v2.3.7

• Move to LLVM 6.0.1
• Move ELL to C++17
• Add support for python 3.7
• Add support for more ONNX operations
• Added instructions for Raspbian Lite (https://microsoft.github.io/ELL/tutorials/Raspberry-Pi-setup/)
• Add new OpenBLAS include path for Ubuntu 18.04
• Fix reorder data node optimization pass
• Fix debugCompiler to work with gallery models
• Improve reordering of data for simple and Winograd depthwise convolutions
• Fixes for "Getting started with image classification on Raspberry Pi using C++" tutorial
• Fix importing of depthwise separable models
• Fix "Repurposing a pretrained image classifier" tutorial

Addresses #166 and #182.

v2.3.6

• Initial version of the ONNX importer
• Tools now have a -v/--verbose option that will control the output of additional helpful logging data
• Demo helper script has been modified to print labels to STDOUT instead of drawing on the image. This is useful for some training sets like CIFAR
• Couple of new options for wrap.py:
  • --convolution_method: this lets you specify the convolution method (it just passes the value on to the compile executable)
  • --llvm-format gets a new option: obj. This, along with --no-llc-tool and --no-opt-tool let you skip the time-consuming llc and opt steps. To retain compatibility with compile, the object file suffix is now .o.

v2.3.5

• Fixed typos in some tutorials (thanks @lisaong )
• Added support for emitting relocatable object code
• Added Travis-CI support (thanks @lisaong )
• Updated profiler to work with models from the ELL gallery
• Fix build break for VS2017.7.3 in StlContainerReferenceIterator
• Fix python interface implementation for LSTM layer nodes
• Fix crash in retarget trainer (issue #141)
• Fix docs for opencv setup on raspberry pi
• Improve support for running models on Raspberry Pi Zero
• make_profiler.sh can now be run from anywhere
• Various bug fixes (#161, #157, #149)

v2.3.4

• Fix importing CNTK models with a Softmax layer
• Add a Reset function to LSTM and GRU models to clear the state

v2.3.3

• Fixes issue #140
• Helps issue #138 by fixing importing of Darknet models with Fully Connected layers
• Simplified and improved LLVM IR optimization in emitted code
• Fix importer for next round of separable darknet models in CNTK
• Improved Winograd convolution implementation
• Fix building the profiler tool
• Fix running build.sh without any command line args
• Miscellaneous bug fixes

v2.3.2

• Fix issue #137
• Added resnet models to gallery
• Add support for naive depthwise separable convolution layers

v2.3.1

• Added more fine grained profiling support for emitted code
• Make find_package for LLVM more resilient to finding the version that ELL needs
• Updated Setting up your Raspberry Pi device for tutorials to move to Raspbian Stretch
• Updated links to YOLOv2 configuration and weights in Getting started with object detection using region of interest networks
• Added better support for importing models from older versions of CNTK

v2.3.0

• New tutorials available:
  • Getting started with object detection using region of interest networks
  • Getting started with audio keyword classification on the Raspberry Pi
• Updated importer to handle the importing of resnet models
• Miscellaneous bug fixes.

v2.2.0

• New tutorial available: Repurposing a pretrained image classifier
  
• Updated Boosting classifier accuracy by grouping categories
  to illustrate how to use callbacks for an asynchronous programming model.
• find_ell.py shouldn't import ell automatically. Should help address issues like #130.
• Added support for importing region detection layers from Darknet models.
• Miscellaneous bug fixes.

v2.1.8

• Fix a bunch of bugs
• Add RegionDetectionLayer that supports darknet regions - see YOLO9000: Better, Faster, Stronger
• Side-by-side tutorial fixes
• Add convenience class that wraps the callback style API for C++ apps
• Add init.py to output of wrap.py so it is easier to import compiled modules
• GRU based audio models are working