This is a framework to train and evaluate models using modular adaptation techniques as proposed in the paper Modular Domain Adaptation (TODO add link).

Using this framework, the model producer trains models using anticipatory domain adaptation techniques, and the model consumer uses the same techniques when evaluating and predicting with the model on a new domain, to address the domain misalignment issue without the need of releasing the training data. Please see the paper for further details.

Clone this repository and use pip at the repo root

pip install -e .

We recommend using conda. All code is developed using python 3.8. Create a conda environment and install the library within it:

conda create -n mda_env python=3.8
conda activate mda_env
pip install -e .

Test out installation is successful, run the following to produce no output:

python -c 'from mda.api import train_lexicon'

CUDA GPUs are supported and used if detected. When using newer CUDA versions, please first install the correct version of torchvision with as instructed at pytroch's installation page using pip, before using pip install -e ..

Please see the demo section for usage of the main APIs.

We release a logistic regression and a RoBERTa model pretrained using all domains in the train split of Media Frames Corpus.

• The logistic regression model is released in the form of a lexicon csv, it is trained using domain-specific normalization and domain-specific bias.
• The RoBERTa model in the form of a torch state dict checkpoint, it is trained using domain-specific bias.

Please follow the example in demos to load, eval, and predict witht them.

If you find this work useful, please include a citation to the following publication:

@inproceedings{chen.2022.modular,
  author={Junshen K. Chen and Dallas Card and Dan Jurafsky},
  title={Modular Domain Adaptation},
  booktitle={Findings of ACL},
  year={2022},
}