保存经过训练的神经网

Question

I need help with saving my neural network.

I'll explain...i programmed multi-layer network in C# ..the part of application is for training and the other part is for testing neural network. Everything works exactly as it should. When i want to train my network i load a set of data from a file. When the training is over i test it on a smaller sample of data and it gives me correct output. But now i would like to be able to train my network and save it, so that i can load it again and use it for further testing.

Answer 1

I will assume you have your machine learning class called Bayes (or whatever). Typically you would mark this as [Serializable]

using System.IO;
[Serializable]
public class NaiveBayes
{
    ...
}

In this class you could then have a method to do your saving

public void Save(Stream stream)
{
    YourBinaryFormatter b = new YourBinaryFormatter();
    b.Serialize(stream, this);
}

YourBinarySerializer here is just some serializer of your choice, you can use another serializer if you wish. Reading these files is the reverse and is equally as easy.

Answer 2

Yes, you could use Serializable as mentioned. But in robotics, there's often a need for language independent (ie easily parsable) knowledge storing which is why I'm adding this answer.

So, how do you save the current state of any data type?
1) Write the type, its state and a possible descriptor 2) Read it.

For an integer int a = 3 , you could write a file with the following content:

integer
a
3

A neural network is an abstract data type, just like an integer. It is defined by a topology and the final weights after training. Let's say you have a MLP with in=3, hid=6, out=2, then you could write a file with the following contents:

3-6-2 // topology
test1 // name of neural network, could also be in filename (or timestamp)
weight matrix [in->hid]
weight matrix [hid->out]

while you would of course write the actual weights instead of "weight matrix". You can fetch the topology at the initialization part of your program or together with the weights, which is at the end of the training stage.

If you want to reconstruct your network, simply parse the written file and use everything you've read to initialize the network as before but now skip training. You should be able to write files on your robot. If you're unable to do so, send the information over wifi to your local computer and write it there.

Answer 3

You can check my implementation of multi-layer network in C# here .

You need to serialize the network structure, weights and biases. Ann library has a helper methods for that:

Step 1. Define layer configuration:

var layerConfig = new LayerConfiguration()
  .AddInputLayer(2)
  .AddHiddenLayer(5)
  .AddHiddenLayer(5)
  .AddOutputLayer(1);

Step 2. Train Model:

model.TrainModel(new List<double> { 0.25, 0.50 }, new List<double> { 1 });
model.TrainModel(new List<double> { 0.75, 0.15 }, new List<double> { 0 });
model.TrainModel(new List<double> { 0.60, 0.40 }, new List<double> { 1 });
...

Step 3. Save trained Model to JSON file:

model.SaveModelToJson("model.json");

You can instantiate new Network object and use previously trained model:

var model2 = new Network("model.json");
List<double> output = model2.UseModel(new List<double> { 0.35, 0.45 });