Reproduce the unsupervised NMT results (De-En)

cindyxinyiwang / deep-latent-sequence-model

Pytorch implementation of "A Probabilistic Formulation of Unsupervised Text Style Transfer" by He. et. al. at ICLR 2020

163 stars 26 forks source link

Reproduce the unsupervised NMT results (De-En) #1

Closed bzhangGo closed 4 years ago

bzhangGo commented 4 years ago

Hi, could you please show more details and instructions on how to reproduce the WMT16 En-De translation results (En-De 26.9, De-En 32.0) with only 5M non-parallel sentences using this codebase? Thanks.

jxhe commented 4 years ago

Hi, the unsupervised NMT results are based on a slight modification of a separate codebase, XLM, for a fair comparison with the SOTA systems. We basically follow their instructions for hyperparameters and data collection, and add a language model based KL loss on that. Unfortunately our modified code is on a cluster which I don't have access to now.

bzhangGo commented 4 years ago

thanks.