I scoured the internet and I can see that this question has been asked by 
many users, but I'm still unable to understand how to do this. 

I have a simple network (as given in the CNN tutorial 
<http://deeplearning.net/tutorial/lenet.html>), except modified to work on 
100*100 images, and a smaller dataset. 

This is basically 

   1. LeNetConvPoolLayer
   2. LeNetConvPoolLayer
   3. HiddenLayer
   4. LogisticRegression (outputs a 0 or 1). 

*1. (Logistically) How do I save the parameters? Do I save them for each of 
the layers or do I simply save the shared variable "params"*?
params = layer3.params + layer2.params + layer1.params + layer0.params

*2. Once saved, do I have to reconstruct this network in another script 
using these params? If so, how do I use these params? I can't find any line 
which assigns the params anywhere.* 

I would like to have the predicted class and the confidence of the 
prediction, to be able to perform verification for face recognition. 



*3. If I want to predict the class for every image, do I simply change the 
batch size to 1? Where can I actually print this prediction? There is a 
function call*
test_losses = [
                    test_model(i)
                    for i in range(n_test_batches)
                ]

*which is defined before as *

# create a function to compute the mistakes that are made by the model
    test_model = theano.function(
        [index],
        layer3.errors(y),
        givens={
            x: test_set_x[index * batch_size: (index + 1) * batch_size],
            y: test_set_y[index * batch_size: (index + 1) * batch_size]
        }
    )

*I can't understand where the class is actually being predicted/which 
variable it is stored in. *


If anyone could help me out with even one of these questions, I'd be 
*extremely* grateful! 
Please let me know if I could provide any other info. 

 

-- 

--- 
You received this message because you are subscribed to the Google Groups 
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to