Multiple network outputs

Is there any way to handle multiple output layers? I’d like to train a network where the first classification result is feed back into the network, e.g.:

example_layer = lasagne.layers.InputLayer(inputs_layer, **params)

...

classification_output = lasagne.layers.DenseLayer(example_layer, nonlinearity=lasagne.nonlinearities.softmax, **params2)

...

another_output_layer = lasagne.layers.DenseLayer(classification_output, nonlinearity=lasagne.nonlinearities.softmax, **params3)

return classification_output, another_output_layer

[DanChianucci]

I have a similar issue, which doesn’t work even if I write my own objective. I have a network which splits into two branches. One branch outputs an image patch (NxCHxHxW), and the other outputs a class (NxCLS) , similar to the network shown below.

   A
   B
 /  \
C    E
D    F

Since the NeuralNet takes a single Layer instance for its layers parameter, one of the branches is completely ignored. For example, if I instantiate the NeuralNet with F as the layers param, then C and D are not placed into the layers_ dictionary. Consequently, I can’t reference C or D from my objective function.

I have fixed this by slightly modifying the initialize_layers method, so that it loops over all Layer instances contained within self.layers, rather than only self.layers[0]. I can submit a pull request if this is of interest.

[dnouri]

Feature implemented in #278 and merged. Continuing in #284: Network.predict works with multi-outputs.