1  from typing import List
 2  
 3  import mnist
 4  from matrix import Matrix
 5  from network import Network
 6  from trainingdata import TrainingData
 7  from util import sigmoid, sigmoid_prime
 8  
 9  
10  class DigitClassifier(Network):
11      def __init__(self, weights: List[Matrix], biases: List[Matrix]):
12          super().__init__(weights, biases)
13  
14      @classmethod
15      def random_network(cls, layer_sizes: List[int]):
16          return super().random_network([28 ** 2] + layer_sizes + [10])
17  
18      @staticmethod
19      def activation(x: float) -> float:
20          return sigmoid(x)
21  
22      @staticmethod
23      def activation_derivative(x: float) -> float:
24          return sigmoid_prime(x)
25  
26      def prepare_expected_output(self, expected) -> Matrix:
27          return Matrix.from_vector([float(i == expected) for i in range(10)])
28  
29      def prepare_input(self, inp) -> Matrix:
30          return Matrix.from_vector(inp)
31  
32      def interpret_output(self, out: Matrix):
33          out = out.to_vector()
34          return out.index(max(out))
35  
36      def evaluate(self, test_data: List[TrainingData]) -> float:
37          result = 0
38          for td in test_data:
39              if self.feedforward(td.input_data) == td.expected_output:
40                  result += 1
41          return result / len(test_data)
42  
43  
44  if __name__ == '__main__':
45      network = DigitClassifier.random_network([10, 10])
46      print("Loading training data ...")
47      training_data = [x for _, x in zip(range(2000), mnist.load_train())]
48      print("Loading validation data ...")
49      validation_data = [x for _, x in zip(range(1000), mnist.load_test())]
50      print("Start training")
51      network.train(training_data, 30, 10, 3, validation_data, "network.json")
52      print(f"Accuracy: {network.evaluate(validation_data)}")