Release 3.1.0
Compatible with Nengo 3.0.0
Compatible with TensorFlow 2.0.0 - 2.1.0
Added
- Added
inference_only=True
option to the Converter, which will allow some Layers/parameters that cannot be fully converted to native Nengo objects to be converted in a way that only matches the inference behaviour of the source Keras model (not the training behaviour). (#119)
Changed
- Improved build time of networks containing lots of
TensorNodes
. (#119) - Improved memory usage of build process. (#119)
- Saved simulation state may now be placed on GPU (this should improve the speed of state updates, but may slightly increase GPU memory usage). (#119)
- Changed Converter
freeze_batchnorm=True
option toinference_only=True
(effect of the parameter is the same on BatchNormalization layers, but also has broader effects). (#119) - The precision of the Nengo core build process will now be set based on the
nengo_dl.configure_settings(dtype=...)
config option. Note that this will override the default precision set innengo.rc
. (#119) - Minimum Numpy version is now 1.16.0 (required by TensorFlow). (#119)
- Added support for the new
transform=None
default in Nengo connections (see Nengo#1591). Note that this may change the number of trainable parameters in a network as the scalar defaulttransform=1
weights on non-Ensemble connections will no longer be present. (#128)
Fixed
- Provide a more informative error message if Layer
shape_in
/shape_out
contains undefined (None
) elements. (#119) - Fixed bug in
Converter
when source model contains duplicate nodes. (#119) - Fixed bug in
Converter
forConcatenate
layers withaxis != 1
. (#119) - Fixed bug in
Converter
for models containing passthroughInput
layers inside submodels. (#119) - Keras Layers inside TensorNodes will be called with the
training
argument set correctly (previously it was always set to the default value). (#119) - Fixed compatibility with
progressbar2
version 3.50.0. (#136)