v3.4.0

Bug fixes and improvements

Global

• Removed implicit softmax from last layer, to allow multi-variate regression (#42)

WebAssembly

• Added fix for Webpack loading of NetWASM.js
• Added net.delete() for clean-up

OutputLayer

• Added OutputLayer class

Examples

• Added example project for loading jsNet through Webpack
• Added example for using multiple WASM network instances

Bug fixes

• WASM misassigned learning rate defaults
• WASM momentum not training

v3.3.0

Misc Improvements

Network

• Added confusion matrix
• Made it possible to pass volume (3D array) input data
• Added callback interval config to .train()
• Added collectErrors config to .train() and .test()

InputLayer

• Added InputLayer class

Examples

• Added MNIST dev enviromnment example

v3.2.0

IMG data, validation, early stopping

Network

• Added weight+bias importing and exporting via images, using IMGArrays
• Added validation config to .train(), with interval config
• Added early stopping to validation, with threshold stopping condition
• Added early stopping patience condition
• Added early stopping divergence condition
• Breaking change: "error" key in training callbacks have been changed to "trainingError"
• Breaking change: Removed ability to use either data keys 'expected' and 'output'. Now just 'expected'.

NetUtil

• Added splitData function
• Added normalize function

NetMath

• Added root mean squared error cost function
• Added momentum weight update function
• Breaking change: Renamed "vanilla update fn" to "vanilla sgd"

v3.1.0

Optimizations

ConvLayer

• Optimized errors structure
• Optimized bias structure
• Optimized activations structure
• Optimized weights structure
• Optimized deltaWeights structure
• Optimized deltaBiase structure

NetUtil

• Optimized convolve

FCLayer

• Optimized weights structure
• Optimized bias structure
• Optimized deltaWeights structure
• Optimized sums structure
• Optimized errors structure and net errors propagation
• Optimized activations structure
• Optimized forward()
• Optimized backward()
• Optimized deltaBias structure

Global

• Changed framework loading to allow choosing between versions at runtime
• Added basic server and browser + nodejs demos for how to load jsNet
• Bug fixes
• Changed the way classes were bundled, to fix some bundler compatibility issues (see #33)

v3.0.0

WebAssembly

WebAssembly

• Complete, rewritten, WebAssembly version of jsNet

Global

• Many bug fixes
• Removed default configuration values: l1, l2
• Added layer specific activation function config for FC layers, and ability to turn it off
• Reworked regularization algorithm
• Reworked layer specific activation function assignments
• Reworked net error propagation, using softmax
• net.forward() function now returns softmax activations by default

JavaScript

• Removed babel transpilation (it's 2018)

v2.1.0

Optimizations

NetUtil

• Optimised addZeroPadding() - ~68% faster
• Optimised uniform() - ~588% faster
• Optimised gaussian() - ~450% faster

FCLayer

• Optimised resetDeltaWeights() and applyDeltaWeights() - ~18% faster (overall)

NetMath

• Optimised softmax() - ~924% faster

ConvLayer

• Restricted filters' dropout maps only to when dropout is configured - less memory usage

Bug Fixes

• Fixed bug caused by minification and disabled name mangling

v2.0.0

Convolutional Networks

Network

• New name: jsNet
• Restructured to allow multiple layer types
• Added conv config for configuring filterSize, zeroPadding, stride ConvLayer default values
• Added pool config for configuring size and stride PoolLayer default values
• Added (input) channels config, used by ConvLayers
• Re-wrote the JSON import/export. Check README for details on backward compatibility
• Removed ability to create a network by just giving layer types in the list
• Can check the version number via Network.version
• Renamed adaptiveLR to updateFn

ConvLayer

• Added ConvLayer.js 🎉 with activation, filterCount, filterSize, zeroPadding and stride configs

Filter

• Added Filter.js

PoolLayer

• Added PoolLayer, with stride and activation configs

NetUtil

• Added NetUtil.js
• Added addZeroPadding
• Added arrayToMap
• Added arrayToVolume
• Added 4 other helper functions

NetMath

• Renamed noadaptivelr to vanillaupdatefn

FCLayer

• Renamed Layer.js to FCLayer.js. Layer still exists as an alias to what is now FCLayer

Bug Fixes

• Fixed training callback giving iteration index, not count (-1)

v1.5.0

Training, Misc

Network

• Made string configs (activation, adaptiveLR, cost, distribution) case/underscore/space insensitive.
• Allow custom activation functions to be configured
• Allow custom cost functions to be configured
• Allow custom weight distribution functions to be configured
• Added time elapsed to training/testing logs and elapsed milliseconds to training callback
• Added log option to training/testing to disable logging for each one
• Added mini batch SGD training, and miniBatchSize config for .train() as its config
• Breaking change: classes are no longer required straight into global context. See readme.
• Added shuffle .train() option
• Added callback .test() option
• Breaking change: Updated default values

NetMath

• Breaking change (if you were using the NetMath functions directly): Made config functions' names lower case

Bug Fixes

• Fixed iteration error logged when testing being the averaged total at that point

v1.4.0

Weights Initialization

Network

• Reworked current weights initialization to be configurable
• Set old weights init to uniform distribution, with additional limit config
• Added mean as weightsConfig option, for gaussian distribution
• Added stdDeviation as weightsConfig option, for gaussian distribution

NetMath

• Added standardDeviation to NetMath
• Added gaussian weights distribution
• Added xavierNormal weights distribution
• Added lecunUniform weights distribution
• Added lecunNormal weights distribution
• Added xavierUniform weights distribution

v1.3.0

Regularization

Network

• Added dropout, with dropout configuration
• Added L2 regularization, via l2 strength config
• Added L1 regularization, via l1 strength config
• Added max norm regularization, via the maxNorm threshold value config

Bug Fixes

• Fixed error value logged accumulating across epochs, instead of resetting to 0
• Fixed epoch counter logged resetting instead of accumulating across training sessions