A Neural Conversational Model 
TensorFlow implementation of Conversation Models.
-
Model
seq2seq_attention: Seq2Seq model with attentional decoder- Encoder
- Unidirectional RNN
- Stack Bidirectional RNN
- Attention
- Bahdanau Attention (option Norm)
- Luong Attention (option Scale)
- Decoder
- Greedy (beam_width = 0)
- Beam Search (beam_width > 0)
- Scheduled Sampling
- Encoder
-
Dataset
Requirements
- Python 3.6
- TensorFlow 1.4
- nltk
- hb-config (Singleton Config)
- tqdm
- Slack Incoming Webhook URL
Project Structure
initiate Project by hb-base
.
βββ config # Config files (.yml, .json) using with hb-config
βββ data/ # dataset path
βββ scripts # download dataset using shell scripts
βββ seq2seq_attention # seq2seq_attention architecture graphs (from input to logits)
βββ __init__.py # Graph
βββ encoder.py # Encoder
βββ decoder.py # Decoder
βββ data_loader.py # raw_date -> precossed_data -> generate_batch (using Dataset)
βββ hook.py # training or test hook feature (eg. print_variables)
βββ main.py # define experiment_fn
βββ model.py # define EstimatorSpec
Reference : hb-config, Dataset, experiments_fn, EstimatorSpec
Todo
- make dataset Korean dialog corpus like Cornell_Movie-Dialogs_Corpus
- Implements CopyNet
- Apply hb-research/notes - Neural Text Generation: A Practical Guide
Config
Can control all Experimental environment.
example: cornell-movie-dialogs.yml
data:
base_path: 'data/cornell_movie_dialogs_corpus/'
conversation_fname: 'movie_conversations.txt'
line_fname: 'movie_lines.txt'
processed_path: 'processed_cornell_movie_dialogs_data'
word_threshold: 2
max_seq_length: 200
sentence_diff: 0.33 # (Filtering with input and output sentence diff)
testset_size: 25000
PAD_ID: 0
UNK_ID: 1
START_ID: 2
EOS_ID: 3
model:
batch_size: 32
num_layers: 4
num_units: 512
embed_dim: 256
embed_share: true # (true or false)
cell_type: gru # (lstm, gru, layer_norm_lstm, nas)
dropout: 0.2
encoder_type: bi # (uni / bi)
attention_mechanism: normed_bahdanau # (bahdanau, normed_bahdanau, luong, scaled_luong)
train:
learning_rate: 0.001
sampling_probability: 0.25 # (Scheduled Sampling)
train_steps: 100000
model_dir: 'logs/cornell_movie_dialogs'
save_checkpoints_steps: 1000
loss_hook_n_iter: 1000
check_hook_n_iter: 1000
min_eval_frequency: 1000
print_verbose: True
debug: False
predict:
beam_width: 5 # (0: GreedyEmbeddingHelper, 1>=: BeamSearchDecoder)
length_penalty_weight: 1.0
slack:
webhook_url: "" # after training notify you using slack-webhookUsage
Install requirements.
pip install -r requirements.txt
First, check if the model is valid.
python main.py --config check_tiny --mode train
Then, download Cornell_Movie-Dialogs_Corpus and train it.
sh scripts/prepare_Cornell_Movie-Dialogs_Corpus.sh
python data_loader --config cornell-movie-dialogs # pre-processing data
python main.py --config cornell-movie-dialogs --mode train_and_evaluate
Experiments modes
β evaluate: Evaluate on the evaluation data.β½ extend_train_hooks: Extends the hooks for training.β½ reset_export_strategies: Resets the export strategies with the new_export_strategies.β½ run_std_server: Starts a TensorFlow server and joins the serving thread.β½ test: Tests training, evaluating and exporting the estimator for a single step.β train: Fit the estimator using the training data.β train_and_evaluate: Interleaves training and evaluation.
After training, start chatting.
python chat.py --config cornell-movie-dialogs
Tensorboard
tensorboard --logdir logs
Conversation Example
> (Human)
> hi.
i ' m sorry . <\s>
> why?
i ' m not interested .
> hmm.. are you human?
nnnnn . <\s>
> you are you?
i ' m not going to marry you . <\s>
> what? me also!
i ' m sorry . <\s>
> anyway, what is your name?
i ' m not going to kill you . <\s>
> i'm asking your name.
i ' m sorry . <\s>
> yeah, so.. where are you from?
i ' m not sure . <\s>
> Have you lost your memory?
i ' m not sure i ' m not going to you . <\s>
> Have you ever been to Florida?
i ' m not sure i ' m not a little . <\s>
> there there.
no , no , no , no ,
> Are you alright?
i ' m not going to die . <\s>
> alsoulutely, you are very healthy.
i don ' t know . <\s>
> you know that? you are robot.
i ' m not sure . <\s>
> i don ' t know what to say.
i ' m not sure i ' m not a little stupid . <\s>
> what do you like to do in your free time?
i ' m not sure i ' m not gonna take you a little more . <\s>
> are you in love?
i ' m not sure i ' m not stupid . <\s>
> i know.
i ' m not going to let you
> what's your name?
i ' m not sure i ' m not a good guy . <\s>
> oh god..
i ' m not going to do that . <\s>
> let's talk about something else.
i ' m not sure . <\s>
> i have to figure out what is the problem.
i ' m not sure . <\s>
> because of you are stupid.
i ' m not going to get you . <\s>
> florida?
he is a homosexual - -
> really?
i ' m not sure i
> can you tell me about more detail about him?
i ' m not a cop . <\s>
> okay.
i ' m not going to
> you always say 'i....'
i ' m not sure i ' m not a murderer . <\s>
> hmm.....
i ' m not sure i ' m not a little popular . <\s>
> you are not popular.
i ' m not sure i ' m not a girl .
> are you boy?
i ' m not sure . <\s>
> are you girl?
i ' m not sure . <\s>
> who are you?
i ' m sorry . <\s>
> i'm sorry.
i ' m not crowding . <\s>
> just kidding.
i ' m going to be a writer
> great!
i don ' t know .
> what do you mean?
i ' m sorry , i ' m sorry . <\s>
> omg
no . <\s>
> florida.
he hasn ' t got any
> music is my life.
you ' re a fool , you ' re a man !
> wow!
i ' m not kidding .
> sorry.
i ' m not kidding .
> yeah, i know it.
i ' m not sure . <\s>
> california?
what ? <\s>
> haha.
it ' s a real competition
Reference
- stanford-tensorflow-tutorials by Chip Huyen
- TensorFlow Neural Machine Translation Tutorial - Tensorflow Official
- Deep Learning for Chatbots, Part 1 β Introduction - WildML
- hb-research/notes - Neural Text Generation: A Practical Guide
Author
Dongjun Lee ([email protected])