jhyuklee / namegen-tensorflow

Draft implementation of conditional name generation

Name-generation

Repository for Generative model for names of Olympic Name Dataset Created at March 2, 2017 Korea Unviversity, Data-Mining Lab

Tensorflow Implementation for Name-GAN(Generative Adversarial Network).

• Jinhyuk Lee ([e-mail])
• Bumsoo Kim ([e-mail])

Requirements

Take a look at the installation instruction for details about installation.

• Install cuda-8.0
• Install cudnn-v5.1
• Install Tensorflow

# Install Tensorflow GPU version
$ sudo apt-get install python-pip python-dev
$ pip install tensorflow-gpu

# If the code above doesn't work, try
$ sudo -H pip install tensorflow-gpu
$ sudo pip install --upgrade

Directories

• utils.py : Progress bar function
• model.py : GAN(encoder+decoder+generator+discriminator model) + load + save
• ops.py : Basic functions for rnn, lstm, feed-forward neural network, dropout
• dataset.py : Crawling name dataset and train the autoencoder(encoder + decoder) model
• main.py : Training script for the GAN model (requires pretrained autoencoder model from dataset.py)

Modules

• Each name string will be represented into an encoded vector consisted of (h+c)
• h : LSTM's hidden state
• C : LSTM's output state

1. Encoder

• Encodes the given string value into a hidden vector h
• Output of the model = (2 x cell-dim) = (LSTM's h) + (LSTM's c)

Division	representation	specifics
input	x	character-level embedding of name strings
output	h	vector-level representation of name strings
model	RNN	input-size * time-step -> (2 x cell-dim)

2. Decoder

• Decodes the given hidden vector into an approximated string value x-hat
• Output of the model will (time_steps x input_dim)

Division	representation	specifics
input	h	vector-level representation of name strings
output	x-hat	near-value reconstruction of 'x'
model	RNN	(cell-dim) -> input-size * time-step

3. Generator (G)

• Generates a fake hidden vector representing a name string
• Output of the model = (2 x cell-dim) = (LSTM's h) + (LSTM's c)

Division	representation	specifics
input	Zc	Random input vector with class info
output	Xc	Generated name hidden vectors
model	Linear	(z_dim + class_dim) => (cell_dim * 2)

4. Discriminator (D)

• Binary classification. Define whether the given input is fake or not.

Division	representation	specifics
input	Xc	Hidden vector for name-class representation
output	Pc	Probabilities whether the input is fake or not
model	Linear	(cell_dim * 2 + class_dim) => p