DJL News - January 2021

[zachgk]

Community

@lanking520 gave a presentation on DJL at Baidu’s PaddlePaddle annual summit (https://www.itdks.com/Home/Act/apply?id=5504, DJL presentation starts at 37”45).

@skirdey wrote the blog post "How Netflix uses Deep Java Library (DJL) for distributed deep learning inference in real-time" (https://aws.amazon.com/blogs/opensource/how-netflix-uses-deep-java-library-djl-for-distributed-deep-learning-inference-in-real-time/).

@skirdey wrote the blog post "Deep Java Library - benchmarking" (https://stanislav-kirdey.medium.com/deep-java-library-benchmarking-924b829051fd).

@zachgk gave the talk “Achieve real-time inference at scale on AWS with Deep Java Library” at AWS re:Invent (https://virtual.awsevents.com/media/0_j32zjk5i).

@FrankLiu will give a presentation in the Virtual JUG (Java User Group) community on Mar 2nd. Watch out for event announcement (https://www.meetup.com/virtualJUG/events/).

@frankfliu and @skirdey will give the presentation “Introducing Deep Java Library for building and deploying AI applications in Java” at the San Francisco Bay JUG (Java User Group) community on Mar 10 (https://www.meetup.com/sfjava/events/275912241/).

@roywei has a presentation scheduled for the Kafka community on using DJL to deploy models from streaming applications via a virtual talk that’s going to be broadcasted to the members in their Slack channel https://join.slack.com/t/confluentcommunity/shared_invite/zt-ljvhn34l-7irJW2Q_Qyr4Eh2b1WgNAg). The date is not confirmed, but should be in February.

@FrankLiu and @skirdey will give a presentation in the Chicago JUG on 2/18 (https://www.meetup.com/ChicagoJUG/events/276120748/).

Feature Enhancements

• Upgraded ONNXRuntime version to 1.5.2
• Added Eclipse support to make it easy to import DJL into Eclipse
• EasyHPO Class to simplify training with Hyperparameters
• Start building DJL-Zero for zero-DL knowledge training and discuss how to handle finding models for zero-DL knowledge inference
• Created a AbstractSymbolBlock to refactor all derived SymbolBlock
• Finished PaddlePaddle engine with all JNI configurations
• Upgraded PyTorch version to 1.7.1
• Released DJL Android package 0.9.0
• Enabled CodeQL check for continuous build
• Merged paddlepaddle engine JNI code
• Refactored model serving status API
• Refactored request parsing
• Added support for eclipse developer to make it easy to import DJL into eclipse
• Rename DJL-Easy to DJL-Zero as wrapper for Zero Deep Learning Knowledge Training
• Merge BERT contribution
• Refactor RNN-related operator to be more generic across PyTorch and MXNet
• Add PyTorch RNN/GRU/LSTM operator support
• StopGradient and ScaleGradient Issue (scale gradient for backward pass #521)
• Add dataset.makeTranslator (Start dataset.makeTranslator pattern #547)
• Add DJL Twitter to djl.ai for live content updates on site
• Simplify Zero with Dataset metadata classes
• Upgrade tflite to 2.4.0
• Upgrade TF engine to 2.4.1 and enable XLA build
• Added rotate90 operators
• Added PaddlePaddle RC1 Linux/Windows updates
• Trackers use builders to match DJL conventions
• DJL serving scale workers automatically

Documentation & Examples

• Added PyTorch training and TensorFlow inference for the Amazon Review model.
• Added paddlepaddle engine README
• Add tflite engine README
• Update adoptopenjdk8 install instruction
• Add javadoc for Criteria builder API
• [D2L] Completed section 2.2: Data Preprocessing
• [D2L] Completed section 2.3: Linear algebra
• [D2L] Completed section 2.4: Calculus

Bug Fixes

• Fixed format java gradle plugin to allow run on non-unicode charset platform: gradle-build java-formatter fails on unicode character (api->ProgressBar) when build run in non utf-8 system (gradle-build java-formatter fails on unicode character (api->ProgressBar) when build run in non utf-8 system  #493)
• Fixed OrtSymbolBlock leak issue
• Closed a bunch of stale github issues
• Addressed CodeQL warnings
• Fixed PyTorch Android build pipeline
• Added SentimentAnalysis unit test
• Fix NDArray Splicing Bug Issue (Possible NDArray Splicing Bug #383)
• Fix windows nightly build pipeline
• Improved TensorFlow SSD post-processing code to cover more models