Pass the loss value to other models in Keras

user_3878377 0
Viewed 1061

Assume I have two models in series with each other. I am using lstm for both. I want to have a prediction from one of them and find the loss between the predicted value and the ground truth and pass this number to update the other model. Can anyone tell me where to add external loss from outside into Keras model.

For example. Consider the loss between predicted and true value is 0.2. I want to pass this 0.2 to another lstm keras model as an external value to add it with the loss from its output. enter image description here

What python command with keras can do this and where should I add this loss to what module. I thought it should be added to model.compile but the loss there is for objective function. How can I take this loss into account for the next layer that it understands to account for it (in updating the parameters using SGD or any optimizer in training)

lstm keras tensorflow deep-learning python
1 Answers

If I understand you correctly, what you are asking for isn't easily achieved (and perhaps isn't what you need).

Taking a step back, a training step in a keras model is roughly:

predictions = model(train_x)
loss = some_loss_function(true_labels, predictions)
do_backprop_etc_to_update_model_in_order_to_minimize_loss(loss_operations)

keras & tf (cleverly) hides much of the implementation of the 3rd of these steps(which I have grossly oversimplified) but nonetheless it happens and is at the core of the training keras/tf does.

That means that if you have a loss from model1 then it cannot sensibly be passed to model2 without including the operations from model1 inside model2 (which is effectively just like having model3 that does everything - including both outputs. This is, I expect, what you really want?)

If, instead of the architecture you have described, you opt for a single model then you can have 2 distinct outputs each with their own loss function. Furthermore your second loss function can be as complex as you like (so could include a computation of the first loss function).

An architecture like this is perhaps an example:

inputs = Input(shape=(17,))
dense1 = Dense(32, activation='relu')(inputs)
dense2 = Dense(18, activation='relu')(dense1)
dense3 = Dense(7, activation='relu')(dense2)
out1 = Softmax()(dense3)
out2 = Dense(1, activation='sigmoid')(dense2)

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

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_2 (InputLayer)            [(None, 17)]         0                                            
__________________________________________________________________________________________________
dense_3 (Dense)                 (None, 32)           576         input_2[0][0]                    
__________________________________________________________________________________________________
dense_4 (Dense)                 (None, 18)           594         dense_3[0][0]                    
__________________________________________________________________________________________________
dense_5 (Dense)                 (None, 7)            133         dense_4[0][0]                    
__________________________________________________________________________________________________
softmax_1 (Softmax)             (None, 7)            0           dense_5[0][0]                    
__________________________________________________________________________________________________
dense_6 (Dense)                 (None, 1)            19          dense_4[0][0]                    
==================================================================================================
Total params: 1,322
Trainable params: 1,322
Non-trainable params: 0
user_24554940
Related