CreateML's image classification is designed to work on natural images, so it's not surprising that it may be sub-optimal on images from user drawings. There is no other dedicated custom model training solution for drawing classification. You can always try some open source drawing classifier (PyTorch implementation), fine tuning onto your own data and convert it to Core AI (or CoreML) to run on device. If you can try CreateML framework instead of the CreateML app you can plug in an open source image feature extractor that is more suited for drawings (than natural images). That path you do not have to worry about setting up training in Python.

I tried simplifying the dataset to just a few classes and short examples in case there was a problem with the data. Can you share this sample training data you use to reproduce this issue in the CreateML app?

great, thx for confirming. We will look into your setup

While we investigate whether the new Linux & DGX environment causes this issue of missing dependency, a workaround you can try is to perform the exporting step (saving .fmadapter) on a Mac with the python environment set up as documented. You need to move over the checkpoints (adapter and draft model) from training and call the exporting step there.

What is the output of pip freeze | grep coremltools, in the Python environment you running this notebook?

Have you tried right click and open a new window/tab (or two-finger click and select open a new window)? This is a known issue and this method seems to work around it.

Can you file a feedback using feedback assistant and attach a problematic USDZ file?

Thank you for trying it out. It is recommended to have at least 60GB disk space to support object tracking model training. The used disk space will be released upon training completion.

Can you file a feedback on https://feedbackassistant.apple.com? Pls attach a sample original image, the cropped image and trained image classifier model (.mlmodel file). There could be some preprocessing in play here, so a code snippet from your application helps too.

Have you had any update on this? We'd like to hear from you. If possible, pls do a screen recording and file a bug via Feedback Assistant, so we can investigate. btw: this sample code was tagged as Vision instead of CreateML, which is why we could not respond in time

Did you get this model trained from CreateML app? If yes, have you tried if you can reproduce the same issue (colorKeys being the same despite providing different inputs) in CreateML app's Preview pane?

It's supported in Playground. The referred thread is caused by a known SDK issue (without knowing your specific OS/Xcode configs, it's hard to confirm). I'd suggest upgrade to latest macOS and Xcode to try again.