cv/nn/nn.py

import numpy as np

class NeuralNetwork:
    def __init__(self, input_layer_size=3, hidden_layer_size=4, output_layer_size=2):
        self.input_layer_size = input_layer_size
        self.hidden_layer_size = hidden_layer_size
        self.output_layer_size = output_layer_size

        self.weights_input_to_hidden = np.random.uniform(-0.5, 0.5, (self.hidden_layer_size, self.input_layer_size))
        self.weights_hidden_to_output = np.random.uniform(-0.5, 0.5, (self.output_layer_size, self.hidden_layer_size))
        self.bias_input_to_hidden = np.zeros((self.hidden_layer_size, 1))
        self.bias_hidden_to_output = np.zeros((self.output_layer_size, 1))

        self.epochs = 3000 # Default
        self.learning_rate = 0.1 # Default

        return
    
    def feedforward(self, data):
        # Forward propagation (to hidden layer)
        hidden_raw = self.bias_input_to_hidden + self.weights_input_to_hidden @ data
        self.hidden = 1 / (1 + np.exp(-hidden_raw)) # sigmoid

        # Forward propagation (to output layer)
        output_raw = self.bias_hidden_to_output + self.weights_hidden_to_output @ self.hidden
        output = 1 / (1 + np.exp(-output_raw))
        return output
    
    def backprop(self, data, output, result):
        # Backpropagation (output layer)
        delta_output = output - result
        self.weights_hidden_to_output += -self.learning_rate * delta_output @ np.transpose(self.hidden)
        self.bias_hidden_to_output += -self.learning_rate * delta_output

        # Backpropagation (hidden layer)
        delta_hidden = np.transpose(self.weights_hidden_to_output) @ delta_output * (self.hidden * (1 - self.hidden))
        self.weights_input_to_hidden += -self.learning_rate * delta_hidden @ np.transpose(data)
        self.bias_input_to_hidden += -self.learning_rate * delta_hidden
        return
    
    def get(self, data):
        data = np.reshape(data, (-1, 1))
        return self.feedforward(data)

    def learning(self, dataset, results, epochs, learning_rate):
        self.epochs = epochs
        self.learning_rate = learning_rate
        e_loss = 0
        e_correct = 0

        # Learning
        for epoch in range(epochs):
            print(f"Epoch {epoch}")

            for data, result in zip(dataset, results):
                data = np.reshape(data, (-1, 1))
                result = np.reshape(result, (-1, 1))
                
                output = self.feedforward(data)

                # Loss / Error calculation
                e_loss += 1 / len(output) * np.sum((output - result) ** 2, axis=0)
                e_correct += int(np.argmax(output) == np.argmax(result))

                self.backprop(data, output, result)

            # print some debug info between epochs
            print(f"Loss: {round((e_loss[0] / len(dataset)) * 100, 3)}%")
            print(f"Accuracy: {round((e_correct / len(dataset)) * 100, 3)}%")
            e_loss = 0
            e_correct = 0

        return
    
    def save(self, filename):
        np.savez(filename,
                 weights_input_to_hidden=self.weights_input_to_hidden,
                 weights_hidden_to_output=self.weights_hidden_to_output,
                 bias_input_to_hidden=self.bias_input_to_hidden,
                 bias_hidden_to_output=self.bias_hidden_to_output
                 )
        return

    def load(self, filename):
        with np.load(filename) as f:
            self.weights_input_to_hidden = f['weights_input_to_hidden'],
            self.weights_hidden_to_output = f['weights_hidden_to_output'],
            self.bias_input_to_hidden = f['bias_input_to_hidden'],
            self.bias_hidden_to_output = f['bias_hidden_to_output']
        return