CODEWITHSUNDEEP
pythonneural-networkfeed-forward
Usman
Usman

Reputation: 3

Is there any range of random initialisation of weights in neural network in classification probelm?

I am implementing a neural network which will take 225 input neurons and it has to classify random numbers from 1 to 7. So for this i need 225 random weights for first output. Suggest me what to do? I have to feed this to feed forward neural network

Upvotes: -1

Views: 205

Answers (1)

artemis
artemis

Reputation: 7281

Note - I am assuming you are using a basic feed forward neural network with backprop. Please state otherwise if you are not.

You will essentially need two sets of weights, one set of weights for your hidden layers and one set of weights for your output layers.

Here is a basic function that should explain what I mean:

# Initialize a network
def initialize_network(n_inputs, hidden_nodes, n_outputs):
    n_inputs = len(training_data[0]) - 1
    n_outputs = len(set([row[-1] for row in training_data]))

    # Create a blank list to hold the network
    network = []
    # Create your hidden layer
    hidden_layer = [{'weights': [random() for i in range(n_inputs + 1)]} for i in range(hidden_nodes)]
    # Append the hidden layer to your network list
    network.append(hidden_layer)
    # Create the output layer
    output_layer = [{'weights': [random() for i in range(hidden_nodes + 1)]} for i in range(n_outputs)]
    # Append that
    network.append(output_layer)

    # Return the network
    return network

A couple of things to remember:

  • hidden_nodes should be a tunable parameter or specified in your project instructions. The # of hidden nodes is different for everyone
  • Your training data will vary in size, but the function above is agnostic of that

Upvotes: 0

Related Questions