敢えてRubyで学ぶ「ゼロから作るDeep Learning」第3章のニューラルネットワークのMNIST推論処理

PyCallで取り込んだNumPyのデータを無理やりNumo::NArrayに変換してRubyで行列計算をする。

require 'numo/narray'
require 'numo/gnuplot'
require 'datasets'
require 'mini_magick'
require 'pycall/import'
require './neuralnet'
include PyCall::Import

def init_network
  pyimport :numpy
  pyimport :pickle
  pkl = open("sample_weight.pkl", "rb")
  pickle.load(pkl)
end

# 独自適当メソッド
# 無理やりnumpy形式をlist -> Ruby標準Array -> NArrayに変換する荒業
def numpy_to_narray(w)
  row = w.shape[0]
  return Numo::NArray.concatenate(Array(w.tolist).flatten) if w.shape.length == 1
  col = w.shape[1]
  Numo::NArray.concatenate(Array(w.tolist).flatten).reshape(row, col)
end

def predict(network, x)
  w1 = numpy_to_narray(network['W1'])
  w2 = numpy_to_narray(network['W2'])
  w3 = numpy_to_narray(network['W3'])
  b1 = numpy_to_narray(network['b1'])
  b2 = numpy_to_narray(network['b2'])
  b3 = numpy_to_narray(network['b3'])

  a1 = x.dot(w1) + b1
  z1 = sigmoid(a1)
  a2 = z1.dot(w2) + b2
  z2 = sigmoid(a2)
  a3 = z2.dot(w3) + b3
  softmax(a3)
end

# MNISTデータ取り込み
# train = Datasets::MNIST.new(type: :train)
test = Datasets::MNIST.new(type: :test)
x = Numo::NArray.concatenate(test.map{|t| t.pixels }).reshape(10000,784)
t = Numo::NArray.concatenate(test.map{|t| t.label })
network = init_network
# とりあえず一気に変換
y = predict(network, x)
accuracy_cnt = 0
10000.times do |i|
  p = y[i,true].to_a.index(y[i,true].max)
  accuracy_cnt += 1 if p == t[i]
end

p "Accuracy:#{Float(accuracy_cnt/10000.to_f)}"

• 参考文献 https://github.com/oreilly-japan/deep-learning-from-scratch/blob/master/ch03/neuralnet_mnist.py