Hi just a couple of questions about the new style transfer in Create ML. Does it use the same methods/underlying models as Turicreate to do the style transfer? Turicreate lets you use up to 8 styles at once (thus reducing the size of the produced MLmodel) i.e. you don''t need to create a MlModel for each style. Does the new CreateML style transfer have the same ability? Thanks
[tags:wwdc20-10642]
13 results found
Post
Replies
Boosts
Views
Activity
Unfortunately no. CreateML style transfer lets you use only one style image per model.
Hi! I am attempting to create a style transfer model for images. My content image folder has about 80,000 images. However, when I click train, the Create ML app displayed the “Processing” message for about a minute before I get the error: Could not create buffer with format ‘BGRA’ (-6662). Hope someone can help me with what this issue is! Is my training data too large or are their pictures in there that don’t conform to the right image format? Thank you!
Hi, You've probably already resolved this, but it seems very likely that your problem is the size of your training data. I'm far from an expert, but can share my observations as I've investigated storage issues I've hit while creating models. The biggest factors seem to be, the number of training images, the style density (more dense seems to take more space and time), the number of iterations. Good luck!
When and where can I get the latest version of Create ML? I can't wait to use the style transfer feature!
Note that the Style Transfer template requires you to be on macOS 11.0 beta as well.
The latest version of Create ML is delivered in Xcode, so you should be able to download and install the Xcode 12 beta to get it.
Good morning, I really enjoyed the session on Build Image and Video Style Transfer models in Create ML. I would like to know if it is possible to train a similar model, very fast for videos, using turicreate. Thanks!
Great to hear that you enjoyed the session. It's definitely an exciting feature. The model performance enhancements that accelerate the training process and enable the real-time processing you saw for video are unique to Create ML. If you're looking to train programmatically rather than through the app UI, I recommend doing to via Swift. We have a session that dives into this specifically using Create ML framework within a Swift playground: Control training in Create ML with Swift - https://developer.apple.com/videos/play/wwdc2020/10156/
Is the source code for the demo app from the Build Image and Video Style Transfer models in Create ML session available?
Hello, I'm creating models using Create ML with the new Style Transfer. In the WWDC video it can be seen that the size of the model can be less than 1MB. Each model that I try to create is 6.7MB of output size, even changing the parameters and sources to train the models, the size remains high and always identical. I would like to know if it is a beta bug or if there are parameters that can reduce the size of the models to be imported into Xcode. Thanks
Hi Apple Devs, I've been really enjoying the style transfer tool and am excited by the results I'm able to produce. Right now, It seems that the beta version limits the output to 512x512. Are there plans to increase the output size? Stay hungry, stay foolish! Also, stay healthy. Joe
I'm getting an Invalid PNG file error. It isn't clear which image is causing this error and I don't know enough about what is required by the system to triage it.