Skip to content

Latest commit

 

History

History
87 lines (69 loc) · 2.65 KB

CHANGELOG.rst

File metadata and controls

87 lines (69 loc) · 2.65 KB
Raw

SNN Toolbox: Release Notes

Version 0.6.0

Support for Tensorflow 2.4. Added support for Conv1D layers. Implemented spiking maxpool layer for tensorflow backend. More example scripts for ResNets and pytorch models. Minor bugfixes.

Version 0.5.0

Added support for Tensorflow 2.2. The toolbox no longer imports stand-alone Keras, but instead uses Keras only from within Tensorflow (tf.keras). Enabled simulating SNNs in the tensorflow-based INIsim using graph mode rather than eager execution, which results in a speed-up of about 7X. Removed support for python 2. Updated various temporal coding backends.

Version 0.4.1

The toolbox now supports input models from the PyTorch library.

Thanks to Pengfei Sun for contributing.

Version 0.4

The toolbox now supports deploying converted networks on the SpiNNaker architecture!

Thanks to ej159, pabogdan, and rbodo for contributing.

Version 0.3.2

Simulation with Brian2 backend now supports:
  • Constant input currents (less noisy than Poisson input)
  • Reset-by-subtraction (more accurate than reset-to-zero).
  • Bias currents

Thanks to wilkieolin for this contribution.

Version 0.3.1

Bugfixes:
  • Setting biases in convolution layers for pyNN and Brian2 simulator backends.
  • Parsing axis parameter in BatchNorm layers.
  • Counting of SNN operations.
  • Minor issues due to updating to latest keras / tensorflow version.
  • Syntax error in equation for membrane potential due to updating Brian2.
  • Restoring a previously saved SNN to run with INIsim now works again.
  • Fixed issue #25 (permutation of weights after flatten layer in models trained with recent Keras version and simulated with Brian2 / pyNN).
Added support for:
  • Intel's neuromorphic platform "Loihi".
  • Tensorflow 2.0.
  • Parsing depthwise-separable convolutions.
  • Strides > 1 for pyNN and Brian2 simulator backends.
  • Parsing a model can be skipped now by loading a previously saved parsed model.
  • Using SNN toolbox more easily from within a python script instead of via terminal only.
  • Save and load functions for Brian2 networks.
Miscellaneous:
  • Added end-to-end examples for creating and training the model, saving the dataset, and setting up the config file to run SNN toolbox.
  • Moved large model files and datasets to separate repository (snntoolbox_applications) to shrink size of core package.
  • Minor refactoring, repo cleanup, and performance improvements.
Contributors:
  • rbodo
  • sflin
  • nandantumu
  • morth
  • wilkieolin
  • Al-pha