A Pytorch implementation of GAN-BERT paper
"bert.py" and "qc-fine_bert.py" files are basically the same with different datasets; this is just my laziness. "labeled_and_unlabeled.tsv" is the mixed version of labeled and unlabeled data from original tensorflow implementation of the paper.
Running this model only once with only one seed might not be enough. In my experiments, out of 10 runs, only 2 yielded reasonable results. I believe this is due to the difference of scheduling of the optimizers between this implementation and the original TensorFlow implementation.