DJL Monthly Update - February 2021

[zachgk]

Released DJL 0.10.0 (Release Notes)

Community

The DJL team uploaded a DJL 101 video series. It is about 1.5 hours total (8 videos of ~10 minutes each) and covers setup, ndarrays, training, and inference https://www.youtube.com/watch?v=J4wUnxOYN8M&list=PLC1JzXeHJitDmZBHupMGZE2zHzU8fEvfD&index=2&t=2s

Patrick Arnold published an article titled "AWS Deep Java Library – Was ist das?" in AUSY (written in German) https://www.ausy-technologies.de/de/insight/aws-deep-java-library---was-ist-das-~i3706

@lanking520 published the blog post "Train “Undying” Flappy Bird Using Reinforcement Learning on Java" https://towardsdatascience.com/train-undying-flappy-bird-using-reinforcement-learning-on-java-98ff68eb28bf

NVIDIA published a blog post on using DJL on GPU-powered Spark for running inference on deep learning models https://developer.nvidia.com/blog/accelerating-deep-learing-with-apache-spark-and-gpus-on-aws/

Events

@FrankLiu and @skirdey presented on DJL to the Chicago JUG on 2/18 https://www.youtube.com/watch?v=gBy99IhjSSA.

@roywei and @skirdey presented in the St. Louis Apache Kafka Meetup on using DJL to deploy models from streaming applications on March 3rd https://www.meetup.com/Saint-Louis-Kafka-meetup-group/events/276466753/?isFirstPublish=true

@goswamig and @skirdey will present on deploying deep learning models using DJL from Scala-based big data pipelines in the Bay Area Big Data and Cloud Meetup group on March 30 https://www.meetup.com/datariders/events/276592801/

Feature Enhancements

• Paddle 2.0.0 published
• Model serving auto scale feature
• Make default engine deterministic when multiple engines are available
• Implemented the norm operator
• Add DJL serving worker idle time, (Serving configure idle timing #613)
• Improve the PyTorch index performance (Use PyTorch index to improve the performance #638)
• Improvements for MultiThreadedBenchmark, (Various fixes and improvements for MultiThreadedBenchmark #617)
• Add support for java-9 automatic module name, (Add automatic module to each jar manifest #627)
• Add PaddlePaddle OCR notebook (improve documentation and fix doc build error #640)
• List input and output classes when loading model zoo model (List input and output classes in ModelZoo lookup #624)
• Make XavierInitializer by default for all weights and improve setInitializer (Make XavierInitializer default value & Improve setInitializer #664)
• Updated SageMaker notebook docker image to support execute system command
• Adding Central add metadata info, (Central add metadata info #674)
• Make embedding operator generic across engines (Make embedding generic #688)
• Added Symbolic Model loading and saving mechanism (SymbolBlock Serialization #687) (DJL parameter saving #669)
• Adds a helper ObjectDetectionDataset (Create ObjectDetectionDataset #683)

Documentation & Examples

• [PaddlePaddle] OCR example (Add DJL Paddle OCR Example #568)
• Updated tutorial and miscellaneous documentation (Add purpose to block tutorial and miscellaneous docs #607)
• Edited community page Minor edits to community page #576 (https://github.com/awslabs/djl/pull/576/files)
• Added demo project to show case who to build java-9 modular application with DJL
• Added SageMaker Studio Notebook setup for DJL
• Document DJL forums (Document DJL forums #646)
• Add guide to engines (Add guide to engines #660)
• Document applications, link from Javadoc to D2L, fix tutorial ([DOCS] Document applications, link from Javadoc to D2L, fix tutorial #673)

D2L Book

• [D2L] Complete Section 2.4: Calculus (deepjavalibrary/d2l-java@c501d07)
• [D2L] Complete Section 2.5: Automatic Differentiation (Adding section 2.5: Automatic Differentiation d2l-java#80)
• [D2L] Completed section 2.6: Probability (Adding section 2.6: probability d2l-java#82)
• [D2L] Complete Section 2.7: Documentation (Adding section 2.7: Documentation d2l-java#81)
• [D2L] Completed section 8.1: Sequence Models (Adding section 8.1: Sequence Models d2l-java#83)
• [D2L] Completed section 8.2: Sequence Models (Adding section 8.2: Text Preprocessing d2l-java#84)
• [D2L] Completed section 8.3: Language Models and the Dataset (Adding section 8.3: Language Models and the Dataset d2l-java#85)
• [D2L] Completed section 8.4: Recurrent Neural Networks (Adding section 8.4: Recurrent Neural Networks d2l-java#86)

Bug Fixes

• Rename CheckpointsTrainingListener to SaveModelTrainingListener (Rename CheckpointsTrainingListener to SaveModelTrainingListener #573)
• Fixed model zoo supported engine bug for DLR, TFLite and OnnxRuntime model zoo
• Upgraded maven wrapper script for Windows to 0.5.6
• Fixed gradle wrapper TLS issue for Windows
• Upgrade TFLite to 2.4.1
• Fixed Model View bug to prevent crashing on a null model value
• Refactor the initialize workflow for easier customization (Refactor initialize method #675)
• Fixed broken jupyter notebook
• Refactor BERT block to remove unnecessary overloading forward (Remove unnessary overloading forward #696)
• Refactored PyTorch pipeline to remove extra upload stage
• Fixed model zoo supported engine bug (Fix model zoo supported engine bug #616)
• Fixed broken jupyter notebooks
• Fixed PyTorch training performance issue (DJL training on CPU is much slower on python #621)
• Investigated and resolved IntelliJ plugin classloader question (Jar incompatibility #622)
• Investigated and resolved java.nio.ReadOnlyBufferException (java.nio.ReadOnlyBufferException when calling TensorFlow model.  #570)
• Fixed softmax from_logit flag and RNN tests (fix softmax flag behavior #639)
• Fixed Jacoco code coverage report to reflect real coverage (exclude example and integration code)
• Fixed PaddlePaddle engine co-exist with MXNet issue
• Fix formatCpp linter (Format Cpp #682)
• Enabled Codecov report for DJL (add codecov #679)
• Updates BERT example including testing (Update BERT example #681)
• Removes old DataManager class (Delete DataManager #691)