…ndition is not satisfied, all models will fail to run.

…ate flag to set wether or not the L2 pooling is used on the utterance encoders output.

…utterance encoder(s) and utterance decoder. Also factorized UtteranceDecoder to make it easier to add future inputs to the it. In particular, I renamed \'hs\' to \'inp\' in UtteranceDecoder.

…antic similarities between targets and samples using SEMILAR: A Semantic Similarity Toolkit. The setup should be directly transferable to clusters for parallel fast evaluation (less than a few hours per model).

…coders and direct-connection-between-encoder-and-decoder.

…merging Alessandro's corrections and refactoring of the beam and stochastic search.

…ed 'index' in data_iterator, so that we can now retrieve the index of each triple. This will be used to pair it with semantic information.

Added DCGM + modified search to remove prev_h

…ialogue module to bootstrap from these.

… last commit.