Class: RubyBrain::Network

Inherits:
Object
  • Object
show all
Extended by:
Forwardable
Defined in:
lib/ruby_brain/network.rb

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(num_units_list) ⇒ Network

Constructor of Network class

Examples:

network structure Array(num_units_list)

[10, 30, 3]       # => 3 inputs, hidden layer 1 with 30 units, 3 outputs
[15, 50, 60, 10]  # => 15 inputs, hidden layer 1 with 50 units, hidden layer 2 with 60 units, 10 outputs

Parameters:

  • num_units_list (Array) —

    Array which describes the network structure



15
16
17
18
19
 
# File 'lib/ruby_brain/network.rb', line 15

def initialize(num_units_list)
  @layers = []
  @num_units_list = num_units_list
  @weights_set = WeightContainer.new(@num_units_list)
end

Instance Attribute Details

#learning_rate ⇒ Object

Returns the value of attribute learning_rate.



6
7
8
 
# File 'lib/ruby_brain/network.rb', line 6

def learning_rate
  @learning_rate
end

Instance Method Details

#dump_weights ⇒ Object 



259
260
261
262
263
 
# File 'lib/ruby_brain/network.rb', line 259

def dump_weights
  @weights_set.each_weights do |weights|
    pp weights
  end
end

#dump_weights_to_yaml(file_name = nil) ⇒ Object

Dumps weights of the network into a file whose format is YAML.

Parameters:

  • file_name (String) (defaults to: nil) —

    The path to the YAML file in which weights are saved.



268
269
270
 
# File 'lib/ruby_brain/network.rb', line 268

def dump_weights_to_yaml(file_name=nil)
  @weights_set.dump_to_yaml(file_name)
end

#get_forward_outputs(inputs) ⇒ Object

Calculate the network output of forward propagation.

Parameters:

  • inputs (Array) —

    Input dataset.



66
67
68
69
70
71
72
73
74
75
76
77
78
 
# File 'lib/ruby_brain/network.rb', line 66

def get_forward_outputs(inputs)
  inputs.each_with_index do |input, i|
    @layers.first.nodes[i].value = input
  end

  a_layer_outputs = nil
  a_layer_inputs = @layers.first.forward_outputs
  @layers.each do |layer|
    a_layer_outputs = layer.forward_outputs(a_layer_inputs)
    a_layer_inputs = a_layer_outputs
  end
  a_layer_outputs
end

#init_network ⇒ Object

Initialize the network. This method creates network actually based on the network structure Array which specified with Constructor.



31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
 
# File 'lib/ruby_brain/network.rb', line 31

def init_network
  @layers = []
  layer = Layer.new
  (@num_units_list[0] + 1).times do
    layer.append Nodes::ConstNode.new
    layer.output_weights = @weights_set.weights_of_order(0)
  end
  @layers << layer

  @num_units_list[1..-2].each_with_index do |num_units, i|
    layer = Layer.new
    layer.input_weights = @weights_set.weights_of_order(i)
    layer.output_weights = @weights_set.weights_of_order(i+1)
    (num_units).times do
      layer.append Nodes::Neuron.new
    end
    layer.append Nodes::ConstNode.new
    @layers << layer
  end

  layer = Layer.new
  layer.input_weights = @weights_set.weights_of_order(@num_units_list.size - 2)
  @num_units_list[-1].times do
    layer.append Nodes::Neuron.new
  end
  @layers << layer
end

#learn(inputs_set, outputs_set, max_training_count = 50, tolerance = 0.0, monitoring_channels = []) ⇒ Object

Starts training with training dataset

Parameters:

  • inputs_set (Array<Array>) —

    Input dataset for training. The structure is 2 dimentional Array. Eatch dimentions correspond to samples and features.

  • outputs_set (Array<Array>) —

    Output dataset for training. The structure is 2 dimentional Array. Eatch dimentions correspond to samples and features.

  • max_training_count (Integer) (defaults to: 50) —

    Max training count.

  • tolerance (Float) (defaults to: 0.0) —

    The Threshold to stop training. Training is stopped when RMS error reach to this value even if training count is not max_training_count.

  • monitoring_channels (Array<Symbol>) (defaults to: []) —

    Specify which log should be reported. Now you can select only ‘:best_params_training`

Raises:



137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
 
# File 'lib/ruby_brain/network.rb', line 137

def learn(inputs_set, outputs_set, max_training_count=50, tolerance=0.0, monitoring_channels=[])
  raise RubyBrain::Exception::TrainingDataError if inputs_set.size != outputs_set.size
  #      raise "inputs_set and outputs_set has different size!!!!" if inputs_set.size != outputs_set.size

  best_error = 9999999999999
  best_weights_array = []
  max_training_count.times do |i_training|
    accumulated_errors = 0.0 # for rms
    inputs_set.zip(outputs_set).each do |t_input, t_output|
      forward_outputs = get_forward_outputs(t_input)
      # for rms start
      total_error_of_output_nodes = forward_outputs.zip(t_output).reduce(0.0) do |a, output_pair|
        a + ((output_pair[0] - output_pair[1])**2 / 2.0)
      end
      # end
      accumulated_errors += total_error_of_output_nodes / forward_outputs.size
      # accumulated_errors += forward_outputs.zip(t_output).reduce(0.0) { |a, output_pair| a + ((output_pair[0] - output_pair[1])**2 / 2.0) } / forward_outputs.size
      # for rms end
      backward_inputs = forward_outputs.zip(t_output).map { |o, t| o - t }
      run_backpropagate(backward_inputs)
      update_weights
    end

    rms_error = Math.sqrt(2.0 * accumulated_errors / inputs_set.size) # for rms
    # rms_error = calculate_rms_error(inputs_set, outputs_set)
    puts "--> #{rms_error} (#{i_training}/#{max_training_count})"
    
    if rms_error < best_error
      puts "update best!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"
      best_error = rms_error
      best_weights_array = @weights_set.get_weights_as_array
    end
    puts "best: #{best_error}"


    break if rms_error <= tolerance
  end
  if monitoring_channels.include? :best_params_training
    File.open "best_weights_#{Time.now.to_i}.yml", 'w+' do |f|
      YAML.dump(best_weights_array, f)
    end
  end
end

#learn2(inputs_set, outputs_set, max_training_count = 50, tolerance = 0.0, monitoring_channels = []) ⇒ Object

Raises:



182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
 
# File 'lib/ruby_brain/network.rb', line 182

def learn2(inputs_set, outputs_set, max_training_count=50, tolerance=0.0, monitoring_channels=[])
  # looks like works well for networks which has many layers... [1, 10, 10, 10, 1], [1, 100, 100, 100, 1]
  # looks like NOT works well for networks which has many units in a layer... [1, 100, 1]
  raise RubyBrain::Exception::TrainingDataError if inputs_set.size != outputs_set.size
  # raise "inputs_set and outputs_set has different size!!!!" if inputs_set.size != outputs_set.size
  initial_learning_rate = @learning_rate

  rms_error = Float::INFINITY
  max_training_count.times do |i_training|
    accumulated_errors = 0.0 # for rms
    inputs_set.zip(outputs_set).each do |t_input, t_output|
      forward_outputs = get_forward_outputs(t_input)
      # for rms start
      total_error_of_output_nodes = forward_outputs.zip(t_output).reduce(0.0) do |a, output_pair|
        a + ((output_pair[0] - output_pair[1])**2 / 2.0)
      end
      # end
      error_of_this_training_data = total_error_of_output_nodes / forward_outputs.size
      accumulated_errors += error_of_this_training_data
      # accumulated_errors += forward_outputs.zip(t_output).reduce(0.0) { |a, output_pair| a + ((output_pair[0] - output_pair[1])**2 / 2.0) } / forward_outputs.size
      # for rms end
      # if error_of_this_training_data > rms_error**2/2.0
      #   @learning_rate *= 10.0
      # end
      backward_inputs = forward_outputs.zip(t_output).map { |o, t| o - t }
      run_backpropagate(backward_inputs)
      update_weights
      # @learning_rate = initial_learning_rate
    end

    rms_error = Math.sqrt(2.0 * accumulated_errors / inputs_set.size) # for rms

    # rms_error = calculate_rms_error(inputs_set, outputs_set)
    puts "--> #{rms_error} (#{i_training}/#{max_training_count})"
    break if rms_error <= tolerance
  end
end

#learn_only_specified_layer(layer_index, inputs_set, outputs_set, max_training_count = 50, tolerance = 0.0) ⇒ Object 



221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
 
# File 'lib/ruby_brain/network.rb', line 221

def learn_only_specified_layer(layer_index, inputs_set, outputs_set, max_training_count=50, tolerance=0.0)
  # looks like works well for networks which has many layers... [1, 10, 10, 10, 1], [1, 100, 100, 100, 1]
  # looks like NOT works well for networks which has many units in a layer... [1, 100, 1]
  raise "inputs_set and outputs_set has different size!!!!" if inputs_set.size != outputs_set.size
  initial_learning_rate = @learning_rate

  rms_error = Float::INFINITY
  max_training_count.times do |i_training|
    accumulated_errors = 0.0 # for rms
    inputs_set.zip(outputs_set).each do |t_input, t_output|
      forward_outputs = get_forward_outputs(t_input)
      # for rms start
      total_error_of_output_nodes = forward_outputs.zip(t_output).reduce(0.0) do |a, output_pair|
        a + ((output_pair[0] - output_pair[1])**2 / 2.0)
      end
      # end
      error_of_this_training_data = total_error_of_output_nodes / forward_outputs.size
      accumulated_errors += error_of_this_training_data
      # accumulated_errors += forward_outputs.zip(t_output).reduce(0.0) { |a, output_pair| a + ((output_pair[0] - output_pair[1])**2 / 2.0) } / forward_outputs.size
      # for rms end
      if error_of_this_training_data > rms_error**2/2.0
        @learning_rate *= 10.0
      end
      backward_inputs = forward_outputs.zip(t_output).map { |o, t| o - t }
      run_backpropagate(backward_inputs)
      update_weights_of_layer(layer_index)
      @learning_rate = initial_learning_rate
    end

    rms_error = Math.sqrt(2.0 * accumulated_errors / inputs_set.size) # for rms

    # rms_error = calculate_rms_error(inputs_set, outputs_set)
    puts "--> #{rms_error} (#{i_training}/#{max_training_count})"
    break if rms_error <= tolerance
  end
end

#load_weights_from(weights_set_source) ⇒ Object 



21
22
23
24
 
# File 'lib/ruby_brain/network.rb', line 21

def load_weights_from(weights_set_source)
  @weights_set.load_from(weights_set_source)
  init_network
end

#load_weights_from_yaml_file(yaml_file) ⇒ Object

Loads weights of the network from existing weights file whose format is YAML.

Parameters:

  • yaml_file (String) —

    The path to the YAML file which includes weights.



275
276
277
 
# File 'lib/ruby_brain/network.rb', line 275

def load_weights_from_yaml_file(yaml_file)
  @weights_set.load_from_yaml_file(yaml_file)
end

#run_backpropagate(backward_inputs) ⇒ Object

Calculate the networkoutput of backward propagation.

Parameters:

  • backward_inputs (Array) —

    Input for backpropagation. Usually it is loss values.



83
84
85
86
87
88
89
90
91
 
# File 'lib/ruby_brain/network.rb', line 83

def run_backpropagate(backward_inputs)
  a_layer_outputs = nil
  a_layer_inputs = backward_inputs
  @layers.reverse[0..-2].each do |layer|
    a_layer_outputs = layer.backward_outputs(a_layer_inputs)
    a_layer_inputs = a_layer_outputs
  end
  a_layer_outputs
end

#update_weights ⇒ Object

Updates weights actually based on the result of backward propagation



95
96
97
98
99
100
101
102
103
 
# File 'lib/ruby_brain/network.rb', line 95

def update_weights
  @weights_set.each_weights_with_index do |weights, i|
    weights.each_with_index do |wl, j|
      wl.each_with_index do |w, k|
        wl[k] = w - (@learning_rate * @layers[i].nodes[j].this_output * @layers[i+1].nodes[k].this_backward_output)
      end
    end
  end
end

#update_weights_of_layer(layer_index) ⇒ Object 



105
106
107
108
109
110
111
112
113
114
115
 
# File 'lib/ruby_brain/network.rb', line 105

def update_weights_of_layer(layer_index)
  layer_index = @weights_set.num_sets + layer_index if layer_index < 0
  @weights_set.each_weights_with_index do |weights, i|
    next if i != layer_index
    weights.each_with_index do |wl, j|
      wl.each_with_index do |w, k|
        wl[k] = w - (@learning_rate * @layers[i].nodes[j].this_output * @layers[i+1].nodes[k].this_backward_output)
      end
    end
  end
end