Note: the repository is not maintained. Feel free to PM me if you'd like to take up the maintainance.
Build a general-purpose conversational chatbot based on a hot seq2seq approach implemented in tensorflow. Since it doesn't produce good results so far, also consider other implementations of seq2seq.
The current results are pretty lousy:
hello baby - hello
how old are you ? - twenty .
i am lonely - i am not
nice - you ' re not going to be okay .
so rude - i ' m sorry .
Disclaimer:
- the answers are hand-picked (it looks cooler that way)
- chatbot has no power to follow the conversation line so far; in the example above it's a just a coincidence (hand-picked one)
Everyone is welcome to investigate the code and suggest the improvements.
Actual deeds
- realise how to diversify chatbot answers (currently the most probable one is picked and it's dull)
Papers
Nice picture
Curtesy of this article.
Setup
git clone git@github.com:nicolas-ivanov/tf_seq2seq_chatbot.git
cd tf_seq2seq_chatbot
bash setup.sh
Run
Train a seq2seq model on a small (17 MB) corpus of movie subtitles:
python train.py
(this command will run the training on a CPU... GPU instructions are coming)
Test trained trained model on a set of common questions:
python test.py
Chat with trained model in console:
python chat.py
All configuration params are stored at tf_seq2seq_chatbot/configs/config.py
GPU usage
If you are lucky to have a proper gpu configuration for tensorflow already, this should do the job:
python train.py
Otherwise you may need to build tensorflow from source and run the code as follows:
cd tensorflow # cd to the tensorflow source folder
cp -r ~/tf_seq2seq_chatbot ./ # copy project's code to tensorflow root
bazel build -c opt --config=cuda tf_seq2seq_chatbot:train # build with gpu-enable option
./bazel-bin/tf_seq2seq_chatbot/train # run the built code
Requirements