1.0.1
Features
- Improvements for Björck initializers.
- Stride handling in convolutionnal layers.
Bug fixes
- Fixed bug with
ScaledL2NormPooling
which was causingnan
values to appear after the first training step.
ScaledL2NormPooling
which was causing nan
values to appear after the first training step.