Neupy Versions Save

Fixes:

• Fixed neupy installation problem for cases when user don't have pre-installed Theano package (#198)

Enhancement:

• Switched to Theano version 1.0
• Use tableprint library instead of neupy table module
• Use progressbar2 library instead of neupy progressbar module
• Removed logic that controls number of outputs in terminal during the training

Bugs:

• Convert output from the Step layer to the integer explicitly in order to avoid boolean outputs
• Fixed issue where message, that tells about training interruption, breaks result table

Features:

• Added module that contains popular DNN architectures, namely Resnet50, VGG19, VGG16, SqueezeNet, AlexNet
• Pre-trained parameters for the new DNN architectures
• Changed format in which neupy stores artitectures.
• To the existed pickle format there were added support for a few new formats, namely hdf5, json and python dictionary

Enhancement:

• Changed API for the mixture of experts ensemble. Now it works from architectures module.
• Save pickle files using protocol compatible with python 2 and 3 versions
• New error messages that explain failures during parameter loading in storage module
• Use different parameter loading strategies in storage module

Bugs:

• Fixed issue with PNN class mapping (#177)
• Added SOFM weight normalization to the cosine similarity measurement

Features:

• Pickle serializer for networks with fixed architectures
• SOFM weight initialization with PCA
• Added hexagon shaped grid types to SOFM
• Added parameter reduction over time for SOFM
• Possibility to set up different step sizes for different neigbour neurons in SOFM
• Added support for N-dimensional grid shapes for SOFM
• Max-norm regularization algorithm

Enhancement:

• Made n_outputs as an optional parameter for SOFM if feature_grid was specified (and vice versa)

Bugs:

• Fixed problem with more complicated cases for inline connections

Features:

• Added LVQ algorithm
• Added LVQ2 algorithm
• Added LVQ2.1 algorithm
• Added LVQ3 algorithm
• Added step reduction algorithm into all LVQ versions

Bugs:

• Changed SciPy version in order to fix problem with golden search algorithm

Features:

• Added LSTM layer
• Added GRU layer
• Added customizable weight and bias initialization for LSTM and GRU layers
• Created pointer to all layers in the network and connection
• Ability to extract layer by its name from the connection

Enhancement:

• Theano 0.9.0 support
• Fixed issues related to float16 data type

Bugs:

• Modified algorithm for layer name generation
• Solved problem with Dropout in Wolfe Search algorithm
• Fixed a few minor bugs

Bugs:

• Fixed input and output layers duplication in the connection
• Fixed training issues for networks with shared weights
• Make valid input order for the compile method
• Fixed progress bar appearance for the training with multiple inputs
• Fixed shuffle data option for networks with multiple inputs

Features:

• Added Leaky Relu layer
• Saliency Map plot
• Ability to specify input and output layers for layer connections (start and end methods)
• Added method that compiles network (compile method)

Enhancements:

• Validate that layer name is unique in the network
• Added ability to train networks with multiple inputs

Features:

• Added Global Pooling layer
• Added Concatenate layer
• Added Element-wise layer
• Added Embedding layer
• Added Local Response Normalization layer
• Added Discrete digits dataset
• Parallel connections
• Added layer_structure plot to visualize relations between layers
• Added an ability to save and load weights from the pickle file

Enhancement:

• Improved and modified layer connection API
• Developed graph structure that stores relations between layers
• Set up bias as an optional parameter
• Skip layers for the layer_structure function
• Assign unique identifier for each layer

Bugs:

• Fixed lots of small bug in the layer connection module
• Fixed bugs with Hessian algorithm
• Fixed summary table

Features:

• Added Restricted Boltzmann Machine (RBM) (#64)

Enhancement:

• PNN mini-batch prediction (#72)
• Check that it's possible to connect two layers during layer connection procedure (#114)
• Add function for RBM that makes Gibbs sampling from the visible input for multiple iterations (#115)
• Add more flexible way to initialize network parameters (#110)

Refactoring:

• Use ParameterProperty class instead of ArrayProperty (#111)

Bugs:

• Fixed Quasi Newton algorithm for the training GPU training (#108)
• NeuPy shows NaN output values in summary tables as a dash symbol (#109)