Keras multiple outputs

[shaha...@gmail.com]

Hi,

I want to build a net for solving a classification problem, But I want to classify K outputs using a softmax classifier, each output can be one of M groups. How can I do that with Keras? 

clarification: 

If for example I have N data entries, I mean I want the classifier to give me K outputs, each output will be a vector of length M with 1 only in the group that the entry belongs to. 

This means my target set size will be NxKxM

[Florian Golemo]

Hello,

I'd do that with the functional API. You create your network like any other network and then you just create several output layers, like so:

from keras.layers import Input, Dense
from keras.models import Model

inputs = Input(shape=(N,)) # N is the width of any input element, say you have 50000 data points, and each one is a vector of 3 elements, then N is 3

x = Dense(64, activation='relu')(inputs) # this is your network, let's say you have 2 hidden layers of 64 nodes each (don't know if that's enough for you)
x = Dense(64, activation='relu')(x)

output1 = Dense(M, activation='softmax')(x) # now you create an output layer for each of your K groups. And each output has M elements, out of which because of 'softmax' only 1 will be activated. (practically this is of course a distribution, but after sufficient training, this usually makes one element close to one and the other elements close to zero)
output2 = Dense(M, activation='softmax')(x)
output3 = Dense(M, activation='softmax')(x)
... #you have to fill in the remaining layers here, or better: use a for loop
outputK = Dense(M, activation='softmax')(x)

model = Model(input=inputs, output=[output1, output2, output3, ..., outputK])


model.compile(optimizer='rmsprop',
              loss='categorical_crossentropy',
              metrics=['accuracy'])

model.fit(inputData, [outputData1, outputData2, outputData3, ... outputDataK], nb_epochs=10, batch_size=64)

[Florian Golemo]

Just to be clear, I assume here, that your dataset actually looks like this:

IxN inputs and KxM outputs,

where
- I is the number of rows in your dataset and in each row:
  - N is the length of the input vector
  - K is the number of outputs
  - M is the length of the each output vector

[shaha...@gmail.com]

Thank you very much for your quick and very helpful answer! 

This is exactly what I wanted and it works excellently.

The only problem I have is that now my metrics are the accuracy for each output separately. I want to use a callback to train and take the net after the epoch with the best validation accuracy and I can't do that, because now the validation metrics are calculated separately for each output. Is there a way to make the fit function calculate the average accuracy over all outputs? 

Thanks you very much!

[viji.ven...@gmail.com]

Hello, I have exactly the same question. Were you able to figure out how to calculate the average accuracy over all outputs and use that in your callbacks?

[monikas...@gmail.com]

Hi...what is the solution for getting a combined accuracy for multiple heads?

[Sergio Vera Martínez]

Hello everyone, 

I have the same problem. I get the total loss, as well as from each output, but not the total/mean/combined accuracy.

I would really appreciate it if someone finds an answer, and comments on this thread.

Many thanks in advance.

Best regards,

Sergi

El dia dilluns, 28 de gener de 2019 a les 10:44:55 UTC+1, monikas...@gmail.com va escriure:

[Sergio Vera Martínez]

Hello everyone,

Fortunately, I've found the solution. I've used a callback method to calculate the mean accuracy. Next, there are the links to the solution:

https://github.com/keras-team/keras/issues/10287

https://stackoverflow.com/questions/50541409/keras-accuracy-for-multi-output-model-is-not-working

Best regards

El dia dilluns, 3 de maig de 2021 a les 20:07:19 UTC+2, Sergio Vera Martínez va escriure: