v2.1.7

• Added Hard Sigmoid activation type to Python interfaces
• Speed up Retarget Trainer
• Wrap.py fixes for C++ interfaces
• Cleanup and (minor) optimization of FFTNode
• Cleanup headers in utilities, emitters, and model to speed up builds
• Python logging instead of print

v2.1.6

• Signal processing function and nodes
• Updated archive name for Map
• Fix bug in predict wrapper to support loop execution
• drivetest distributed testing enhancements

v2.1.5

• Raspberry Pi Zero support (by Eric Tang, #123)
• model_test segfault fix for VS 2017
• New interface for emitting control-flow instructions
• IRLocalValue classes to simplify emitting IR
• Rename DynamicMap to Map
• SWIG generation of step-based user programming model
• Fix to #125
• ccache support

v2.1.4

• Math library updates to unify Vector, Matrix, and Tensor Operations functions
• ProtoNNTrainer V2
• Support the "Step" programming model in ell.py
• Fix Javascript code for the ELL_* SWIG interface change
• Customization fixes for compile.exe

v2.1.3

• Multi-node retarget support.
• Renamed python namespaces to use lowercase "ell".
• New "dr" models to the gallery.

v2.1.2

• macOS build fix.
• prefer OpenBLAS libraries by default.
• in-progress work on steppable maps.

v2.1.1

• 6 new "dcda" 64x64 ILSVRC2012 models in the gallery.
• Property-bag-type metadata and archiver improvements.
• Renamed ELL python model to "ell" (lowercase).
• Support for combining ELL models into the same C++ application.
• Fix 32 bit Linux build.

v2.1.0

• Improved support for ProtoNN trainer and compiling ProtoNN models are now available. Introductory tutorial here.
• Added the following models:
  • stacked: dsf_I128x128x3CMCMCMCMCMCCCMC1AS, dst_I128x128x3CMCMCMCMCMCCCMC1AS, dsb_I128x128x3CMCMCMCMCMCCCMC1AS
  • no maxpool: dscs2_I128x128x3CCCCCCC1AS, ddcs2_I128x128x3CCCCCCCCCCCCCC1AS
• Laying the foundation for RNN, LSTM, and GRU layers
• Improved steppable maps. Tutorial soon.
• Added the datasetFromImages tool, which aids in creating ELL compatible datasets from images, which can then be used by ELL's trainers to produce a trained predictor.
• Fixes for bugs and compiler warnings throughout the codebase

v2.0.5

Fixes to import tutorial.
IRMapCompiler fixes.
Logging support.
Math library updates.
Early retargeting work (in progress).
Early protoNN work (in progress).

v2.0.4

Corrected naming of 64x64 stacked models.