louisjc / mnist-neural-network Goto Github PK

       def backpropagate(self, input_vector, target):
        """Reduce error for one input vector:
        Calculating the partial derivatives for each coeff then subtracts"""
        c = 1./(self.iteration + 10)  # Learning coefficient
        hidden_outputs, outputs = self.feed_forward(input_vector)

        # Calculation of partial derivatives for the output layer and subtraction
        output_deltas = outputs * (1 - outputs) * (outputs - target)

        # Calculation of partial derivatives for the hidden layer and subtraction
        hidden_deltas = hidden_outputs * (1 - hidden_outputs) * \
            np.dot(np.delete(self.layers[-1], 300, 1).T, output_deltas)
        self.layers[-1] -= c*np.outer(output_deltas, np.append(hidden_outputs, 1))
        self.layers[0] -= c*np.outer(hidden_deltas, np.append(input_vector, 1))