Branched output Keras

问题

My model is such that it branches into 2 output layers as follows:

Input -> L1 -> L2 -> L3 -> out1

Input -> L1 -> L2 -> L3 -> out2

I am using it this way because I want out1 and out2 to have 2 different activation functions. Therefore, I have created a model:

model = Model(inputs=[input_layer], outputs=[out1, out2])

I am compiling it using:

model.compile(Adam(lr=1e-2), loss=[loss_fn1, loss_fn2], metrics=[accuracy])

loss functions are defined this way:

def loss_fn1(y_true, y_pred):
    #send channel 1 to get bce dice loss
    loss1 = binary_crossentropy(y_true[:, :, :, 0:1], y_pred[:, :, :, 0:1])   
    return loss1

def loss_fn2(y_true, y_pred):    
    #l2 loss for channels 2 and 3
    loss2 = mean_squared_error(y_true[:, :, :, 1:3], y_pred[:, :, :, 1:3])
    return loss2

Does this use loss_fn1 on out1 and loss_fn2 on out2 tensor? Because, that is what I intend to do, but I am unsure regarding whether my code actually does that. Any pointers would help.

I want to use loss_fn1 on out1 tensor and loss_fn2 function on out2 tensor.

EDIT:

loss value from loss_fn1 range: [0, 1] - sigmoid activation.

loss value from loss_fn2 range: [0, inf] - no activation

Is there a way to reduce loss_fn1 and loss_fn2 independently, without using separate models? I am afraid that loss1 + loss2 would eventually only cause a decrease in value of loss2 as loss1 has a low value in comparison to loss2.

回答1:

Yes, your interpretation is right. From the Keras documentation:

If the model has multiple outputs, you can use a different loss on each output by passing a dictionary or a list of losses. The loss value that will be minimized by the model will then be the sum of all individual losses.

来源：https://stackoverflow.com/questions/47774552/branched-output-keras