CoreML Model Zoo

This is a collection of Machine Learning models converted into the CoreML framework. These models can run on Apple devices under one of the following operating systems: iOS, macOS, tvOS, watchOS.

All the models in the repo:

• have pre-trained weights
• are intended mainly for inference
• are optimized for mobile devices

Below there are few sections:

• Format and Available Models describes all the models available and details about them
• Usage gives examples/links/information on how to use the models

Format and Available models

All models should be run on one of the following systems: iOS (11.0+), macOS (10.13+), tvOS (11.0+), watchOS (4.0+). While the specified versions are the minimum required versions to be able to run CoreML models at all, it is highly recommended to use the latest available OS version. Otherwise, some of the models might run much slower or might even fail to run if they require some of the new features.

Currently only Computer Vision models are presented. In future, Natural Language Processing and Speech Processing models might be ported as well and any contribution is welcomed.

Each model has:

• Model file in the format of *.mlmodel
• Model size in MB
• Input size in pixels
• Quality metrics both for source and converted versions
• Performance metrics such as latency/RPS for different mobile devices (iPhone 7, iPhone 11 Pro) and accelerators (CPU, GPU, ANE).

For more info on how the performance is measured please look in the coreml-performance repo.

Computer Vision models

All the Computer Vision models are placed under vision folder and are split by problem they are solving.

Inputs to all the models are unified:

• 3-channel RGB
• [0-255] pixel range

All the necessary preprocessing steps are already embedded into the model so that the user should not worry about that.

Classification

Model	Reference	Year
AlexNet	"One weird trick for parallelizing convolutional neural networks"	2012
VGG	"Very Deep Convolutional Networks for Large-Scale Image Recognition"	2014
ResNetV1	"Deep Residual Learning for Image Recognition"	2015
DenseNet	"Densely Connected Convolutional Networks"	2017
MobileNetV2	"MobileNetV2: Inverted Residuals and Linear Bottlenecks"	2018
ResNeSt	"ResNeSt: Split-Attention Networks"	2020

Usage

The easiest way to run the Coputer Vision models on device is to use Apple Vision library and auto-generated code for the *.mlmodel file. Below is an example for resnet18_v1_torchvision model (see this tutorial for more). Note that there is no error processing in this example which should be a must-have in real application.

import Vision
import CoreImage

let model = try VNCoreMLModel(for: resnet18_v1_torchvision().model)

let request = VNCoreMLRequest(model: model)
request.imageCropAndScaleOption = .centerCrop

let handler = VNImageRequestHandler(ciImage: ciImage)
try! handler.perform([request])

let results = emotionsRequest.results! as! [VNClassificationObservation]

There are other ways to use raw CoreML model without any frameworks. One might find example of this in the coreml-performance repo.