Source code of paper: Composable Text Controls in Latent Space with ODEs

https://arxiv.org/abs/2208.00638

Code is coming soon...

We recommend to create a new conda enviroment (named latentops) by:

conda create -n latentops python==3.9.1 pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch

Then activate latentops and install the required packages by running:

conda activate latentops
bash build_envs.sh

Download and process the datasets by running the script:

bash download_datasets.sh

Download and process the pretrained model by running the script:

bash download_pretrained_models.sh

Download and process the external classifiers by running the script:

bash download_classifiers.sh

You can do conditional generation (default Yelp) by running:

cd code
bash conditional_generation.sh $1 $2

$1 represents operators (1 for sentiment, 4 for tense, 33 for formality). $2 represents desired labels:

• sentiment: 0-negative, 1-positive
• tense: 0-past, 1-present, 2-future
• formality: 0-informal, 1-formal

For examples, you can run:

# for positive sentences
bash conditional_generation.sh 1 1
# for past sentences
bash conditional_generation.sh 4 0
# for positive & future sentences
bash conditional_generation.sh '1,4' '1,2'
# for positive & future & informal
bash conditional_generation.sh '1,4,33' '1,2,0'
# for positive & future & informal and negative & future & informal
bash conditional_generation.sh '1,4,33' '1,2,0;0,2,0'

The generated files can be found in ../ckpts/model/sample/ (default: ../ckpts/large_yelp/sample/sampling*.txt)

Modify the path of data file in code/train_vae.sh

dataset=your_dataset_name
# e.g., dataset=yelp
TRAIN_FILE=path_to_train_data_file 
# e.g., TRAIN_FILE=../data/datasets/yelp_data/train.shuf.merge
TEST_FILE=path_to_test_data_file
# e.g., TEST_FILE=../data/datasets/yelp_data/test.merge

The structure of the data file: one line one sentence. See ../data/datasets/yelp_data/test.merge for example.

Then run the script to train a VAE

cd code
bash train_vae.sh

The checkpoints will be saved in ../ckpts/LM/$dataset/$name by default. You also can find the tensorboard logs in code/runs/$dataset

After training VAE, you can train the GAN and classifiers to do some operations.

You need to specify some key arguments:

train_cls_gan='gan'

ckpt_path=path_to_vae_ckpts  # e.g., ckpt_path=../ckpts/base_yelp

TRAIN_FILE=path_to_test_gan_data_file 
# e.g., TRAIN_FILE=../data/datasets/yelp_data/train_gan.txt

TRAIN_FILE=path_to_test_gan_data_file 
# e.g., TEST_FILE=../data/datasets/yelp_data/test.txt

The GAN training and test data file should have the line format (exclude bracket []): [0]\t[text], where the [0] is not used and meaningless in the training and it can be any other integer. See the example in ../data/datasets/yelp_data/train_gan.txt

Then run the below command to train GAN:

cd code
bash train_classifier_latent.sh

You need to specify some key arguments:

train_cls_gan='cls'

ckpt_path=path_to_vae_ckpts  # e.g., ckpt_path=../ckpts/base_yelp

TRAIN_FILE=path_to_test_cls_data_file 
# e.g., TRAIN_FILE=../data/datasets/yelp_data/train_sentiment.txt

TRAIN_FILE=path_to_test_cls_data_file 
# e.g., TEST_FILE=../data/datasets/yelp_data/test_sentiment.txt

cls_step=identifier_of_classifier
# identifier of this classifier, the classifier will be stored in path_to_vae_ckpts/checkpoint-cls-1 if cls_step=1

n_classes=number_of_classes
# number of classes,  e.g., if it contains 2 classes, n_classes=2

The Classifiers training and test data file should have the line format (exclude bracket []):[class_label]\t[text], where [class_label] should be the class label of the text, it should be a integer. If you have 2 classes, the [class_label] should be 0 or 1. If you have 3 classes, it should be 0, 1, or 2. See ../data/datasets/yelp_data/train_sentiment.txt for example.

Then run the below command to train Classifiers:

cd code
bash train_classifier_latent.sh

To facilitate comparison, we provide the output files of text editing with single attribute (text style transfer) in ./outputs folder.

@misc{liu2022composable,
      title={Composable Text Controls in Latent Space with ODEs}, 
      author={Guangyi Liu and Zeyu Feng and Yuan Gao and Zichao Yang and Xiaodan Liang and Junwei Bao and Xiaodong He and Shuguang Cui and Zhen Li and Zhiting Hu},
      year={2022},
      eprint={2208.00638},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}