faithful-nmn

This repository contains the code associated with the paper Obtaining Faithful Interpretations from Compositional Neural Networks, published at ACL 2020.

Dependencies

PyTorch 1.2.0 is a dependency. Please see the requirements.txt file for the other dependencies. Note that the required AllenNLP version is 0.9.0. Also, significant amounts of code were reused from LXMERT and Learning to Count Objects in Natural Images for Visual Question Answering

Setup Steps

We recommend the following steps to get started using this repository.

1. Create a new Python environment. Using anaconda, you can do this with conda create -n faithful_nmn python=3.6 pip
2. Install the dependencies. First, follow the instructions on the PyTorch website to install PyTorch with the correct CUDA version. Then run pip install -r requirements.txt.
3. Download the necessary data and put it in the dataset/nlvr2/ directory (see next section).
4. Download the LXMERT pre-trained model:

wget --no-check-certificate https://nlp1.cs.unc.edu/data/model_LXRT.pth -P lxmert_pretrained/

NLVR2 Data preparation

Please download the following files and put them in the dataset/nlvr2/ directory. They were created by the LXMERT authors (or by running the LXMERT data preparation code).

wget https://faithful-nmn.s3-us-west-2.amazonaws.com/train.json dataset/nlvr2/
wget https://faithful-nmn.s3-us-west-2.amazonaws.com/valid.json dataset/nlvr2/
wget --no-check-certificate https://nlp1.cs.unc.edu/data/lxmert_data/nlvr2_imgfeat/train_obj36.zip dataset/nlvr2/
unzip dataset/nlvr2/train_obj36.zip -d dataset/nlvr2/ && rm dataset/nlvr2/train_obj36.zip
wget --no-check-certificate https://nlp1.cs.unc.edu/data/lxmert_data/nlvr2_imgfeat/valid_obj36.zip dataset/nlvr2/
unzip dataset/nlvr2/valid_obj36.zip -d dataset/nlvr2/ && rm dataset/nlvr2/valid_obj36.zip

Pre-trained models

• LayerCount NMN

• SumCount NMN

• CountModule NMN

• CountModule+Decontextualization NMN

• CountModule+Pretraining NMN

• CountModule+Pretraining+Decontextualization NMN

Running pre-trained model

To get model predictions on the validation set, use the following command

allennlp predict MODEL_URL dataset/nlvr2/valid.json --include-package lib --include-package predictors --use-dataset-reader --dataset-reader validation --output-file predictions.json --predictor full --cuda-device 0

, where MODEL_URL should be replaced with the URL for the model that you want to use to obtain predictions. For instance, for the CountModule+Pretraining+Decontextualization NMN, the command should be

allennlp predict https://faithful-nmn.s3-us-west-2.amazonaws.com/countmodule_pretraining_decontextualization_nmn_model.tar.gz dataset/nlvr2/valid.json --include-package lib --include-package predictors --use-dataset-reader --dataset-reader validation --output-file predictions.json --predictor full --cuda-device 0

Evaluating pre-trained model

To evaluate model for accuracy on the validation set, use the following command.

allennlp evaluate model.tar.gz dataset/nlvr2/valid.json --include-package lib

To evaluate model for faithfulness on the portion of the validation set with annotated module outputs, use the following command.

allennlp evaluate model.tar.gz dataset/nlvr2/valid.json --include-package lib -o "{"validation_dataset_reader": {"box_annotations_path": "\"dataset/nlvr2/nlvr2_test_box_annotations.csv\"", "only_with_box_annotation": "true"}}"

Annotation preparation

We provide annotations in the format necessary for our models in dataset/nlvr2/all_annotations_round4.tsv. If you would like to generate these annotations yourself from the QDMR annotations, please run the following commands:

cd dataset/nlvr2
python process_raw_data_scripts/process_annotations.py nlvr2_qdmr_programs_11k.csv
python process_raw_data_scripts/filter_annotations.py train_annotations_from_qdmr_rawnumber.tsv dev_annotations_from_qdmr_rawnumber.tsv test1_annotations_from_qdmr_rawnumber.tsv simple_count_annotations.tsv
mv filtered_annotations.tsv all_annotations.tsv

Note that the resulting file will not be identical to dataset/nlvr2/all_annotations_round4.tsv because it will include test annotations and around 200 extra programs for simple count questions that are in simple_count_annotations.tsv.

Training a model

To train a model, please use the following command: allennlp train train_configs/nlvr2.jsonnet -s serialized_dir --include-package lib

DROP Experiments

Data: dataset/drop contains the manual annotations for intermediate module outputs for DROP.

Code: Text-NMN-faithful -- Model code, evaluation scripts, etc. can be found here.

Citation

If you find our work relevant/useful to yours, please cite our paper:

@inproceedings{subramanian-etal-2020-obtaining,
    title = "Obtaining Faithful Interpretations from Compositional Neural Networks",
    author = "Subramanian, Sanjay  and
      Bogin, Ben  and
      Gupta, Nitish  and
      Wolfson, Tomer  and
      Singh, Sameer  and
      Berant, Jonathan  and
      Gardner, Matt",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.495",
    pages = "5594--5608"
}