jChampions JSR381 Visual recognition for java
- 1. JSR 381 Visual Recognition for Java Finally, A Java-Friendly ML API Gran Sasso
- 2. Frank Greco @frankgreco Chair NYJavaSIG - NY Java User Group JC, Cloud/Mobile Architect Zoran Severac @neuroph AI Researcher Univ of Belgrade, Serbia JC, Apache NetBeans PMC NullPointerExceptional
- 3. Session Topics ● Patterns and Machine Learning (ML) ● Why is ML Different than Conventional Programming ● Java and ML ● VisRec JSR 381 ● VisRec Architecture and Implementation ● Examples
- 5. Machine Learning - Huge Impact on Everything...
- 6. Patterns...
- 8. “One of the most interesting aspects of the world is that it can be considered to be made up of patterns” Norbert Wiener - 1894-1964 - MIT Cybernetics (1948) - early steps to modern AI This was AI before it was called “AI”
- 11. Learning a Foreign Language Understand All Rules Observe/Follow Patterns
- 12. Where Else Does This Apply?
- 13. Where Else Does This Apply? Any application that has a large number of parameters and lots of data
- 15. Ecommerce Language Translation Autonomous Transportation Online Search Call Centers Social Media Monitoring
- 16. Session Topics ● Patterns and Machine Learning (ML) ● Why is ML Different than Conventional Programming ● Java and ML ● VisRec JSR 381 ● VisRec Architecture and Implementation ● Examples
- 17. T F T F T F T F FT FT FT FT FT T F
- 21. ML - Basic Workflow Obtaining, Organizing, Cleaning, Removing Bias, etc. Selecting ML Algorithm, Parameter Experimentation, Analyze Errors Deploy Model with jar/container
- 22. ML - Basic Workflow Obtaining, Organizing, Cleaning, Removing Bias, etc. Selecting ML Algorithm, Parameter Experimentation, Analyze Errors Deploy Model with jar/container
- 23. ML - Basic Workflow Obtaining, Organizing, Cleaning, Removing Bias, etc. Selecting ML Algorithm, Parameter Experimentation, Analyze Errors Deploy Model with jar/container
- 24. Session Topics ● Patterns and Machine Learning (ML) ● Why is ML Different than Conventional Programming ● Java and ML ● VisRec JSR 381 ● VisRec Architecture and Implementation ● Examples
- 25. You can write your Machine Learning app in any language you want as long as its Python.
- 26. You can write your Machine Learning app in any language you want as long as its Python. HEY!
- 28. Java Needs to be a First Class Citizen for ML ● Don’t want another JavaScript “Doomsday” Scenario occurring in our Lifetime ● Python is a very good language. There should be others to express the ML world’s creativity ● Java is a great language with a huge ecosystem with 12M developers ● Existing ML APIs aren’t Java-friendly or AppDev-friendly
- 29. Everyone we talked was interested to collaborate
- 30. Session Topics ● Patterns and Machine Learning (ML) ● Why is ML Different than Conventional Programming ● Java and ML ● VisRec JSR 381 ● VisRec Architecture and Implementation ● Examples
- 31. Don’t boil the ocean… [heat up some coffee first] VisRec - JSR #381 Visual Recognition for Java Spec Leads and Expert Group Zoran Sevarac (ML researcher - Univ of Belgrade) Frank Greco (NYJavaSIG, Consultant) Kevin Berendsen (OpenValue, Senior Software Eng)
- 32. Other contributors and Advisors ● Frank Liu (Amazon) ● Constantin Drabo ● Amit Nagesh ● Marissa Staller ● Eric Bruno ● Anakar Parida ● Nikita Ivanov (Gridgain) ● James Weaver (IBM) ● Werner Keil ● Jyoti Buddha ● Guillaume Laforge (Google) ● Ed Burns (fmr Oracle) ● Nishant Raut (Mumbai JUG) ● Sandhya Kapoor (fmr IBM)
- 33. JSR 381 API, Architecture, Implementations, and Examples
- 34. JSR #381 - Visual Recognition API ● ● ● ● ● ● More readable and easier to use VisRec JSR #381 https://github.com/JavaVisRec/visrec-api/wiki/Getting-Started-Guide
- 35. Issues with Existing Offerings ● ● ● ●
- 36. VisRec Design goals and principles ● ● ● ●
- 37. What Can you Do with It? 1) Classify images and data/objects classification task 2) Estimate numeric value regression task 3) Object Detection classification + regression(position)
- 38. Generic ML Layer Visual Recognition Image Abstraction VisRec Public API Apps/Libs Visual Recognition Layer JSR381 Service Provider Interface ML Engine 1 Service Provider Impl ML Engine 2 Service Provider Impl JSR 381 Architecture
- 39. Core Classes & Design Generic ML Layer (common abstractions, interfaces for general ML tasks) Visual Rec Layer (specialization of generic ML tasks for specific objects / domain) Implementation Based on specific ML engine DataSet<E> ImageSet extends DataSet<ExampleImage> Classifier<T, R> BinaryClassifier<T> MutliClassClassifier<T, R> ImageClassifier<IMAGE_CLASS> AbstractImageClassifier<IMAGE_CLASS , MODEL> implements ImageClassifier<IMAGE_CLASS> ImageClassifierNetwork extends AbstractImageClassifier<BufferedImage, ConvolutionalNetwork> Regressor<T, R> SimpleLinearRegression< MODEL_CLASS> implements Regressor<Float, Float>, SimpleLinearRegressionNetwork extends SimpleLinearRegression<FeedForwardN etwork>
- 40. How to classify image Build Classifier ImageClassifier<BufferedImage> dukeClassifier = NeuralNetImageClassifier.builder() .inputClass(BufferedImage.class) .imageHeight(64) .imageWidth(64) .labelsFile(new File("labels.txt")) .trainingFile(new File("index.txt")) .networkArchitecture(new File("duke_net.json")) .exportModel(Paths.get("duke.dnet")) .maxError(0.05f) .maxEpochs(1000) .learningRate(0.01f) .build() Use Classifier BufferedImage image = ... Map<String, Float> results = classifier.classify(image); Full Example on GitHub
- 41. Running the Demo Epoch:1, Time:686ms, TrainError:0.68538016, TrainErrorChange:0.68538016, TrainAccuracy: 0.91150445 Epoch:2, Time:505ms, TrainError:0.4474997, TrainErrorChange:-0.23788047, TrainAccuracy: 0.96460176 Epoch:3, Time:506ms, TrainError:0.28403035, TrainErrorChange:-0.16346934, TrainAccuracy: 0.96460176 Epoch:4, Time:479ms, TrainError:0.1824384, TrainErrorChange:-0.101591945, TrainAccuracy: 0.9734513 Epoch:5, Time:486ms, TrainError:0.123548426, TrainErrorChange:-0.058889978, TrainAccuracy: 0.99115044 Epoch:6, Time:493ms, TrainError:0.08493269, TrainErrorChange:-0.038615733, TrainAccuracy: 1.0
- 42. Minimal Configuration ● Image class ● Image dimensions ● Category labels (labels file) ● List of image files with corresponding category labels (index file) ● Convolutional neural network architecture for image classification ● Training parameters (max error, max epochs, learning rate) ● File name for the trained model
- 43. Under the hood: Convolutional Neural Network for Image Recognition High dimensional data + automated preprocessing and feature extraction
- 44. Reference Implementation: Deep Netts ● Pure Java implementation of main Deep Learning algorithms: convolutional and feed forward neural networks ● Open source (LGPL) and professional edition ● Pro Edition includes Deep Learning IDE ○ Visual ML Tools, Wizards and ● Starting today Pro Edition is available Under FREE DEVELOPMENT license! https://www.deepnetts.com/download
- 45. Image classification - Convolutional network using Deep Netts API ConvolutionalNetwork neuralNet = ConvolutionalNetwork.builder() .addInputLayer(imageWidth, imageHeight, 3) .addConvolutionalLayer(3, 3, 12) .addMaxPoolingLayer(2, 2) .addFullyConnectedLayer(30) .addFullyConnectedLayer(20) .addOutputLayer(labelsCount, ActivationType.SOFTMAX) .hiddenActivationFunction(ActivationType.TANH) .lossFunction(LossType.CROSS_ENTROPY) .randomSeed(123) .build();
- 47. Next Steps for You! github.com/JavaVisRec/visrec-api/wiki/Getting-Started-Guide
- 48. JSR 381 Visual Recognition for Java Finally, A Java-Friendly ML API Please provide feedback and contribute! No reason why Java cannot be used for ML!
- 49. Thank You!