Style transfer

Goal: implement a tool changing style of a provided text.

Sources:

• Style Transfer from Non-Parallel Text by Cross-Alignment

  • Paper: https://arxiv.org/pdf/1705.09655.pdf
  • Code: https://github.com/shentianxiao/language-style-transfer

• Adversarial Text Generation via Feature-Mover’s Distance

  • Paper: https://papers.nips.cc/paper/7717-adversarial-text-generation-via-feature-movers-distance.pdf
  • Code: https://github.com/shentianxiao/language-style-transfer [1]

Progress:

1. Implemented data downloading (imdb) that will later be used for training. (17.07.19)
2. Changed code in language-style-transfer from python2 to python3. (17.07.19)
3. Found out how to run language-style-transfer code (python3.6 and tensorflow==1.3.0). Added casting to tf.int32 to fix a bug. (19.07.19)
4. Fixed some other bugs in language-style-transfer regarding running script, created running instructions in readme. Downloaded and tested model according to instructions in readme created by authors (python3.6 style_transfer.py --online_testing ../data/yelp/sentiment.test --output ../tmp/sentiment.test --vocab ../model/yelp.vocab --model ../model/model --load_model true --beam 8). It's possible to run this model in interactive mode: after writing it turns out it outputs translation to similar sentence and sentence with style transfer. It is not that good, especially considering that I use model pretrained by authors. (30.09.19)

TODO:

1. Run chainger-fm-gan code (write requirements file / docker) on some corpora.

[1] Reimplementation of original code. Link to author's repository at http://zhegan27.github.io/Paper.html) was not available.