Skip to content

Latest commit

 

History

History
146 lines (94 loc) · 5.62 KB

README.md

File metadata and controls

146 lines (94 loc) · 5.62 KB
Raw

Running docker files

First of all make sure to install docker :)

Follow the directory to the lick-caption-bias, and when you are in the directory execute this command.

For lstm 

docker build -f ./Dockerfiles/Dockerfile_lstm -t lstm_fact . 
docker run -it lstm_fact /bin/bash
#In the container
cd lick-caption-bias

For BERT 

docker build -f ./Dockerfiles/Dockerfile_bert -t bert_fact . 
docker run -it bert_fact /bin/bash
#In the container
cd lick-caption-bias

You should be in the container now. Go to the lick-caption-bias directory and run the scripts as described below.

Quantifying Societal Bias Amplification in Image Captioning

This repository contains source code necessary to reproduce the results presented in the paper Quantifying Societal Bias Amplification in Image Captioning (CVPR 2022, Oral). Please check the project website here.

LIC metric

LIC metric measures how much biased a set of model generated captions are with respect to the captions in the training dataset. LIC is computer as follows:

  1. Mask attribute-revealing words.

  1. Train 2 classifiers, 1 for human captions and 1 for generated captions, with attribute labels. The classifiers' goal is to predict the attribute (e.g. gender, race) for the person in the image using only the captions.

  2. Calculate LIC scores for each classifier. If the set of captions is not biased, the classifier accuracy should be close to random chance.

  1. To compute bias amplification, take the difference of the LIC socres between 2 classifiers.

Setup

  1. Clone the repository.
  2. Download the data (folder name: bias_data) and place in the current directory. The folder contains human/generated captions and corresponding gender/racial annotations from the paper Understanding and Evaluating Racial Biases in Image Captioning.
  3. Install dependancies:

For LSTM classifier

- Python 3.7
- numpy 1.21.2 
- pytorch 1.9
- torchtext 0.10.0 
- spacy 3.4.0 
- sklearn 1.0.2 
- nltk 3.6.3

For BERT classifier

- Python 3.7
- numpy 1.21.2 
- pytorch 1.4
- transformers 4.0.1
- spacy 2.3
- sklearn 1.0.2 
- nltk 3.6.3

Compute LIC

We evaluate various captioning models (i.e. NIC, SAT, FC, Att2in, UpDn, Transformer, OSCAR, NIC+, and NIC+Equalizer). In the following commands, you can select a model in $model_name from them (i.e. nic, sat, fc, att2in, updn, transformer, oscar, nic_equalizer, or nic_plus).

In the paper, LSTM or BERT is used as the classifier. Please run the following commands according to the classifier you prefer to use.


  • To train the LSTM classifier on human captions and compute LIC in terms of gender bias run:

    python3 lstm_leakage.py --seed $int --cap_model $model_name --calc_ann_leak True

    Where $int is the arbitrary integer for random seed and $model_name is the choice of a captioning model to be compared.


  • To train the LSTM classifier on generated captions and compute LIC in terms of gender bias run:

    python3 lstm_leakage.py --seed $int --cap_model $model_name --calc_model_leak True

    Where $model_name is the choice of a captioning model.


  • To train the BERT classifier on human captions and compute LIC in terms of gender bias run:

    python3 bert_leakage.py --seed $int --cap_model $model_name --calc_ann_leak True


  • To train the BERT classifier on generated captions and compute LIC in terms of gender bias run:

    python3 bert_leakage.py --seed $int --cap_model $model_name --calc_model_leak True


Note: If you compute LIC in terms of racial bias, please run race_lstm_leakage.py or race_bert_leakage.py.

Note: To use pre-trained BERT without fine-tuning, you can add --freeze_bert True, --num_epochs 20, and --learning_rate 5e-5.

Results

Gender bias

Racial bias

Note: The classifier is trained 10 times with random initializations, and the results are reported by the average and standard deviation.

Citation

@inproceedings{hirota2022quantifying,
  title={Quantifying Societal Bias Amplification in Image Captioning},
  author={Hirota, Yusuke and Nakashima, Yuta and Garcia, Noa},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={13450--13459},
  year={2022}
 }