Let’s say we have trained a model
m1 = Chain(Dense(10, 10, relu), Dense(10,2), softmax)
and we have then substituted
Dense(10,4) and trained the latter layer only, getting a model
m2 = Chain(Dense(10, 10, relu), Dense(10,4), softmax)
m1 == m2.
Let’s say that we want now to combine them in a model
m3 = Chain(Dense(10, 10, relu), Dense(10,6), softmax)
Dense(10,6) is the union of the output neurons of
m1, with their respective input weights from their shared 10-neurons input layer.
What is the best way to construct