2.3.6

• Fixed bug affecting PCA initialization
• Fixed documentation of decay functions thanks to @mariajmolina
• Fixed possible propagation of numerical errors in computing the quantization_error thanks to @lorenzoferre
• Improved PCA initialization making it faster and more robust to non scaled data, thanks to @lorenzoferre
• Implemented the offline batch training algorithm, thanks to @lorenzoferre

2.3.5

This is an amendment to release 2.3.4. In 2.3.4 the distortion measure is incorrectly called divergence measure.

2.3.4

• Distortion measure implemented.

2.3.3

• sigma now has a different decay function and it can be selected via the parameter sigma_decay_function
• the user is now warned when sigma is higher than sqrt(x^2+y^2)
• a selection of decay functions for both sigma and learning rate has been implemented by @BrandonGarciaWx
• the train method now accept fixed_points so that the algorithm can be force to train specific neurons for specific samples

2.3.2

• improved structure of hexagonal topology - thanks to @aznt00th
• improved pca initialization
• fixed bug affecting the topographic error
• the main notebooks containing the examples are now covered by CI

2.3.1

• Introduce use_epochs parameter, when True it will keep learning rate and decay constant for an entire epoch. Thanks to @jriege555 for submitting this change
• Topographic Error can now be computed also for hexagonal grids. Thanks to @TharindaDilshan
• Fixed issue with pca initialization

2.3.0

• Activation functions can now be defined via callables, thanks to @chicodelarosa
• The method distance_map now has the parameter neighbour_average that allows the user to normalize the distances. Thanks to @lbugnon

2.2.9

License change.