Custom loss function using intermediate layer output and output layers

[dfdf]

I need to create a loss function that uses values from a intermediate layer based on DeepID2.

Basically I have this network:

net = NeuralNet(
    layers=[
        ('input', layers.InputLayer),
        ('conv1', Conv2DLayer),
        ('pool1', MaxPool2DLayer),
        ('dropout1', layers.DropoutLayer),
        ('conv2', Conv2DLayer),
        ('pool2', MaxPool2DLayer),
        ('dropout2', layers.DropoutLayer),
        ('conv3', Conv2DLayer),
        ('pool3', MaxPool2DLayer),
        ('dropout3', layers.DropoutLayer),
        ('conv4', Conv2DLayer),
        ('dropout4', layers.DropoutLayer),
        ('flatten1', layers.FlattenLayer),
        ('flatten2', layers.FlattenLayer),
        ('concat',layers.ConcatLayer),
        ('hidden4', layers.DenseLayer),
        ('output', layers.DenseLayer),
        ],

And I need to change the loss function to something like:

1 - Input two images file for each mini-batch

2 - If the images are from the same person (same ID) the loss function will be:

Loss_function:

Error from Image 1 (categorical_crossentropy[Image 1, ID1])
plus
Error from Image 2 ( categorical_crossentropy[Image 2, ID1])
plus
Error between the values of hidden4_layer for the Image 1 and Image 2
(Ex: get_output from hidden4 for image 1 and image 2 and apply squared_error )

3 - If the images are from different person:

Loss_function:

Error from Image 1 (categorical_crossentropy[Image 1, ID1])
plus
Error from Image 2 ( categorical_crossentropy[Image 2, ID2])

Is possible to do this in nolearn?

[BenjaminBossan]

It is definitely possible, though not out of the box. Have a look at the objective function in nolearn.lasagne.base. It is responsible for determining the loss. There you may specify a loss that depends on whatever layer(s) you wish. Then pass the new objective to nolearn's NeuralNet at initialization.

[charbelc]

dfdf have you solved the problem?
If yes i will appreciate if you share what have you learned. I have the same problem...
Thanks in advance