Add a custom layer at the end of a trained Model

[jenniew]

From user group issue, how to add layer to end of existing dllib keras graph model. Using nodes(someNames) or node(someName) can only apply to layers having name. The getInputs() and getOutputs() are private now. Should we make them public to easily manipulate input and output layers?
Related issue: https://groups.google.com/u/1/g/bigdl-user-group/c/G8HvbLpZCZ8

[jason-dai]

Can we add the model and additional layers to Sequential?

Or something like

in = Input(...)
out = model(in)
...

[emartinezs44]

I could add more examples using the functional api in the doc. I usually use stackoverflow, because there are a lot of spark community and it will help for bigdl to reach to more people.

[emartinezs44]

An example would be:

  val input = Variable[Float](Shape(1, 10))
  val dense1 = Dense[Float](2).from(input)
  
  // existing model
  val model = Model(input, dense1)

  val newLayer = Dense[Float](2)
  // Add a new layer to existing model
  val sequential =
    Sequential[Float]()
    .add(model)
    .add(newLayer)

Original model:
------------------------------------------------------------------------------------------------------------------------
Layer (type)                            Output Shape              Param #       Connected to                          
========================================================================================================================
Inputeadea940 (Input)                   (None, 1, 10)             0                                                   
________________________________________________________________________________________________________________________
Densee7b602d8 (Dense)                   (None, 1, 2)              22            Inputeadea940                         
________________________________________________________________________________________________________________________
Total params: 22
Trainable params: 22
Non-trainable params: 0


Model with a new layer:
------------------------------------------------------------------------------------------------------------------------
23-02-02 18:40:50 [main] INFO  InternalDistriOptimizer$:182 - Model Summary:
------------------------------------------------------------------------------------------------------------------------
Layer (type)                            Output Shape              Param #       Connected to                          
========================================================================================================================
Inputb1d1e4a4 (Input)                   (None, 1, 10)             0                                                   
________________________________________________________________________________________________________________________
Modele623acc9 (Model)                   (None, 1, 2)              22            Inputb1d1e4a4                         
________________________________________________________________________________________________________________________
Dense38b91e4d (Dense)                   (None, 1, 2)              6             Modele623acc9                         
________________________________________________________________________________________________________________________
Total params: 28
Trainable params: 28
Non-trainable params: 0
------------------------------------------------------------------------------------------------------------------------

[jason-dai]

I could add more examples using the functional api in the doc. I usually use stackoverflow, because there are a lot of spark community and it will help for bigdl to reach to more people.

stackoverflow sounds like a great idea :-)