v0.8.2

Main changes:

• Improved functionality for layer/network copying
• Added repeat function
• Fixed output shape for the conv layers when input is unknown
• Typo fixes

v0.8.1

Main changes:

• Added Tensorflow 1.13 support
• Added DropBlock layer
• Added define method for all layers

v0.8.0

Main changes:

• New design for the layer graphs and layers
• Changed inline operators > (still supported) to >> and lists to| operator for parallel connections
• Added 2 new step decay algorithms, namely exponential_decay and polynomial_decay
• New regularizers
• New signals functionality, can be extended with classes (signals attribute)
• Tensorflow optimizers are used instead of self-implemented functions
• Added GroupNorm layer
• Changed behavior for the the predict methods
• Added show method to the networks
• Added plot_errors for the optimizers with better and more accurate visualizations

Small changes:

• The error parameters for optimizers was renamed to loss
• Changed summary format in the logs
• Removed training_errors and validation_errors and it were replaced with new errors attribute
• Renamed the prediction_error method to score

Misc:

• Big code refactoring
• Better exceptions

Removed

• Add-ons were removed
• Tuple of integer as network configuration
• Removed RBFKmeans
• Removed train_end_signal and epoch_end_signal
• Removed plots.error_plot function
• Removed plots.network_structure function

v0.7.3

• Fixed problem that occurred when GNG was loaded from pickled file. Issue #229
• Improved representation for the GNG graph elements
• Added extra exceptions into GNG graph methods in order to get informative fail when unexpected action was applied

v0.7.2

Changes:

• Added missing dilation rates for ResNet-50 architecture
• Fixed fan computation for weight initializers
• Added variable initializer in the predict method
• Change XavierNormal to HeNormal default initializer

v0.7.1

Changes:

• Speed up network initialization via lazy parameter initialization using tensorflow's tensors instead numpy's arrays
• Global pooling layer accepts two string arguments that point to different tensorflow functions.
• Fixes for the reshape layer when used with unknown input shape
• Fixed cross entropy loss functions for spatial inputs
• Removed input blocker during the training
• Combined GradientDescent and MinibatchGradientDescent into one class GradientDescent

v0.7.0

Changes:

• Backend was moved to Tensorflow
• Pickle storage for the weights has been replaced with HDF5
• Changed order of the dimensions for the convolutional filter (channel expected to be in the last dimension)
• The compile method was removed
• Added wolfe search to conjugate gradient
• Fixes for the training algorithms

Removed:

• Linear models
• Quickprop training algorithm
• Ensemble algorithms

v0.6.5

Bugs:

• Fixed iterative updates for discrete hopfield networks: #218

v0.6.4

Features:

• Added Growing Neural Gas (GNG) algorithm
• Added 2 new improvements to the base GNG algorithms

Bug:

• Set up default value for epoch_time in order to prevent crashes when training was stopped on first iteration.

v0.6.3

Fixes:

• Use old table-style, otherwise table breaks for some of the font styles and it doesn't look good in ipython notebook
• Adam optimizer fix for GPU (#200)