8
\$\begingroup\$

I'm looking for some general tips on code practices, doing things properly in idiomatic pythonic way. But mostly I want to know if this was build correctly. I tried making neural nets in the past, but failed horrible, but this one seems to be working fine.

The backpropagation method is still not developed because I still haven't figured out the math behind it (if anyone have a good material for a non-engineer I would be very glad).

#! /usr/bin/env python
"""
    This module is a framework for a Artificial Neural Network.

    :param NeuralNetwork: See this documentation for how to use this module..
    :type NeuralNetwork: NeuralNetwork

    Author: Fernando Rodrigues dos Santos
"""
import random
import math


class NeuralNetwork:
    """USAGE:
         - Create a NeuralNetwork object and set it's parameters at your will.
         - Use the method start_net(inputs) with the right amount of inputs to calculate the first iteration of the net.
         - Use the get_result() method to extract the result of each output node.
         - Use the get_weights() method to extract the weight of each node.
         - Use the set_weights(weights) method to set the new weight for each node in the net.

         # NOT IMPLEMENTED

         - Use the back_propagate_error(error) method to propagate back the result with the quantified error
    """
    def __init__(self, n_input, n_hidden_layer, n_hidden_nodes, n_output):
        """Creates each layer of the net (input, hidden and output) based on the set parameters
        :param n_input: Number of inputs nodes of the net
        :type n_input: int
        :param n_hidden_layer: Number of hidden layers
        :type n_hidden_layer: int
        :param n_hidden_nodes: Number of nodes per hidden layer
        :type n_hidden_nodes: int
        :param n_output: Number of output nodes of the net
        :type n_output: int
        """
        self.input_layer = Layer(n_input, "Input")

        self.hidden_layers = []
        for n in xrange(n_hidden_layer):
            self.hidden_layers.append(
                Layer(n_hidden_nodes, "Hidden %i" % (n+1))
            )

        self.output_layer = Layer(n_output, "Output")

        self.connect_layers()

    def connect_layers(self):
        """Connects each layer to it previous and next layer."""

        # Input layers are only connected to the first hidden layer
        for node in self.input_layer.nodes:
            node.set_next_layer(self.hidden_layers[0])

        # hidden layers are connected in both directions
        for i, hidden_layer in enumerate(self.hidden_layers):
            for node in hidden_layer.nodes:
                # if its the first hidden layer set its previous layer as the input layer
                # else set its previous layer as the last hidden layer
                if i == 0:
                    previous_layer = self.input_layer
                else:
                    previous_layer = self.hidden_layers[i-1]

                # if its the last hidden layer, set its next layer as the output layer
                # else set  its next layer as the next hidden layer
                if i == len(self.hidden_layers) - 1:
                    next_layer = self.output_layer
                else:
                    next_layer = self.hidden_layers[i+1]

                node.set_next_layer(next_layer)
                node.set_previous_layer(previous_layer)

        # Output layers are only connected to the last hidden layer
        for node in self.output_layer.nodes:
            node.set_previous_layer(self.hidden_layers[-1])

    def start_net(self, input_values):
        """Pass the initial input values to the neural net and let it compute the result.
        :param input_values: Input values in a list form
        :type input_values: list
        """

        # Set input values to input layer nodes
        for i, node in enumerate(self.input_layer.nodes):
            node.set_value(input_values[i])

        # Feed-forward input values (weighted) to first hidden layer nodes
        self.input_layer.feed_forward()

        # Feed-forward result (weighted again) each subsequent hidden layer nodes
        for layer in self.hidden_layers:
            layer.feed_forward()

        # Finally feed-forward result (once again weighted) to the output layer nodes
        self.output_layer.feed_forward()

    def get_result(self):
        """Return the value of output nodes of the neural net after computing the input
        :rtype : list
        """
        return [node.value for node in self.output_layer.nodes]

    def get_weights(self):
        """Return weights of each node of the neural net. (to use in a GA)"""
        hidden_layers = []

        for layer in self.hidden_layers:
            hidden_weights = []
            for node in layer.nodes:
                hidden_weights.append(node.weights)
            hidden_layers.append(hidden_weights)
        out_weights = []

        for node in self.output_layer.nodes:
            out_weights.append(node.weights)

        return hidden_layers, out_weights

    def set_weights(self, weights):
        output_weights = weights[1]
        for i, node in enumerate(self.output_layer.nodes):
            node.set_weights(output_weights[i])

        for i, layer in enumerate(self.hidden_layers):
            hidden_weights = weights[0][i]
            for j, node in enumerate(layer.nodes):
                node.set_weights(hidden_weights[j])

    def back_propagate_error(self, error):  # TODO: to be implemented
        pass


class Layer:
    def __init__(self, n_nodes, layer_name):
        self.nodes = []

        for n in xrange(n_nodes):
            self.nodes.append(
                Node(layer_name + " | Node: %i" % (n+1))
            )

    def feed_forward(self):
        for node in self.nodes:
            node.feed_forward()


class Node:
    def __init__(self, layer_name):
        self.name = layer_name
        self.value = None
        self.previous_layer = None
        self.next_layer = None
        self.weights = []
        self.iteration = 0

    def set_previous_layer(self, layer):
        self.previous_layer = layer

    def set_next_layer(self, layer):
        self.next_layer = layer

    def set_value(self, value):
        self.value = value

    def get_values(self):
        value = []
        for node in self.previous_layer.nodes:
            value.append(node.value)
        return value

    def set_weights(self, weights):
        self.weights = weights

    def get_weights(self):
        weights = []
        if self.iteration == 0:
            # randomize the weights at start
            for _ in xrange(len(self.previous_layer.nodes)):
                weight = random.uniform(-1, 1)
                weights.append(weight)
        else:
            weights = self.weights

        self.iteration += 1

        return weights

    def feed_forward(self):
        """
        1) Get the values of the previous layers
        2) Multiple them by the weights of the node
        3) Sum it all together
        4) Pass the result to the activation function
        5) Set the node value to the returned value of activation function"""

        if self.previous_layer:
            values = self.get_values()
            self.weights = self.get_weights()
            # print self.weights, self.name

            weighted_sum = sum(
                v * w for v, w in zip(values, self.weights)
            )
            self.value = self.activation(weighted_sum)

    @staticmethod
    def activation(x):
        """Sigmoid function"""
        try:
            return 1 / (1 + math.e ** -x)
        except OverflowError:
            return 0

ANN = NeuralNetwork(n_input=1, n_hidden_layer=2, n_hidden_nodes=2, n_output=2)
inputs = [1, 2]
ANN.start_net(inputs)
output = ANN.get_result()
net_weights = ANN.get_weights()
ANN.set_weights(net_weights)
\$\endgroup\$
3
  • 1
    \$\begingroup\$ Back propagation is kind of inherently horrendously complicated. I've never found a resource that explained it well for people whose multivariable calculus is a bit rusty. The third edition of Artificial Intelligence: A Modern Approach has a pretty good discussion of it, and there's also a book by Tom Mitchell that explains it pretty clearly, but I've never found a decent online resource. \$\endgroup\$
    – tsleyson
    Commented May 5, 2015 at 4:10
  • \$\begingroup\$ In my experience, neural networks become much easier to implement when you start thinking of them in terms of linear algebra and matrices. \$\endgroup\$
    – tsleyson
    Commented May 5, 2015 at 4:12
  • \$\begingroup\$ See this question and its answer. \$\endgroup\$ Commented May 5, 2015 at 10:35

1 Answer 1

2
\$\begingroup\$

First off, the below section of code here:

self.hidden_layers = []
for n in xrange(n_hidden_layer):
    self.hidden_layers.append(
        Layer(n_hidden_nodes, "Hidden %i" % (n+1))
    )

Can be shortened to the below, using a generator expression:

self.hidden_layers = [
    Layer(n_hidden_nodes, "Hidden {0}".format(n + 1)) for n in xrange(n_hidden_layer)
]

You also have other places where generator expressions could be used, like here, as a simple example:

for node in self.output_layer.nodes:
    out_weights.append(node.weights)

Do note the use of str.format as well. Using % for string formatting in any version of Python after 2.6 is deprecated, and str.format should be used instead. Here's an example of how str.format is used:

# str.format without positional or named parameters
print "{} {}".format("Hello", "world")

# str.format with positional parameters
print "{1} {0}".format("world", "Hello")

# str.format with named parameters
print "{word1} {word2}".format(word1="Hello", word2="world")

Finally, in any version of Python 2.x, you need to explicitly have all classes inherit from object, like this: class MyClass(object):. If you're using Python 3.x or above, it's okay to just type stuff like clas MyClass:.

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.