Open shane-huang opened 5 years ago
The current parameter representation is not very good to represent conditional relationships. e.g. seq2seq and vanilla lstm model each has a set of parameters, we choose one of them and the paramters of the other should not be sampled.
search space pruning also needs consideration here.
The current parameter representation is not very good to represent conditional relationships. e.g. seq2seq and vanilla lstm model each has a set of parameters, we choose one of them and the paramters of the other should not be sampled.