seqGAN

Why we included this project

This is a compact, readable PyTorch reimplementation of the SeqGAN paper, aimed at people learning how generative adversarial networks handle discrete sequence data. The author stripped the original reference code down to its essentials, swapping rollout-based policy gradients for a single reward per sentence and using a bidirectional GRU discriminator, so the training loop is much easier to follow. The comments are thorough, and the repo documents the hacks that actually helped stabilize training, like training the discriminator more than the generator and giving each network its own optimizer. Everything runs on synthetic data, so you can reproduce the learning curve and watch how MLE pretraining followed by adversarial training plays out without hunting down a large dataset.