srikris
commented
6 years ago CreateML uses a different base model by default. You can use the same base model as Create ML using the following:
model = turicreate.image_classifier.create(data,
target='is_expensive', model = 'VisionFeaturePrint_Screen')More details here: https://apple.github.io/turicreate/docs/api/generated/turicreate.image_classifier.create.html#turicreate.image_classifier.create
Some performance numbers are available here: https://apple.github.io/turicreate/docs/userguide/image_classifier/how-it-works.html
A while ago I created an image classification model that ended up being about 100 MB in size, trained with just a few thousand images. At WWDC it was announced you can now quantize your Core ML models to significantly reduce the file size. It seems Create ML does this automatically. Anyways, I am wondering if Turi Create will automatically quantize weights, or is there a way to specify it should be quantized using 8, 4, or 1-bit? I understand there’s some trade offs with accuracy to consider, but my naive mind seems to think Turi Create could automatically quantize it based on the data to achieve the best results. Thanks!