Skip to content
[unmaintained] Make seq2seq for keras work
Branch: master
Clone or download
Latest commit b22cde3 Dec 20, 2016
Type Name Latest commit message Commit time
Failed to load latest commit information.
bin trade ssh for https Jan 20, 2016
configs vocab_max_size fix; batch_size reduced May 21, 2016
data initial commit Dec 3, 2015
lib unnecessary model_save() removed Jul 21, 2016
utils initial commit Dec 3, 2015
.gitignore initial commit Dec 3, 2015 update readme Dec 19, 2016
requirements.txt minor fixes Feb 9, 2016 bump Seq2seq version Jan 19, 2016

debug seq2seq

Note: the repository is not maintained. Feel free to PM me if you'd like to take up the maintainance.

Make seq2seq for keras work. And also give a try to some other implementations of seq2seq.

The code includes:

  • small dataset of movie scripts to train your models on
  • preprocessor function to properly tokenize the data
  • word2vec helpers to make use of gensim word2vec lib for extra flexibility
  • train and predict function to harness the power of seq2seq


  • The code has bugs, undoubtedly. Feel free to fix them and pull-request.
  • No good results were achieved with this architecture yet. See 'Results' section below for details.


Nice picture



git clone
cd debug_seq2seq
bash bin/
python bin/

and then

python bin/


No good results were achieved so far:

[why ?] -> [i ' . . $$$ . $$$ $$$ $$$ $$$ as as as as i i]
[who ?] -> [i ' . . $$$ . $$$ $$$ $$$ $$$ as as as as i i]
[yeah ?] -> [i ' . . $$$ . $$$ $$$ $$$ $$$ as as as as i i]
[what is it ?] -> [i ' . . $$$ . $$$ $$$ $$$ $$$ as as as as as i]
[why not ?] -> [i ' . . $$$ . $$$ $$$ $$$ $$$ as as as as i i]
[really ?] -> [i ' . . $$$ . $$$ $$$ $$$ $$$ as as as as i i]

My guess is that there are some foundational problems in this approach:

  • Since word2vec vectors are used for words representations and the model returns an approximate vector for every next word, this error is accumulated from one word to another and thus starting from the third word the model fails to predict anything meaningful... This problem might be overcome if we replace our approximate word2vec vector every thimestamp with a "correct" vector, i.e. the one that corresponds to an actual word from the dictionary. Does it make sence? However you need to dig into seq2seq code to do that.

  • The second problem relates to word sampling: even if you manage to solve the aforementioned issue, in case you stick to using argmax() for picking the most probable word every time stamps, the answers gonna be too simple and not interesting, like:

are you a human?			-- no .
are you a robot or human?	-- no .
are you a robot?			-- no .
are you better than siri?  		-- yes .
are you here ?				-- yes .
are you human?			-- no .
are you really better than siri?	-- yes .
are you there 				-- you ' re not going to be
are you there?!?!			-- yes .

Not to mislead you: these results were achieved on a different seq2seq architecture, based on tensorflow.

Sampling with temperature could be used in order to diversify the output results, however that's again should be done inside seq2seq library.

You can’t perform that action at this time.