dapascual / k2t Goto Github PK

Python 92.08% Shell 7.92%

k2t's Introduction

Keyword2Text

This repository contains the code of the paper: "A Plug-and-Play Method for Controlled Text Generation", if you find this useful and use it for your own research, please cite us.

Setup

Download and unzip the repository.
Create a new conda environment and install the required libraries from the requirements.txt file.

conda create -n k2t python=3.6
conda activate k2t
pip install -r requirements.txt

A GPU will be required to run the experiments. Make sure you have a results folder.

Run Model

Hyperparameter Study

Uncomment the appropriate lines of run.sh to run the hyperparameter experiments from the paper. For example,

python main.py -mode='next' -file_name=/data/50_keywordsets_eval/word_sets.txt -results_subfolder=guide_vs_no_guide_beams -weight=10.0 -top_p=0.9 -n_generated_sentences=90 -do_guarantee=True

runs K2T with ordered guide words (mode='next') on the random keywords dataset. It runs with lambda=weight=10, nucleus sampling with top-p=0.9, number of generated tokens = 90, and no weight annealing to guarantee word appearance. The results are saved in results/tmp

ROC Story dataset

Uncomment the appropriate line of run.sh to run the model on the ROC story dataset:

python main.py -mode='max' -file_name=/data/ROC/ROCStories_20_storylines_500_0.txt -results_subfolder=final4_ -weight=5.0 -top_p=0.9 -n_generated_sentences=-7 -n_beams=4 -do_guarantee=True -task='ROC'

News Article dataset

Uncomment the appropriate line of run.sh to run the model on the News Article story dataset:

python main_DBS.py -mode='max' -file_name=/data/keyword_to_articles -results_subfolder=tmp -weight=5.0 -top_p=0.9 -n_generated_sentences=-15 -n_beams=4 -do_guarantee=True -task='key2article'

├── data
│   ├── 50_keywordsets_eval
│   │   └── word_sets.txt
│   ├── keyword_to_articles
│   │   ├── test_10.txt
│   │   ├── test_12.txt
│   │   ├── test_13.txt
│   │   ├── test_14.txt
│   │   ├── test_15.txt
│   │   ├── test_16.txt
│   │   ├── test_4.txt
│   │   ├── test_5.txt
│   │   ├── test_8.txt
│   │   └── test_9.txt
│   └── ROC
│       └── ROCStories_20_storylines_500_0.txt
├── encode_keywords.py
├── encode_keywords_word2vec.py
├── main.py
├── metrics_degen.py
├── metrics_degen_run.sh
├── perplexity.py
├── README.md
├── requirements.txt
├── results
├── run.sh
└── utility_gpt.py

k2t's People

Contributors

Stargazers

Watchers

Forkers

tannonk calandula techthiyanes cahya-wirawan alronz saurabh3949 minji-o-j poisonbox haruboring minhduc19 saradhix

k2t's Issues

numpy==1.21.2 not compatible with python3.6

I am trying to install numpy==1.21.2

pip install numpy==1.21.2 shows Version not found

installing from source says requires python version >= 3.7

Generate with context sentence as input

Hi thank you for great work,

I have a question on how to generate with context sentence as input of your model as you did in your paper appendix.

For example:

Text 1

Context: Johnny Depp will be asked to make a fifth Pirates Of The Caribbean film if the fourth
instalment is a success. Producer Jerry Bruckheimer said he already has a screenplay

Keywords: fifth, audience, embraces, character, original, trilogy, continue, story, digital, cameras,
pirates, fresh, new, carry

OURS + GPT2-774M: Johnny Depp will be asked to make a fifth Pirates Of The Caribbean film if
the fourth instalment is a success. Producer Jerry Bruckheimer said he already has a screenplay with
Pirates new trilogy character Johnny Depp and digital animation studio Story Studio. The fifth film
will continue the original story from the first three films and will be set in the Caribbean. "We are very
excited to carry on the story of the audience’s favourite Pirates of the Caribbean characters," said
Bruckheimer. "We are fresh off the success of Pirates of the Caribbean: On Stranger Tides and we
are looking forward to embracing the new generation of fans with a new film that will be even more
exciting than the first three films." The Pirates of the Caribbean: On Stranger Tides cameras have been
filming in Barranquilla since April 11. Filming will continue for two months to reach completion by
May 23

Thank you again for your great work!

variable result_subsequences shape should be number_of_beam * number_of_concurrent_sentences * values_list?

As you are selecting best from number of concurrent sentences from each beam.
currently the shape is 2D (number_of beam x number_of_concurrent_sentences) * value_list

Let me know If I'm understanding something wrong.

Inference time

Hi, I am testing your code, the result looks very good, however when I tried it (several times), your code need around 1 to 4 minutes to generate sentence for one article "data/keyword_to_articles/test_5.txt" (similar times for other articles). Is it expected? I use and old gpu (nvidia 1080), but I think it wouldn't make a big different with newer gpu.

License

Hi,
thanks for the clean and easy to read implementation! I would like to use and modify the code and add custom conditions for deleting branches, masking logits etc.
As far as I can see there is not any license applied to your codes, so officially I would not be allowed to do any of these and my university is strict in this concern. Would I be allowed to do above and present our work - of course with proper citation? It would be a building brick in a larger model.