Hacker News Reader

Comment by jagged-chisel | Hacker News Reader

For the ML-impaired, how portable is this process? Does it work on other ML platforms? Are the digital assets (specifically this MLMODEL file/bundle) easily transmuted to work with, say, TF?

View on HN

It's not a pre-existing standard, but it is a documented format. It's based on protocol buffers, so I would expect it's fairly easy to write your own reader if necessary.

https://apple.github.io/coremltools/coremlspecification/

That project also has code for converting from quite a few formats (https://coremltools.readme.io/docs) to CoreML.

Having that (3-clause BSD-licensed) code will help in writing the reverse converters.

I’m also ML impaired, but would people really train the model on a Mac? Surely you would want to train in the cloud or on something with a proper discrete GPU. It seems that a better article would be how to take e.g. yolo and convert it.

people do run Macs in datacenters, so if you wanted to have it run in a dc I’m sure you could. Secondly all pro level Macs have discrete graphics, so I’m not sure what you mean by “proper discrete gpu”

Discrete nvidia GPU with CUDA. While there is Tensorflow compiled for radeon GPUs, majority of libraries are based on CUDA

... The models created by this are designed to be used in apps for on-device ML, on Macs, iPhones, iPads, etc.

Tensorflow was recently released for Metal

Has there been any major updates to using TF for AMD GPUs recently? Or is it still too much of a headache to do it?

Not sure what you mean by "all pro level macs". But the macbook pro M1 has no discrete GPU. It does however have a "neural processor" to accelerate ML.

Ah right I always forget about the 13” Ugly step sister - it was igpu only for the intel models too.

However the other pro lines (MBP15/16 iMac Pro; Mac Pro) are all equipped with dGPUs

Presumably this can be trained on the neural engines on iPhone/iPad chips and the new M1s.

Also my Macbook Pro has an 8GB GPU.

Typically the model is pre-trained and you're just training the last layer, or adding a new layer, which is much much less computationally expensive.

The article is not explicit about it but I think they’re using transfer learning. They take a general model trained, for instance, on ImageNet and use it to train a more specialised model to classify whatever you want. The training of the specialised model is way faster than the training of the general model.

Apple's main goal with Create ML is to make something app developers can use to add machine learning powered features to their apps. So it makes sense that they'd want something that you can train with the hardware you already have as an app developer.

It's actually pretty performant. I trained an object detector (from scratch) on my 2016 MBP in about 6 hours[1].

As of macOS 11 also support transfer learning for object detection which starts from pre-trained weights and should converge more quickly if your dataset is similar to COCO[2] but I haven't tested it out yet.

[1] https://blog.roboflow.com/createml/ [2] https://blog.roboflow.com/coco-dataset/

I did a quick Google search (so maybe I’m missing something) but there is no official support for converting Core ML — which is the foundation for Create ML — models to Tensorflow or PyTorch. You need to look for third-party libraries. Although there exist libraries to convert Tensorflow models to Core ML.

You can convert CoreML models to ONNX which is widely supported cross-platform for doing inference. https://github.com/onnx/onnxmltools

sorry my lack of understanding: So you think an image classifier trained on Create ML can be exported to another framework of not?

It’s not entirely clear to me exactly what was meant by the grandparent post by converting to PyTorch or Tensorflow.

There are two discrete concepts: training and inference. You can think of training a bit like the source code and inference a bit like running the compiled binary. The different frameworks have their own serialization formats for their model weights.

If the goal is to do inference using CoreML weights trained in CreateML on non-Apple platforms (eg on a server or android device), converting them to ONNX is a way to do that.

You probably won’t be able to pick up training on another framework though.

Reply on news.ycombinator.com