Tensorflow r0.10
Cuda 7.5 (for GPU)
nltk python package
For full information, see the paper:
SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient (http://arxiv.org/abs/1609.05473)
We provide example codes to repeat the synthetic data experiments with oracle evaluation mechanisms. Move to MLE_SeqGAN folder and run
python pretrain_experiment.py
will start maximum likelihood training with default parameters. In the same folder, run
python sequence_gan.py
will start SeqGAN training. After installing nltk python package, move to pg_bleu folder and run
python pg_bleu.py
will start policy gradient algorithm with BLEU score (PG-BLEU), where the final reward for MC search comes
from a predefined score function instead of a CNN classifier.
Finally, move to schedule_sampling folder and run
python schedule_sampling.py
will launch SS algorithm with default parameters.