2coreml. I'm not sure how much it's interrelated but maybe you should cite it?
There are new models so I'll take this!

[backnotprop]

Hey it’s sited in my blog post, and I state that I use Prisma's approach as a basis— but I’ll add the same to the repo. Very similar for sure. But mine uses a TensorFlow model and converter (a lot of people have been trying to convert this to CoreML) by Logan Engstrom, whereas Prisma demonstrates converting the Torch model by Justin Johnson. Thanks!!

[likedan]

Now I see the difference. Great work!

[devajith]

will you able to convert this style transfer model :
https://github.com/vanhuyz/CycleGAN-TensorFlow/releases/download/v0.1-alpha/apple2orange.pb

[backnotprop]

@devajith it looks like that graph has cycles, so probably not (at least not easily)