# A reproduction of the wonderful step by step backprop example at

# https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/

require "bundler/setup"

require "rann"

# inputs

inputs = Array.new(2){ |i| RANN::Neuron.new "input #{i}", 0, :input, :sig }

# hidden layer

hiddens = Array.new(2){ |i| RANN::Neuron.new "hidden #{i}", 3, :standard, :sig }

hidden_bias = RANN::Neuron.new "bias", 0, :bias

# output layer

outputs = Array.new(2){ |i| RANN::Neuron.new "output #{i}", 3, :output, :sig }

output_bias = RANN::Neuron.new "bias", 0, :bias

# connect it all w/initial weights

connections = []

connections << RANN::Connection.new(inputs[0], hiddens[0], 0.15.to_d)

connections << RANN::Connection.new(inputs[0], hiddens[1], 0.25.to_d)

connections << RANN::Connection.new(inputs[1], hiddens[0], 0.2.to_d)

connections << RANN::Connection.new(inputs[1], hiddens[1], 0.3.to_d)

connections << RANN::Connection.new(hidden_bias, hiddens[0], 0.35.to_d)

connections << RANN::Connection.new(hidden_bias, hiddens[1], 0.35.to_d)

connections << RANN::Connection.new(hiddens[0], outputs[0], 0.4.to_d)

connections << RANN::Connection.new(hiddens[0], outputs[1], 0.5.to_d)

connections << RANN::Connection.new(hiddens[1], outputs[0], 0.45.to_d)

connections << RANN::Connection.new(hiddens[1], outputs[1], 0.55.to_d)

connections << RANN::Connection.new(output_bias, outputs[0], 0.6.to_d)

connections << RANN::Connection.new(output_bias, outputs[1], 0.6.to_d)

network = RANN::Network.new connections

backprop = RANN::Backprop.new network

inputs = [0.05.to_d, 0.10.to_d]

targets = [0.1.to_d, 0.99.to_d]

outputs = network.evaluate [0.05.to_d, 0.10.to_d]

puts "forward prop outputs: #{outputs.map(&:to_f).inspect}"

network.reset!

gradients, error =

RANN::Backprop.run_single(

network,

[0.05.to_d, 0.10.to_d],

[0.01.to_d, 0.99.to_d]

)

puts "error: #{error.to_f}"

puts "backprop gradients & updates:"

gradients.each do |cid, g|

c = network.connections.find{ |c| c.id == cid }

puts "#{c.input_neuron.name} -> #{c.output_neuron.name}: g = #{g.to_f}, u = #{(c.weight - 0.5.to_d * g).to_f}"

end