Language Generation with Recurrent Generative Adversarial Networks without Pre-training

Code for training and evaluation of the model from "Language Generation with Recurrent Generative Adversarial Networks without Pre-training".

Additional Code for using Fisher GAN in Recurrent Generative Adversarial Networks

Sample outputs (32 chars)

" There has been to be a place w
On Friday , the stories in Kapac
From should be taken to make it 
He is conference for the first t
For a lost good talks to ever ti

Training

To start training the CL+VL+TH model, first download the dataset, available at http://www.statmt.org/lm-benchmark/, and extract it into the ./data directory.

Then use the following command:

python curriculum_training.py

The following packages are required:

• Python 2.7
• Tensorflow 1.1
• Scipy
• Matplotlib

The following parameters can be configured:

LOGS_DIR: Path to save model checkpoints and samples during training (defaults to './logs/')
DATA_DIR: Path to load the data from (defaults to './data/1-billion-word-language-modeling-benchmark-r13output/')
CKPT_PATH: Path to checkpoint file when restoring a saved model
BATCH_SIZE: Size of batch (defaults to 64)
CRITIC_ITERS: Number of iterations for the discriminator (defaults to 10)
GEN_ITERS: Number of iterations for the geneartor (defaults to 50)
MAX_N_EXAMPLES: Number of samples to load from dataset (defaults to 10000000)
GENERATOR_MODEL: Name of generator model (currently only 'Generator_GRU_CL_VL_TH' is available)
DISCRIMINATOR_MODEL: Name of discriminator model (currently only 'Discriminator_GRU' is available)
PICKLE_PATH: Path to PKL directory to hold cached pickle files (defaults to './pkl')
ITERATIONS_PER_SEQ_LENGTH: Number of iterations to run per each sequence length in the curriculum training (defaults to 15000)
NOISE_STDEV: Standard deviation for the noise vector (defaults to 10.0)
DISC_STATE_SIZE: Discriminator GRU state size (defaults to 512)
GEN_STATE_SIZE: Genarator GRU state size (defaults to 512)
TRAIN_FROM_CKPT: Boolean, set to True to restore from checkpoint (defaults to False)
GEN_GRU_LAYERS: Number of GRU layers for the genarator (defaults to 1)
DISC_GRU_LAYERS: Number of GRU layers for the discriminator (defaults to 1)
START_SEQ: Sequence length to start the curriculum learning with (defaults to 1)
END_SEQ: Sequence length to end the curriculum learning with (defaults to 32)
SAVE_CHECKPOINTS_EVERY: Save checkpoint every # steps (defaults to 25000)
LIMIT_BATCH: Boolean that indicates whether to limit the batch size  (defaults to true)
GAN_TYPE: String Type of GAN to use. Choose between 'wgan' and 'fgan' for wasserstein and fisher respectively

Parameters can be set by either changing their value in the config file or by passing them in the terminal:

python curriculum_training.py --START_SEQ=1 --END_SEQ=32

Monitoring Convergence During Training

In the wasserstein GAN, please monitor the disc_cost. It should be a negative number and approach zero. The disc_cost represents the negative wasserstein distance between gen and critic.

Generating text

The generate.py script will generate BATCH_SIZE samples using a saved model. It should be run using the parameters used to train the model (if they are different than the default values). For example:

python generate.py --CKPT_PATH=/path/to/checkpoint --DISC_GRU_LAYERS=2 --GEN_GRU_LAYERS=2

Evaluating text

To evaluate samples using our %-IN-TEST-n metrics, use the following command, linking to a txt file where each row is a sample:

python evaluate.py --INPUT_SAMPLE=/path/to/samples.txt

Experimental Features (not mentioned in the paper)

To train with fgan with recurrent highway cell:

python curriculum_training.py --GAN_TYPE fgan --CRITIC_ITERS 2 --GEN_ITERS 4 \
--PRINT_ITERATION 500 --ITERATIONS_PER_SEQ_LENGTH 60000 --RNN_CELL rhn

Please note that for fgan, there may be completely different hyperparameters that are more suitable for better convergence.

Monitoring Convergence

To measure fgan convergence, gen_cost should start at a positive number and decrease. The lower, the better.

Warning: in the very beginning of training, you may see the gen_cost rise. Please wait at least 5000 iterations and the gen_cost should start to lower. This phenomena is due to the critic finding the appropriate wasserstein distance and then the generator adjusting for it.

Acknowledgments

This repository is based on the code published in Improved Training of Wasserstein GANs.