How do you make a prediction (forecast) from a trained network for a given input?

Question

Below is my code for a neural network,with 3 inputs and 1 hidden layer and 1 output:

#Data 
ds = SupervisedDataSet(3,1)

myfile = open('my_file.csv','r')

for data in tf.myfile ():
   indata =  tuple(data[:3])
   outdata = tuple(data[3])
   ds.addSample(indata,outdata)

net = FeedForwardNetwork() 
inp = LinearLayer(3) 
h1 = SigmoidLayer(1) 
outp = LinearLayer(1)

# add modules 
net.addOutputModule(outp) 
net.addInputModule(inp) 
net.addModule(h1)

# create connections 
net.addConnection(FullConnection(inp, h1))  
net.addConnection(FullConnection(h1, outp))

# finish up 
net.sortModules()

# initialize the backprop trainer and train 
trainer = BackpropTrainer(net, ds)
trainer.trainOnDataset(ds,1000) trainer.testOnData(verbose=True)

print 'Final weights:',net.params

My question is,if you want to use this trained neural network to make a forecast based on specific inputs,how do you do it?

score 4 · Accepted Answer · answered Jan 17 '12 at 01:09

4

According to the documentation, you can test specific inputs with the activate method on your network. Assuming your input looks something like (1 2 3) your code would look like

net.activate((1,2,3))

answered Jan 17 '12 at 01:09

tjarratt

1,692
9
17

To be honest, I had to look at their documentation for a while before that become obvious. That's really frustrating. – tjarratt Jan 17 '12 at 01:10
Thanks.But a quick question if I execute my specific code twice for the same specific 3 inputs,i would get the same result?Why not? – IordanouGiannis Jan 18 '12 at 01:24
I'm not sure I understand. If you run the same input through several times you should get the same result *most* of the time, but not all of the time. Are you expecting to get the same result 100% of the time or for the result to change more often? Either way, it will depend on how you've trained the network. – tjarratt Jan 18 '12 at 03:15

score 0 · Answer 2 · edited Jun 20 '20 at 09:12

If I understand you correctly, your data has a time order. What I do for making forecast is to shift the data table, in order to present the next output as target for the training. For example, if you have this kind of data:

w1 x1 y1 z1

w2 x2 y2 z2

w3 x3 y3 z3

w4 x4 y4 z4

. . .

and you want to predict z2, you construct a table like:

w1 x1 y1 z1 | z2

w2 x2 y2 z2 | z3

w3 x3 y3 z3 | z4

. . .

Then you present the last column as target for the training. Of course, you lose one line at the end of your table.

You can also improve the output by giving the difference between steps as an additional input (gives you the dynamical effect:

w2 x2 y2 z2 (w2-w1) (z2-z1) | z3

w3 x3 y3 z3 (w3-w2) (z3-z2) | z4

. . .

How do you make a prediction (forecast) from a trained network for a given input?

2 Answers2