Implementation of Xiaodong Gu, Kyunghyun Cho, Jung-Woo Ha, Sunghun Kim, DialogWAE: Multimodal Response Generation with Conditional Wasserstein Auto-Encoder, Published as a conference paper at ICLR 2019 https://arxiv.org/pdf/1805.12352.pdf
http://yanran.li/files/ijcnlp_dailydialog.zip
http://nlp.stanford.edu/data/glove.twitter.27B.zip
In the IPython notebook:
Version 1 is my implementation of the model and data loading and batching. It achieves very low accuracy; about 0.05 BLEU score on DailyDialog.
Version 2 is my implementation of the model integrated with the author's dataloader. It is able to reproduce the paper's results; about 0.3 BLEU score on DailyDialog.
The difference between the two dataloaders is that the author's use a type of bucketing. Dialogues of same number of turns are batched together. This makes a big difference in performance.