This small release includes support for custom metrics and objects. Prior to this release, DeepReplay expected a metric name like acc as argument and would fail if an instance of a metric object were passed as argument instead. Now, it takes either a metric name or an instance of a metric object (custom or not).

Fixes

• Setting default movie writer back to FFMPEG
• Applied identity function to input tensor to comply with changes introduced by TensorFlow 1.8 - thanks to @albb762 :-)

ReplayData Callback

• model parameter included, for fetching model initial conditions without having to fit it

Replay

• new visualizations included:
  • violin plot for the weights (build_weights method)
  • violint plots for the Z-values/Activations (build_outputs method)
  • violin plot for the gradients (build_gradients method)
• new properties
  • weights_std: standard deviation of the weights per layer
  • gradients_std: standard deviation of the gradients per layer

Datasets

• two new datasets included:
  • ballfor generating points with zero mean and unit standard deviation inside a n-ball
  • hypercube for generating points for the corners of a n-dimensional hypercube