Wrapping a model using Lambda layer corrupts performance

[stigersh]

Summary

In the debugging process I checked the following: I have a keras model which I wrapped in a lambda. From some reason doing it made the optimization start from a worse accuracy point, and get stuck on that value, whereas without the lambda wrapper all is good.

Works bad:

class mymodel(object):
    def __init__():
        def lambda_wrapper(x):
            model = Sequential()
            #some layers like:
            cell = LSTM(100, return_sequences=True)
            layer = Bidirectional(cell)
            model.add(layer)
            return model(x)

       lambda_model_layer = Lambda(lambda x: lambda_wrapper(x))
       self.model = Sequential()
       self.model.add(lambda_model_layer)

Works fine:

class mymodel(object):
    def __init__():
        model = Sequential()
        #some layers like:
        cell = LSTM(100, return_sequences=True)
        layer = Bidirectional(cell)
        model.add(layer)
        self.model = model

Environment

• Python version: 3.6
• Keras version: 2.2.4
• Keras backend with version: tensorflow 1.14.0

Logs or source codes for reproduction

[stigersh]

I think I found the problem. Lambda layers weren't designed to be used with trainable variables. Thus everything inside the lambda becomes untrainable. That's why the code with the lambda is stuck on the same accuracy value. For writing a custom layer with trainable variables this link should be used: writing-your-own-keras-layers