ml-lab / neural-chatbot

A chatbot based on seq2seq architecture done with tensorflow.

neural-chatbot

A chatbot based on seq2seq architecture.

About

This is the successor to my previous torch project 'seq2seq-chatbot'. Portions of this project were directly adapted from the tensorflow translation example.

This is based off the research in these papers: Sutskever et al., 2014. and Vinyals & Le, 2015.

Dependencies

1. Python 2.7
2. TensorFlow
3. (optional) matplotlib $ pip install matplotlib

How to use

Getting data

To use this (currently) you will have to provide your own data. This is temporary until I add download functionality to an existing online dataset. Your data should be in a format where each succeeding line in a text file is in response to the one above it. You also should have one text file per conversation. In order to help you choose bucket sizes I've written a script to plot the sequence lengths of the source and target sequences. To use this do:

$ python sequencelengthplotter.py --data_dir="path/to/raw_data_files"

The options for the above command are:

Name	Type	Description
data_dir	string	path to raw data files
plot_histograms	boolean	plots histograms for target and source sequences
plot_scatter	boolean	plot x-y scatter plot for target length vs. source length

in buckets.cfg you can then modify the bucket values accordingly. You can add or remove buckets. All bucket values in the [buckets] subheading will be parsed in. Each line under [buckets] should be of the format:

bucket_name: source_length,target_length

In the same configuration file the data settings can also be changed under the [max_data_sizes] subheading.

Name	Type	Description
num_lines	int	number of lines in conversation to go back
max_target_length	int	max length of target sequences
max_source_length	int	max length of source sequences

A configuration file was used because it was a mess trying to find out how to pass bucket values via command line. This seemed like a half-way decent solution. It also enables (in the future) sequencelengthplotter.py the ability to autonomously change these values without requiring any user input.

Training Network

Once you are satisfied with bucket sizes you can then run the optimizer. This will also clean and process your raw text data. To do this:

$ python train.py

There are several options that can be employed:

Name	Type	Description
hidden_size	int	number of hidden units in hidden layers
num_layers	int	number of hidden layers
batch_size	int	size of batchs in training
max_epoch	int	max number of epochs to train for
learning_rate	float	beggining learning rate
steps_per_checkpoint	int	number of steps before running test set
lr_decay_factor	float	factor by which to decay learning rate
batch_size	int	size of batchs in training
checkpoint_dir	string	directory to store/restore checkpoints
dropout	float	probability of hidden inputs being removed
grad_clip	int	max gradient norm
max_train_data_size	int	Use a subset of available data to train
vocab_size	int	max vocab size
train_frac	int	percentage of data to use for training (rest is used for testing)
raw_data_dir	int	raw conversation text files stored here
data_dir	int	Directory data processor will store train/test/vocab files
is_discrete	boolean	True if your data is discrete pairs of input/output where inputs are not also outputs

Tensorboard Usage

After training loss and perplexity graphs can be seen:

$ tensorboard --logdir=/tmp/tb_logs_chatbot/

Sampling output

To have a real conversation with your bot you can begin an interactive prompt by doing:

*Note the sample.py file does not currently read in bucket sizes set from the config file (this will be fixed shortly)

$ python sample.py --checkpoint_dir="path/to/checkpointdirectory" --data_dir="path/to/datadirectory"

A prompt will open up something like:

>

The user can then type in to the prompt

ex/ > Hello chatbot, how are you?

Then hit 'enter' for the response.

Summary of options below:

Name	Type	Description
checkpoint_dir	string	path to saved checkpoint
data_dir	string	path to directory where vocab file is

Results

So far using a Titan X, after about 12 hours of training, it achieves a perplexity of ~30 on a 'relatively small' network. Results with human testing so far haven't been too great. I am trying to find the right set of parameters to get something 'ok', that can be trained in 24 hours or less on my GPU. Results will be added here when found.

Future Features

-Automatic bucket selection and making it easier to change bucket sizes

-Automatic download and preparation of an existing conversational dataset