It appears that the initialization of the weights of either the similarity layer or the mex layer in Keras are fixed to default values and cannot be changed. This is a very critical aspect of using networks in practice, and although the unsupervised initialization mitigate this problem (partially) for the similarity layer, it does not for the MEX layer.
Please add support for specifying the initializers of the parameters, according to Keras standard interface (as shown here).
It appears that the initialization of the weights of either the similarity layer or the mex layer in Keras are fixed to default values and cannot be changed. This is a very critical aspect of using networks in practice, and although the unsupervised initialization mitigate this problem (partially) for the similarity layer, it does not for the MEX layer.
Please add support for specifying the initializers of the parameters, according to Keras standard interface (as shown here).