In Lasagne, Layer class has field params which, according to the documentation, stores all registered parameters along with their tags in the ordered dictionary. However, for the layer of the type RecurrentLayer input-to-hidden and hidden-to-hidden weight matrices are absent in params. Why is that so?
Initially, I wanted to keep the values of the hidden-to-hidden matrix fixed, in other words, make hidden_to_hidden.W not trainable. It can be done for parameters in params dict by removing the trainable tag.
