Navigating the Apple ML Landscape

Sep 12, 20181 like333 views

The document discusses the practical applications of machine learning (ML) in various tasks, emphasizing on-device inference benefits such as user privacy and low latency. It outlines a specific ML approach for classifying board game pieces using rectangle detection and image classification, detailing the steps for implementing this in a coding environment. Additionally, it touches on data collection, model training, and setting up a Python virtual environment for advanced ML tasks such as style transfer and activity classification.

Engineering

Navigating the
!
ML Landscape
Meghan Kane, @meghafon
developer @novoda, Berlin
Swift & Fika, 9/2018

ML tasks
鈫� detection: rectangles, face, barcode, text
鈫� classi铿乧ation: image, text, activity, column data
鈫� style transfer
鈫� image similarity
鈫� object detection
鈫� natural language processin
鈫� regression

Advantages of on device inference
1. user privacy
2. low latency user experience

ML approach
Label: "Border"
1. Detect game box (rectangle
detection)
2. Classify Agricola piece within
rectangle (image classi铿乧ation)

Capabilities
... built right into the Vision
framework, no model training
needed
鈫� detection: rectangles, face,
barcode, text
鈫� object tracking
鈫� image alignment

How?
1. create handler
2. create task speci铿乧 request
3. send request to handler
4. handle results

// 1. Create handler
let handler = VNImageRequestHandler(cgImage: image,
orientation: orientation,
options: [:])

// 2. Create request
let request = VNDetectRectanglesRequest(completionHandler: self.handleDetectedRectangles)

// 3. Send request to handler
do {
try handler.perform([request])
} catch let error as NSError {
// handle error
return
}

// 4. Handle results
func handleDetectedRectangles(request: VNRequest?, error: Error?) {
if let results = request?.results as? [VNRectangleObservation] {
// Do something with results [*bounding box coordinates*]
}
}

Why restrict input image to the box?
鈫� easier to train an accurate model
鈫� faster to collect image data

Capabilities
In Xcode playground, train custom
model for:
鈫� image classi铿乧ation
鈫� text classi铿乧ation
鈫� classi铿乧ation & regression of
column data

Collect Data
鈫� Collect images representative of real world use
cases
鈫� Vary angle & lighting
鈫� >10 images per label, but ideally more
鈫� Equal # images for each label
鈫� Recommended: >299x299 pixels

Collecting Data Quickly
// Extract .jpg frames from .mov @ 5 frames/sec
ffmpeg -i stone.mov -r 5 data/stone/stone_%04d.jpg

Prepare Data
鈫� Split data: 80% train / 20% test

Capabilities
鈫� perform predictions using
model
鈫� quantized weights (32 bit -> 16,
8, 4... bit)
鈫� perform batch predictions
鈫� create custom model layer

Vision + Core ML
1. create Vision Core ML model
2. create handler
3. create task speci铿乧 request
4. send request to handler
5. handle results

// 1. Create Vision Core ML model
let model = AgricolaPieceClassifier()
guard let visionCoreMLModel = try? VNCoreMLModel(for: model.model) else { return }

// 2. Create handler
let handler = VNImageRequestHandler(cgImage: cgImage, orientation: cgImageOrientation)

// 3. Create request
let request = VNCoreMLRequest(model: visionCoreMLModel,
completionHandler: self.handleClassificationResults)

// 4. Send request to handler
do {
try handler.perform([request])
} catch let error as NSError {
// handle error
return
}

// 5. Handle results
func handleClassificationResults(request: VNRequest?, error: Error?) {
if let results = request?.results as? [VNClassificationObservation] {
// Do something with results
}
}

Problem solved
鉁�
Label: "Border"
1. Detect game box (rectangle
detection)
2. Classify Agricola piece within
rectangle (image classi铿乧ation)

Capabilities
...robust Python training
framework
鈫� style transfer
鈫� activity classi铿乧ation
鈫� image similarity
鈫� recommmendations
鈫� object detection

Setting up python
virtualenv venv
source venv/bin/activate
pip install requests==2.18.4 turicreate==5.0b2 jupyter

The document discusses augmented reality (AR) development tools, particularly focusing on the ARToolKit library, which is an open-source tool for marker-based tracking in AR applications. It highlights various ARToolKit versions, their functionality, and application development processes including marker detection and 3D object overlay. Additionally, it touches on limitations of ARToolKit, alternative libraries, and related AR authoring tools.

Drone pptChangik Choi

听

The document describes a parking drone project that uses computer vision techniques to detect parking permits on vehicles. The drone would fly predefined routines to check for valid permits in different parking lots. A Haar feature-based cascade classifier is trained on positive and negative images to detect permits. An SVM model achieves 90% accuracy in recognizing permit numbers. The group developed an Android app to control the drone and process images using OpenCV. Future work includes improving character recognition and adding license plate detection.

Cutting Edge Computer Vision for EveryoneIvo Andreev

听

The document outlines the Global AI Bootcamp focused on computer vision, detailing various aspects such as scenarios for real-life applications, typical tasks, and the deep learning methods involved. It highlights the use of Azure ML services, custom vision for machine learning projects, and training processes, alongside pricing details and example workflows. The content also includes various resources and references to tools for effective implementation in this domain.

5 Ways to Improve Your LiDAR WorkflowsSafe Software

听

This document discusses 5 ways to improve LiDAR workflows using FME software. It begins with an overview of LiDAR and point clouds before addressing each of the 5 ways: 1) simplifying point cloud transformations with FME transformers, 2) preparing data, 3) automating surface model creation, 4) visualizing solutions through 3D city modeling, and 5) expanding tools with FME Hub and third party tools like LAStools. The presentation concludes by emphasizing how FME can help simplify and scale LiDAR workflows.

Lecture 5,6(Comp Vision) [Auto-saved].pptxSadiqaKhan13

听

Towards Machine IntelligenceDaniyal Shahrokhian

听

The document outlines a process for developing a machine learning model, which includes defining the problem, preparing data, choosing a learning algorithm, training a model, evaluating its performance, improving results, deploying it, and establishing a feedback loop. It provides a specific example related to object classification using convolutional neural networks, detailing steps like data collection via web scraping, model training with gradient descent, and deployment on Google Cloud Platform. The emphasis is on continuous improvement through iteration and prototyping.

Advanced Game Development with the Mobile 3D Graphics APITomi Aarnio

听

This document provides an overview of the Mobile 3D Graphics API (M3G), which was designed for 3D graphics on mobile devices. It discusses why developers should use M3G and highlights some of its key features, including scene graphs, dynamic meshes, animation, textures, and more. The document also provides code examples for common tasks like setting up a camera, rendering a rotating cube, and creating animated keyframe sequences.

Annotation tools for ADAS & Autonomous DrivingYu Huang

听

Django Good PracticesSolution4Future

听

The document discusses best practices in Django development, focusing on models, views, templates, and tools like Celery for asynchronous tasks. It covers critical concepts such as signals, migrations, and the importance of class-based views for better code reusability and organization. Additionally, it highlights the importance of avoiding race conditions and includes examples of efficient customizations for user profiles and database operations.

Unsupervised Aspect Based Sentiment Analysis at ScaleAaron (Ari) Bornstein

听

The document outlines the process of performing aspect-based sentiment analysis using Azure Machine Learning services, detailing the end-to-end steps involved in preparing data, training models, and deploying them. Key features include the creation of Azure ML workspaces, configuring compute targets, and utilizing hyperparameter tuning for model optimization. Additionally, it provides examples of training a logistic regression model and deploying it as a web service for predictions.

patternfinders-141205004957-conversion-gate02.pptxsnnit

听

The document discusses the CIFAR-10 dataset for object recognition, which includes 60,000 images divided into 10 mutually exclusive classes, each containing various objects with significant visual variability. It outlines two feature extraction methods for classifying images using SVM and KNN classifiers, highlighting their respective processes and challenges, such as handling different angles and lighting conditions. Results indicate varying levels of accuracy between methods based on the chosen dimensions and classifiers used.

Easy path to machine learningwesley chun

听

The document provides an overview of a presentation about Google Cloud developer tools and an easier path to machine learning. It introduces the speaker and their background and experience. It then outlines the agenda which includes introductions to machine learning and Google Cloud, Google APIs, Cloud ML APIs, and other APIs to consider. It provides examples of using various Cloud ML APIs like Vision, Natural Language, and Speech for tasks like image labeling, text analysis, and speech recognition. The goal is to demonstrate how APIs powered by machine learning can help ease the burden of learning machine learning by allowing users to leverage pre-built models if they can call APIs.

Analytics Zoo: Building Analytics and AI Pipeline for Apache Spark and BigDL ...Databricks

听

The document discusses the integration of Intel's BigDL with Apache Spark to build efficient analytics and AI pipelines, emphasizing the importance of infrastructure over machine learning code. It outlines a structured approach to the AI production journey, including data preparation, modeling, and deployment phases, while illustrating various use cases across industries. The content also highlights the capabilities of Analytics Zoo for simplifying the development of deep learning workflows through high-level APIs and built-in algorithms.

Viktor Tsykunov: Azure Machine Learning ServiceLviv Startup Club

听

The document provides a technical overview of Azure Machine Learning, detailing the requirements, lifecycle processes, and capabilities for deploying machine learning applications, including deep learning solutions. It covers the setup and use of Azure Machine Learning service, including workspace management, experiment tracking, and model deployment, underpinned by practical coding examples. The text also highlights various real-world applications for machine learning and outlines Azure's infrastructure for building, training, and deploying models efficiently.

Ember.js Tokyo event 2014/09/22 (English)Yuki Shimada

听

This document outlines a game development project using Ember.js and WebGL, authored by Yuki Shimada. It details the creation of a web service that enables RPG development, distinguishes between basic and advanced developers, and explains the architecture involving Ruby on Rails, Ember.js, and WebGL technologies. The document also discusses data handling, UI design, and the integration of game mechanics through computed properties and dynamic data binding.

Gitter marionette deckMike Bartlett

听

The document discusses common performance pitfalls and software design mistakes encountered in web development, particularly using Backbone.js and Marionette. It emphasizes the importance of pre-rendering, optimizing event handling, and avoiding tight coupling between components to maintain clean code architecture. Technical debt from gradual transitions between frameworks can lead to significant issues as applications scale.

Marker-based Augmented Monuments on iPhone and iPadEnrico Micco

听

This document discusses marker-based augmented reality on mobile devices. It describes using black and white markers to recognize 3D models and render them over the camera view in real-time. The key steps are importing 3D models, recognizing markers in video frames using OpenCV, and rendering the associated 3D models using OpenGL. The application loads 3D models from XML files, detects markers to identify models, and displays the augmented reality view on iPhone/iPad.

Leaving Flatland: Getting Started with WebGL- SXSW 2012philogb

听

This document discusses getting started with WebGL. It begins with an introduction to WebGL, explaining that it allows 3D graphics in browsers similarly to OpenGL. It then provides examples of what can be done with WebGL, such as data visualization, games, 3D modeling, and more. The document proceeds to explain the basic graphics pipeline and JavaScript API used in WebGL. It concludes by discussing how to set up a basic 3D scene and choose a WebGL library like Three.js or PhiloGL to get started creating WebGL applications.

5 Ways to Optimize Your LiDAR DataSafe Software

听

The document outlines five strategies to optimize lidar data workflows, emphasizing the importance of simplifying data transformation, preparation, surface model creation, and utilizing visualization techniques in 3D city modeling. It also highlights the integration of FME tools and third-party applications to enhance lidar data management, including tips for effective use of these tools. Participants are encouraged to engage in community activities and explore additional resources for improving their lidar data capabilities.

Apple Machine LearningDenise Nepraunig

听

The document discusses machine learning tools and frameworks developed by Apple, particularly Create ML and Turi Create, which simplify the process of integrating machine learning into applications without needing extensive knowledge of math or coding. It outlines various functionalities including regression, text classification, and object detection, highlighting the user-friendly nature of these tools that allow for easy model training and deployment. Additionally, the document provides links for further exploration of these technologies.

Overcoming The Impedance Mismatch Between Source Code And ArchitecturePeter Friese

听

The document discusses the limitations of UML and traditional software modeling tools, highlighting the complexity and inefficiency they introduce. It advocates for using Domain-Specific Languages (DSLs) as a flexible and precise alternative that improves code generation, collaboration, and project communication. The author emphasizes the importance of choosing the right abstractions to enhance software development and avoid creating 'shelfware'.

Mirko Lucchese - Deep Image ProcessingMeetupDataScienceRoma

听

This document summarizes a presentation on deep image processing and computer vision. It introduces common deep learning techniques like CNNs, autoencoders, variational autoencoders and generative adversarial networks. It then discusses applications including image classification using models like LeNet, AlexNet and VGG. It also covers face detection, segmentation, object detection algorithms like R-CNN, Fast R-CNN and Faster R-CNN. Additional topics include document automation using character recognition and graphical element analysis, as well as identity recognition using face detection. Real-world examples are provided for document processing, handwritten letter recognition and event pass verification.

Scala.IO 2024: Mill builds in Scala 3, a migration storyJames Thompson

听

Javascript Design PatternsIv谩n Fern谩ndez Perea

听

This document provides an overview of JavaScript design patterns based on Addy Osmani's book "Essential JavaScript & jQuery Design Patterns". It begins with background on design patterns and defines what a design pattern is. It describes the structure of design patterns and discusses anti-patterns. It then covers common JavaScript design patterns including creational, structural, and behavioral patterns as well as MV* patterns like MVC, MVP, and MVVM. Specific patterns like Module, Observer, Command, Constructor & Prototype, and examples using Backbone.js, Spine.js, and Knockout.js are summarized.

Easy path to machine learning (Spring 2021)wesley chun

听

The document discusses the role of Google Cloud Developer Tools in facilitating machine learning for developers and educators. It covers various Google Cloud services, APIs, and tools that enable users to access and utilize machine learning without extensive technical knowledge, emphasizing the ease of using APIs. Additionally, it highlights examples of machine learning applications and offers guidance on getting started with Google Cloud's infrastructure and resources.

Cloudera Data Science ChallengeMark Nichols, P.E.

听

The document describes the Cloudera Data Science Challenge, which involves solving three data science problems using large datasets. For the first problem, Smartfly, the goal is to predict flight delays using historical flight data and machine learning algorithms like logistic regression and SVM. The second problem, Almost Famous, involves statistical analysis of web log data and filtering for spam. The third problem, Winklr, requires analyzing a social network graph to recommend users to follow. The document discusses the approaches, tools, and results for each problem.

Data Science Challenge presentation given to the CinBITools Meetup GroupDoug Needham

听

The document describes the Cloudera Data Science Challenge, which involves solving three data science problems using large datasets. For the first problem, Smartfly, the goal is to predict flight delays using historical flight data and machine learning algorithms like logistic regression and SVM. The second problem, Almost Famous, involves statistical analysis of web log data and filtering for spam. The third problem, Winklr, requires social network analysis to recommend users to follow on a social media platform based on click data. The document discusses the approaches, tools, and algorithms used to solve each problem at scale using Apache Spark and Hadoop technologies.

From Zero to Hero 鈥� Web PerformanceSebastian Springer

听

The document discusses various aspects of optimizing web performance, particularly focusing on JavaScript and the rendering process within browsers. Key topics include the critical rendering path, the impact of caching and service workers, strategies to minimize render-blocking, and the performance advantages of using CSS animations over JavaScript animations. Additionally, it highlights issues like memory leaks and garbage collection while offering insights into how modern JavaScript engines enhance execution efficiency.

FUNDAMENTALS OF COMPUTER ORGANIZATION AND ARCHITECTUREShabista Imam

听

Modern multi-proposer consensus implementationsFran莽ois Garillot

听

Multi-proposer consensus protocols let multiple validators propose blocks in parallel, breaking the single-leader throughput bottleneck of classic designs. Yet the modern multi-proposer consensus implementation has grown a lot since HotStuff. THisworkshop will explore the implementation details of recent advances 鈥� DAG-based approaches like Narwhal and Sui鈥檚 Mysticeti 鈥� and reveal how implementation details translate to real-world performance gains. We focus on the nitty-gritty: how network communication patterns and data handling affect throughput and latency. New techniques such as Turbine-like block propagation (inspired by Solana鈥檚 erasure-coded broadcast) and lazy push gossip broadcasting dramatically cut communication overhead. These optimizations aren鈥檛 just theoretical 鈥� they enable modern blockchains to process over 100,000 transactions per second with finality in mere milliseconds鈥� redefining what is possible in decentralized systems.

More Related Content

Similar to Navigating the Apple ML Landscape (20)

Django Good PracticesSolution4Future

听

Unsupervised Aspect Based Sentiment Analysis at ScaleAaron (Ari) Bornstein

听

patternfinders-141205004957-conversion-gate02.pptxsnnit

听

Easy path to machine learningwesley chun

听

Analytics Zoo: Building Analytics and AI Pipeline for Apache Spark and BigDL ...Databricks

听

Viktor Tsykunov: Azure Machine Learning ServiceLviv Startup Club

听

Ember.js Tokyo event 2014/09/22 (English)Yuki Shimada

听

Gitter marionette deckMike Bartlett

听

Marker-based Augmented Monuments on iPhone and iPadEnrico Micco

听

Leaving Flatland: Getting Started with WebGL- SXSW 2012philogb

听

5 Ways to Optimize Your LiDAR DataSafe Software

听

Apple Machine LearningDenise Nepraunig

听

Overcoming The Impedance Mismatch Between Source Code And ArchitecturePeter Friese

听

Mirko Lucchese - Deep Image ProcessingMeetupDataScienceRoma

听

Scala.IO 2024: Mill builds in Scala 3, a migration storyJames Thompson

听

Javascript Design PatternsIv谩n Fern谩ndez Perea

听

Easy path to machine learning (Spring 2021)wesley chun

听

Cloudera Data Science ChallengeMark Nichols, P.E.

听

Data Science Challenge presentation given to the CinBITools Meetup GroupDoug Needham

听

The document describes the Cloudera Data Science Challenge, which involves solving three data science problems using large datasets. For the first problem, Smartfly, the goal is to predict flight delays using historical flight data and machine learning algorithms like logistic regression and SVM. The second problem, Almost Famous, involves statistical analysis of web log data and filtering for spam. The third problem, Winklr, requires social network analysis to recommend users to follow on a social media platform based on click data. The document discusses the approaches, tools, and algorithms used to solve each problem at scale using Apache Spark and Hadoop technologies.

From Zero to Hero 鈥� Web PerformanceSebastian Springer

听

Django Good PracticesSolution4Future

听

Unsupervised Aspect Based Sentiment Analysis at ScaleAaron (Ari) Bornstein

听

patternfinders-141205004957-conversion-gate02.pptxsnnit

听

Easy path to machine learningwesley chun

听

Analytics Zoo: Building Analytics and AI Pipeline for Apache Spark and BigDL ...Databricks

听

Viktor Tsykunov: Azure Machine Learning ServiceLviv Startup Club

听

Ember.js Tokyo event 2014/09/22 (English)Yuki Shimada

听

Gitter marionette deckMike Bartlett

听

Marker-based Augmented Monuments on iPhone and iPadEnrico Micco

听

Leaving Flatland: Getting Started with WebGL- SXSW 2012philogb

听

5 Ways to Optimize Your LiDAR DataSafe Software

听

Apple Machine LearningDenise Nepraunig

听

Overcoming The Impedance Mismatch Between Source Code And ArchitecturePeter Friese

听

Mirko Lucchese - Deep Image ProcessingMeetupDataScienceRoma

听

Scala.IO 2024: Mill builds in Scala 3, a migration storyJames Thompson

听

Javascript Design PatternsIv谩n Fern谩ndez Perea

听

Easy path to machine learning (Spring 2021)wesley chun

听

Cloudera Data Science ChallengeMark Nichols, P.E.

听

Data Science Challenge presentation given to the CinBITools Meetup GroupDoug Needham

听

From Zero to Hero 鈥� Web PerformanceSebastian Springer

听

Recently uploaded (20)

FUNDAMENTALS OF COMPUTER ORGANIZATION AND ARCHITECTUREShabista Imam

听

Modern multi-proposer consensus implementationsFran莽ois Garillot

听

Introduction to Python Programming Languagemerlinjohnsy

听

Deep Learning for Natural Language Processing_FDP on 16 June 2025 MITS.pptxresming1

听

This gives an introduction to how NLP has evolved from the time of World War II till this date through the advances in approaches, architectures and word representations. From rule based approaches, it advanced to statistical approaches. from traditional machine learning algorithms it advanced to deep neural network architectures. Deep neural architectures include recurrent neural networks, long short term memory, gated recurrent units, seq2seq models, encoder decoder models, transformer architecture, upto large language models and vision language models which are multimodal in nature.

How Binning Affects LED Performance & Consistency.pdfMina Anis

听

馃攳 What鈥檚 Inside: 馃摝 What Is LED Binning? 鈥� The process of sorting LEDs by color temperature, brightness, voltage, and CRI 鈥� Ensures visual and performance consistency across large installations 馃帹 Why It Matters: 鈥� Inconsistent binning leads to uneven color and brightness 鈥� Impacts brand perception, customer satisfaction, and warranty claims 馃搳 Key Concepts Explained: 鈥� SDCM (Standard Deviation of Color Matching) 鈥� Recommended bin tolerances by application (e.g., 1鈥�3 SDCM for retail/museums) 鈥� How to read bin codes from LED datasheets 鈥� The difference between ANSI/NEMA standards and proprietary bin maps 馃 Advanced Practices: 鈥� AI-assisted bin prediction 鈥� Color blending and dynamic calibration 鈥� Customized binning for high-end or global projects

Intro狠狠撸s-June-GDG-Cloud-Munich community gathering@Netlight.pdfLuiz Carneiro

听

special_edition_using_visual_foxpro_6.pdfShabista Imam

听

IPL_Logic_Flow.pdf Mainframe IPLMainframe IPLKhadijaKhadijaAouadi

听

Tesla-Stock-Analysis-and-Forecast.pptx (1).pptxmoonsony54

听

Cadastral MapsGoogle

听

A Cluster-Based Trusted Secure Multipath Routing Protocol for Mobile Ad Hoc N...IJCNCJournal

听

Mobile Ad Hoc Network (MANET) is a self-organizing and flexible system. MANET systems manage sensitive data from many distinct applications in various domains. Its dynamic nature increases its vulnerability to numerous types of security threats. Many of the present approaches using indirect approaches provide false approximations of trust degrees. It is significantly required a good routing system that meets Quality of Service (QoS) standards and enhances network performance. In this paper purposed cluster-based trustworthy safe multipath routing (CTSMP-Routing) for mobile ad hoc networks (MANETs). Load balancing challenge is addressed by using a modified proportional topology optimization (MPTO) approach using geographical data related to network nodes. The Enhanced Seeker Search Optimization (ESSO) approach is used to compute trust degrees after the clustering phase considering numerous network constraints including node mobility, received signal strength, energy consumption, and cooperation rate. Assumed to be the service node, the node showing the highest degree of trust manages inter-cluster routing. We have developed a hybrid soft computing approach termed the multi-layer deep recurrent neural network (ML-DRNN) to enhance the optimal path-finding process. This method selects, among many routes between source and destination nodes, the best one quickly. The outcomes of this paper demonstrate that CTSMP-Routing provides effective protection against several attack paths within the MANET environment and displays better performance in regard to quality of service (QoS) requirements.

Industrial internet of things IOT Week-3.pptxKNaveenKumarECE

听

AI_Presentation (1). Artificial intelligenceRoselynKaur8thD34

听

NEW Strengthened Senior High School Gen Math.pptxDaryllWhere

听

60 Years and Beyond eBook 1234567891.pdfwaseemalazzeh

听

Fundamentals of Digital Design_Class_12th April.pptxdrdebarshi1993

听

Center Enamel can Provide Aluminum Dome Roofs for diesel tank.docxCenterEnamel

听

Complete guidance book of Asp.Net Web APIShabista Imam

听

Unlock the full potential of modern web development with the Complete Guidance Book of ASP.NET Web API鈥攜our all-in-one resource for mastering RESTful services using Microsoft鈥檚 powerful ASP.NET Core framework. This book takes you on a step-by-step journey from beginner to expert, covering everything from routing and controllers to security, performance optimization, and real-world architecture.

Industry 4.o the fourth revolutionWeek-2.pptxKNaveenKumarECE

听

Fundamentals of Digital Design_Class_21st May - Copy.pptxdrdebarshi1993

听

FUNDAMENTALS OF COMPUTER ORGANIZATION AND ARCHITECTUREShabista Imam

听

Modern multi-proposer consensus implementationsFran莽ois Garillot

听

Introduction to Python Programming Languagemerlinjohnsy

听

Deep Learning for Natural Language Processing_FDP on 16 June 2025 MITS.pptxresming1

听

How Binning Affects LED Performance & Consistency.pdfMina Anis

听

Intro狠狠撸s-June-GDG-Cloud-Munich community gathering@Netlight.pdfLuiz Carneiro

听

special_edition_using_visual_foxpro_6.pdfShabista Imam

听

IPL_Logic_Flow.pdf Mainframe IPLMainframe IPLKhadijaKhadijaAouadi

听

Tesla-Stock-Analysis-and-Forecast.pptx (1).pptxmoonsony54

听

Cadastral MapsGoogle

听

A Cluster-Based Trusted Secure Multipath Routing Protocol for Mobile Ad Hoc N...IJCNCJournal

听

Industrial internet of things IOT Week-3.pptxKNaveenKumarECE

听

AI_Presentation (1). Artificial intelligenceRoselynKaur8thD34

听

NEW Strengthened Senior High School Gen Math.pptxDaryllWhere

听

60 Years and Beyond eBook 1234567891.pdfwaseemalazzeh

听

Fundamentals of Digital Design_Class_12th April.pptxdrdebarshi1993

听

Center Enamel can Provide Aluminum Dome Roofs for diesel tank.docxCenterEnamel

听

Complete guidance book of Asp.Net Web APIShabista Imam

听

Industry 4.o the fourth revolutionWeek-2.pptxKNaveenKumarECE

听

Fundamentals of Digital Design_Class_21st May - Copy.pptxdrdebarshi1993

听

Navigating the Apple ML Landscape

1. Navigating the ! ML Landscape Meghan Kane, @meghafon developer @novoda, Berlin Swift & Fika, 9/2018

4. ! before we pack our bags..

5. why use machine learning .. at all?

6. ML tasks 鈫� detection: rectangles, face, barcode, text 鈫� classi铿乧ation: image, text, activity, column data 鈫� style transfer 鈫� image similarity 鈫� object detection 鈫� natural language processin 鈫� regression

8. Advantages of on device inference 1. user privacy 2. low latency user experience

9. Let's start our road trip

11. Classify board game pieces

12. ML approach Label: "Border" 1. Detect game box (rectangle detection) 2. Classify Agricola piece within rectangle (image classi铿乧ation)

13. Capabilities ... built right into the Vision framework, no model training needed 鈫� detection: rectangles, face, barcode, text 鈫� object tracking 鈫� image alignment

14. How? 1. create handler 2. create task speci铿乧 request 3. send request to handler 4. handle results

15. // 1. Create handler let handler = VNImageRequestHandler(cgImage: image, orientation: orientation, options: [:])

16. // 2. Create request let request = VNDetectRectanglesRequest(completionHandler: self.handleDetectedRectangles)

17. // 3. Send request to handler do { try handler.perform([request]) } catch let error as NSError { // handle error return }

18. // 4. Handle results func handleDetectedRectangles(request: VNRequest?, error: Error?) { if let results = request?.results as? [VNRectangleObservation] { // Do something with results [*bounding box coordinates*] } }

19. ML approach Label: "Border" 1. Detect game box (rectangle detection) 2. Classify Agricola piece within rectangle (image classi铿乧ation)

20. Why restrict input image to the box? 鈫� easier to train an accurate model 鈫� faster to collect image data

21. Capabilities In Xcode playground, train custom model for: 鈫� image classi铿乧ation 鈫� text classi铿乧ation 鈫� classi铿乧ation & regression of column data

22. Collect Data 鈫� Collect images representative of real world use cases 鈫� Vary angle & lighting 鈫� >10 images per label, but ideally more 鈫� Equal # images for each label 鈫� Recommended: >299x299 pixels

23. Collecting Data Quickly // Extract .jpg frames from .mov @ 5 frames/sec ffmpeg -i stone.mov -r 5 data/stone/stone_%04d.jpg

24. Prepare Data 鈫� Split data: 80% train / 20% test

28. what's happening behind the scenes?

29. Transfer Learning

31. Evaluate

33. Save model

34. Capabilities 鈫� perform predictions using model 鈫� quantized weights (32 bit -> 16, 8, 4... bit) 鈫� perform batch predictions 鈫� create custom model layer

35. Vision + Core ML 1. create Vision Core ML model 2. create handler 3. create task speci铿乧 request 4. send request to handler 5. handle results

36. // 1. Create Vision Core ML model let model = AgricolaPieceClassifier() guard let visionCoreMLModel = try? VNCoreMLModel(for: model.model) else { return }

37. // 2. Create handler let handler = VNImageRequestHandler(cgImage: cgImage, orientation: cgImageOrientation)

38. // 3. Create request let request = VNCoreMLRequest(model: visionCoreMLModel, completionHandler: self.handleClassificationResults)

39. // 4. Send request to handler do { try handler.perform([request]) } catch let error as NSError { // handle error return }

40. // 5. Handle results func handleClassificationResults(request: VNRequest?, error: Error?) { if let results = request?.results as? [VNClassificationObservation] { // Do something with results } }

41. Problem solved 鉁� Label: "Border" 1. Detect game box (rectangle detection) 2. Classify Agricola piece within rectangle (image classi铿乧ation)

42. Next challenge...

44. Capabilities ...robust Python training framework 鈫� style transfer 鈫� activity classi铿乧ation 鈫� image similarity 鈫� recommmendations 鈫� object detection

45. Setting up python virtualenv venv source venv/bin/activate pip install requests==2.18.4 turicreate==5.0b2 jupyter

46. Train style transfer jupyter notebook