GTAE: Graph-Transformer Based Auto Encoder for Text Style Transfer

Benchmark

We also provide a benchmark at https://github.com/ykshi/text-style-transfer-benchmark

Preprocessing:

jdk 1.8
stanford-corenlp (http://nlp.stanford.edu/software/stanford-corenlp-full-2018-02-27.zip)
nltk==3.5
tqdm

Training:

generate the linguistic adjacency matrices using stanford nlp
- go into the directory of stanford nlp
  - 'cd stanford-corenlp-full-2018-02-27'
- start server
  - 'java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -preload tokenize,ssplit,pos,lemma,ner,parse,depparse -status_port 9000 -port 9000 -timeout 30000 &>/dev/null'
- extract raw adjacency file for a text file, e.g.
  - python utils_preproc/stanford_dependency.py data/yelp/sentiment.train.text data/yelp/sentiment.train.adjs 9000
  - 9000 is the server port, which should be consistent with the previous step
  - for large text file, split into multiple sub-files first and run stanford_dependency in multi-processes
- build adjacency matrices from raw adjacency file, e.g.
  - ＇python utils_preproc/dataset_read.py data/yelp/sentiment.train.adjs data/yelp/sentiment.train_adjs.tfrecords data/yelp/sentiment.train_identities.tfrecords'
generate vocab of trainning data, e.g.
- 'python get_vocab.py data/yelp/sentiment.train.text data/yelp/vocab_yelp'
Configure your data paths and model parameters as specified in 'config_gtt.py'
Training:
- 'CUDA_VISIBLE_DEVICES=0 python main.py --config config --out output_path --lambda_t_graph 0.05 --lambda_t_sentence 0.02 --pretrain_nepochs 10 --fulltrain_nepochs 3'
- --out is necessary
- checkpoints/ is not saved to output_path automatically (too large). Save this folder manually if necessary, otherwise it will be erased every time we run main.py