Skip to content

Latest commit

 

History

History
61 lines (30 loc) · 1.83 KB

README.md

File metadata and controls

61 lines (30 loc) · 1.83 KB
Raw

Neural-Conversational-Model

Tensorflow Implementation of Neural Conversational Model by Vinyals et.al. - http://arxiv.org/pdf/1506.05869v3.pdf

Includes optimizations like TFRecords, dynamic_rnn implementation and pretrained word-embeddings to speed up training. Trained on the excellently formated Cornell Movie Dialogure corpus available at - http://www.cs.cornell.edu/~cristian/Cornell_Movie-Dialogs_Corpus.html

####PRE-REQUISITES:

  1. Tensorflow

  2. CUDNN, CUDA

  3. Python's NLTK package

  4. Stanford's Pretrained GloVe Vectors available here. Code currently only supports 300-dim vectors

  5. Cornell's Movie Dialog dataset

####HOW TO RECREATE : Requires a Tensorflow supported GPU. CPU version not supported. With all pre-requisites installed

  1. In loadWordVecs.py set paths to downloaded GloVe file and output path in function generateWordDict,

  2. CUDA_VISIBLE_DEVICES=1 python chatbot.py

####RESUME TRAINING WITH SAVED CHECKPOINT FILE :

CUDA_VISIBLE_DEVICES=1 python chaatbot.py --run-mode 0 --load-chkpt <path to valid checkpoint file>

####TO EVALUATE (Requires a checkpoint file):

NOTE : Currently out of vocabulary words are not supported. So for some inputs (some proper nouns, uncommon words) it will throw an error. Will add support soon.

CUDA_VISIBLE_DEVICES=1 python chaatbot.py --run-mode 1 --load-chkpt <path to valid checkpoint file>

####EXAMPLE CONVERSATIONS AFTER ROUGHLY 8 EPOCHS :

Convo 1:

User : Hi!

Chatty: Hi, Charlie

Convo 2:

User : How are you?

Chatty : I'm not sure.

Convo 3:

User : Are you alive?

Chatty : I'm not going to hurt you.

User : ... (quietly backs out of chat)

NOTE : The example conversations were slightly edited to correct misspellings (ex: "I' m") in the training corpus hence the learnt vocabulary. But otherwise unaltered in any way.