NeuralNet 0.0.0

checksums.yaml

@@ -0,0 +1,7 @@
+---
+SHA256:
+  metadata.gz: e26b008d3e48e32c5bd9e2b258f014736291f725a5ee73f8701dd63ddba83037
+  data.tar.gz: 12845989f991bfd9e9100679f4f167b098557986681ce37734800524ec14a8e7
+SHA512:
+  metadata.gz: fee7708d275efa089dde0677e27fc79af3938f7fd60a60bc654a7b7bad70ea33ab4f9f98c5cd78872735ed4df07a0ac4cc0131aa9be3d50c8fa68382864a2856
+  data.tar.gz: 4c2d72e51bbadc97293739ebfb416104cc11eb19b007297a64a62b15ba6811555640cb126ac608713b42d0a5e3ab04088c60d0e4c5c79d19c9a9dea4015eeefc

data/lib/NeuralNet.rb

@@ -0,0 +1,153 @@
+require "numo/narray"
+require "numo/linalg/use/openblas"
+require "numo/linalg/use/lapack"
+
+include Numo
+class NeuralNetwork
+
+  def initialize(inputNodes:0,hiddenNodes:[],outputNodes:0,
+    learningRate: 0.01,activation: "sigmoid",oneHot: false,
+    reluFactor: 1,momentum: 0.0)
+    @nInputs = inputNodes
+    @nHidden = hiddenNodes
+    @nHiddenLayers = hiddenNodes.length
+    @nOutputs = outputNodes
+    @learningRate = learningRate
+    @momentum = momentum
+    @hiddenWeights = []
+    @hiddenBias = []
+    @prevHWeightDeltas = []
+    @prevHBiasDeltas = []
+
+    tmp1,tmp2 = @nInputs,@nHidden[0]
+    @hiddenWeights[0] = DFloat.new([tmp2,tmp1]).rand * 2 - 1
+    @hiddenBias[0] = DFloat.new([tmp2,1]).rand * 2 - 1
+
+    # Hidden Layers Update Matrix (momentum)
+    @prevHWeightDeltas[0] = DFloat.zeros([tmp2,tmp1])
+    @prevHBiasDeltas[0] = DFloat.zeros([tmp2,1])
+
+    for i in (1...@nHiddenLayers)
+      tmp1,tmp2 = @nHidden[i-1],@nHidden[i]
+      @hiddenWeights[i] =  DFloat.new([tmp2,tmp1]).rand * 2 - 1
+      @hiddenBias[i] =  DFloat.new([tmp2,1]).rand * 2 - 1
+      @prevHWeightDeltas[i] = DFloat.zeros([tmp2,tmp1])
+      @prevHBiasDeltas[i] = DFloat.zeros([tmp2,1])
+    end
+
+    @outputWeights = DFloat.new([@nOutputs,@nHidden[@nHiddenLayers-1]]).rand * 2 - 1
+    @outputBias =  DFloat.new([@nOutputs,1]).rand * 2 - 1
+
+    # Output Layer Update Matrix (momentum)
+    @prevOWeightDeltas =  DFloat.zeros([@nOutputs,@nHidden[@nHiddenLayers-1]])
+    @prevOBiasDeltas = DFloat.zeros([@nOutputs,1])
+
+    @hiddenActivation = activation
+    @hiddenActivationDerv = activation + "_prime"
+    if activation == "relu"
+      @reluFactor = reluFactor.to_f
+    end
+
+    if oneHot
+      @outputActivation = "softmax"
+      @outDel = "softmax_out_delta"
+    else
+      @outputActivation = @hiddenActivation
+      @outputActivationDerv = @hiddenActivationDerv
+      @outDel = "_out_delta"
+    end
+  end
+
+  def train!(data,label)
+    x = DFloat[data].transpose
+    y =  DFloat[label].transpose
+    activations = [x]
+    #feed forward
+    for i in(0...@nHiddenLayers)
+      x1 = method(@hiddenActivation).call(@hiddenWeights[i].dot(activations[-1])+@hiddenBias[i])
+      activations.push(x1)
+    end
+    output = method(@outputActivation).call(@outputWeights.dot(activations[-1])+@outputBias)
+    #backpropagation
+    diff = output - y
+    outdelta = method(@outDel).call(output,diff)
+
+    @prevOBiasDeltas = @momentum * @prevOBiasDeltas + outdelta
+    @prevOWeightDeltas = @momentum * @prevOWeightDeltas + outdelta.dot(activations[-1].transpose)
+
+    @outputBias -= @prevOBiasDeltas
+    @outputWeights -= @prevOWeightDeltas
+
+    delta = @outputWeights.transpose.dot(outdelta)
+    (@nHiddenLayers-1).downto(0) do |i|
+      delta = delta*(method(@hiddenActivationDerv).call(activations[i+1]))
+      @prevHWeightDeltas[i] = @momentum * @prevHWeightDeltas[i] + delta.dot(activations[i].transpose)
+      @prevHBiasDeltas[i] = @momentum * @prevHBiasDeltas[i] + delta
+
+      @hiddenWeights[i] -= @prevHWeightDeltas[i]
+      @hiddenBias[i] -= @prevHBiasDeltas[i]
+      delta = @hiddenWeights[i].transpose.dot(delta)
+    end
+  end
+
+  def predict(data)
+    x = DFloat[data].transpose
+    activations = [x]
+    #feed forward
+    for i in(0...@nHiddenLayers)
+      x1 = method(@hiddenActivation).call(@hiddenWeights[i].dot(activations[-1])+@hiddenBias[i])
+      activations.push(x1)
+    end
+    out = method(@outputActivation).call(@outputWeights.dot(activations[-1])+@outputBias).flatten.to_a
+    out
+  end
+
+  def self.load(path)
+    if File.exist?(path)
+      t  = File.binread(path)
+      return Marshal.load(t)
+    else
+      raise Errno::ENOENT , path
+    end
+  end
+
+  def save(path)
+    File.binwrite(path,Marshal.dump(self))
+  end
+
+  private
+  def relu(x)
+    (x * (x>0))*@reluFactor
+  end
+
+  def relu_prime(y)
+    @reluFactor * (y > 0)
+  end
+
+  def softmax(x)
+    v = NMath.exp(x)
+    return (v*(v.sum**-1))
+  end
+
+  def softmax_prime(y)
+    phi = y.dot DFloat.ones(1,@nOutputs)
+    return phi*( DFloat.eye(@nOutputs) - phi.transpose)
+  end
+
+  def sigmoid(x)
+    return (NMath.exp(-x) + 1)**-1
+  end
+
+  def sigmoid_prime(y)
+    return y*(1-y)
+  end
+
+  # cross-entropy loss for softmax
+  def softmax_out_delta(output,diff)
+    diff * @learningRate
+  end
+  
+  def _out_delta(output,diff)
+    method(@outputActivationDerv).call(output)*diff * @learningRate
+  end
+end

metadata

@@ -0,0 +1,44 @@
+--- !ruby/object:Gem::Specification
+name: NeuralNet
+version: !ruby/object:Gem::Version
+  version: 0.0.0
+platform: ruby
+authors:
+- zaki98
+autorequire: 
+bindir: bin
+cert_chain: []
+date: 2018-03-05 00:00:00.000000000 Z
+dependencies: []
+description: Simple Feed Forward Neural Network
+email: ounissizakaria@gmail.com
+executables: []
+extensions: []
+extra_rdoc_files: []
+files:
+- lib/NeuralNet.rb
+homepage: 
+licenses:
+- MIT
+metadata: {}
+post_install_message: 
+rdoc_options: []
+require_paths:
+- lib
+required_ruby_version: !ruby/object:Gem::Requirement
+  requirements:
+  - - ">="
+    - !ruby/object:Gem::Version
+      version: '0'
+required_rubygems_version: !ruby/object:Gem::Requirement
+  requirements:
+  - - ">="
+    - !ruby/object:Gem::Version
+      version: '0'
+requirements: []
+rubyforge_project: 
+rubygems_version: 2.7.4
+signing_key: 
+specification_version: 4
+summary: NeuralNetwork
+test_files: []