python==2.7 cuda 9.0
nltk
TensorFlow == 1.3.0
numpy
pickle
Rake (pip install python-rake)
zpar (pip install python-zpar, download model file from https://github.com/frcchang/zpar/releases/download/v0.7.5/english-models.zip and extract it to POS/english-models)
python source/run.py --exps_dir exps-sampling --exp_name test --use_data_path data/quoradata/test.txt --mode kw-bleu --N_repeat 1 --save_path sa.txt --batch_size 1 --gpu 0 --search_size 50
python source/evaluate.py --reference_path quora-results/ref.txt --generated_path quora-results/gen.txt
To evaluate the generation speed of each method, we have compared the average number of generated paraphrases of each model within a minute with either CPU (eight Intel Core i7-4790K CPUs) or GPU (NVIDIA GeForce GTX TITIAN X GPU) settings. In particular, for each method, we use a well-trained model to produce 20,000 paraphrases (with a batch size of 20) and record the consumed time. The results are shown in the following table.
| Model | # samples/minute on GPU | # samples/minute on CPUs |
|---|---|---|
| VAE | 1309.61 | 88.72 |
| Lag. VAE | 1275.20 | 87.29 |
| CGMH | 20.36 | 5.11 |
| UPSA | 40.28 | 10.93 |