Updating arguments to Lambda Layer after compiling

[nicksam112]

Is there any way to be able to pass in new arguments to a Lambda layer after a model is compiled?

I'm looking to pass in a different external argument every time I run the model, however it looks like the arguments passed in are locked once it's compiled

Any and all help would be appreciated, thank you

[kgrm]

No, Lambda layers are for applying parameter-free backend functions on input tensors.

[akshaychawla]

Deprecated : check my comment below.

You could create a new layer that implements the same lambda functionality. The external argument can be created as a keras variable (K.placeholder) and passed to the layer. So when you have defined your new layer, you will add it to the model somewhat like this:

external_argument = K.placeholder(shape=(1,))
inputs = Input(shape=(784,))
x = Dense(64, activation='relu')(inputs)
x = Dense(64, activation='relu')(x)
x = MyLayer()( [x, external_argument ] ) # keras layers expect a list of inputs to the def call function in the layer definition 
predictions = Dense(10, activation='softmax')(x)

# mention your new external argument as an input when creating your model
model = Model(inputs=[inputs, external_argument], outputs=[predictions])
model.compile(optimizer='sgd', loss='categorical_crossentropy', metrics=['accuracy'])

note that I haven't tested this but it should work. here is a list of sources:

1. https://keras.io/layers/writing-your-own-keras-layers/ : Read this on how to implement your layer. Also note the def call function in the layer, the "x" can be a list of tensors rather than a single tensor. In your case it will be the tensor coming from above the graph AND the external_argument that you have defined as a keras tensor.
2. https://keras.io/getting-started/functional-api-guide/ : since you will use the functional api instead of sequential api.
3. https://keras.io/backend/#using-the-abstract-keras-backend-to-write-new-code : A primer on using the keras backend, especially using the K.placeholder functionality.

[axeper]

Has anyone succeeded in implementing the lambda layer as described by @akshaychawla ?
I am currently facing the same issue as @nicksam112 and am quite lost about creating such a custom layer with placeholders.

Let's say I want a lambda layer that does out = a * in + b but a and b are modified at the end of each epoch using a callback. How would one go about that?

Any help would be really appreciated.

EDIT: I've received some help here.

[akshaychawla]

@axeper I Wrote a small gist which implements the function out = a*x + b in a Lambda layer where "a", "b" and "x" are inputs.
In this gist I generate the inputs (x,a,b) before training, but to implement the functionality that you described you can have a class that contains a keras callback AND a keras data generator. The data being generated from the generator can be changed by the callback which will run after every epoch.