tongda / windbag

A chatbot based on Deep Learning. Mainly for proof of concept, not for production.

windbag

A chatbot based on Deep Learning. Mainly for proof of concept, not for production.

Train

$ python app/train.py --help
usage: train.py [-h] [--use-attention] [--num-steps NUM_STEPS]
                [--write-summary] [--tag TAG]

Windbag trainer.

optional arguments:
  -h, --help            show this help message and exit
  --use-attention       Flag of whether to use attention.
  --num-steps NUM_STEPS
                        Number of steps.
  --write-summary       Flag of whether to write summaries.
  --tag TAG             Tag of experiment.

A sample run command:

python3 app/train.py --use-attention --tag v3 --num-steps 360000 --batch-size 128 --steps-per-eval 30000 --learning-rate 0.0005

Run

When you have a trained model, run app/chatbot.py can start a chatbot using you model.

Note: you have to change the path in app/chatbot.py to your model path.

Here is a sample output of the chatbot:

> hi
oh , honey ?
> what are you doing ?
watch my to to .
> i love you
i love you really . <pad> <pad> <pad> <pad> <pad> pony <pad> gay loser barney content <pad> <pad> <pad> durden bless imbecile caring
> i feed bad
how ? swann ? ? ? <pad> <pad> <pad> <pad> impressive -
> i am happy for you
you you you happy
> yes , i am happy for you .
why may to to understand
> you look so stupid !
i ruined a sorry .
> are you sorry ?
yes . . .
> are you stupid ?
no .
> are you a computer ?
please .
> happy birthday
sherry . wow <pad> <pad> ward necessarily boatwright
> happy birthday !
yes .
> happy birthday ?
yes .
> happy birthday .
hunh huh
> sure .
why ? dinner <pad>

reference

This project is inspired by Stanford CS20si assignment 3.

The implementation refer to Google Seq2Seq project.