Keras: Custom loss with multiple model outputs

[Jamesswiz]

I have a model with multiple outputs from different layers: O: output from softmax layer; y1,y2: from intermediate hidden layer.

  `m = keras.models.Model (inputs=x, outputs=[O,y1,y2])`

I want to compute cross-entropy loss between O and true labels and MSE loss between y1 and y2. My custom loss function will be like:

  ```
  def Myloss(O_true, O_pred, y1, y2):
        MSE = mean_squared_error(y1, y2)
        CE = categorical_crossentropy(O_true, O_pred)
        return MSE + CE 

  ```

Now how do I pass the model outputs to
model.compile(loss='Myloss', ...)

[jannctu]

model.compile(loss={'o': Myloss, 'y1': Myloss, 'y2': Myloss}, optimizer='adam')
hope it helped

[bairavi26]

I am trying to do something similar. @Jamesswiz when I run the MyLoss function, I get an error saying that 'O' (O_true, O_pred) is a missing argument into the loss function.

How are the outputs of the model related to the loss, and compile functions?

[jannctu]

@bairavi26 please post your code here.
which fit wrapping do you use? fit, fit_generator, or train_on_batch?

[jvishnuvardhan]

Please fill the issue template here. Could you update them if they are relevant in your case, or leave them as N/A? Along with the template, please provide as many details as possible to find the root cause of the issue. It would be great if you can provide a small code to reproduce the error. Thanks!

[ymodak]

Automatically closing due to lack of recent activity. Please update the issue when new information becomes available, and we will reopen the issue. Thanks!

[chandu8542]

Here is what I did to get to multi output thing working. suppoe your model is:
model = Model(inputs=[input1,input2], outputs=[output1,output2])
build loss function 1

def loss_fun1(y_true, y_pred):
     # do whatever you want
    loss  = tf.nn.sigmoid_cross....
    return loss

build loss function 2
````
def loss_fun2(y_true, y_pred):
# do whatever you want
loss = tf.nn.sigmoid_cross....
return loss

define loss dict

losses ={'output_layer1':loss_fun1
'output_layer2:loss_fun2
}

define loss weight

lossWeights={'output_layer1':0.5
'output_layer2:0.5
}

create placeholder for targets (keras throw error if not supplied, be careful)

target1 = tf.placeholder(dtype='int32', shape=()) # shapes of output1 your target has
target2 = tf.placeholder(dtype='float32', shape=()) # shapes of output2 your target has

model.compile(optimizer='rmsprop',
loss=losses,
loss_weights= lossWeights,
target_tensors=[target1, target2]

         )

```

[kennychenfs]

so the output_layer1 should be the layer's object or the name in layer(units=3,name='name')?

[chandu8542]

its 'name' of the output layer.

[QaziAmmar]

But if I want to write a combined loss function for both layers then what should I do ?

`def combined_loss(y_true, y_pred):

multi_pred, binary_pred = y_pred
multi_true, binary_true = y_true
multiloss = tf.losses.categorical_crossentropy(y_true, y_pred)
binary_loss = tf.losses.binary_crossentropy(binary_true, binary_pred)
#
loss = multiloss + binary_loss
return loss

`