This repository contains code for the paper Data Feedback Loops: Model-driven Amplification of Dataset Biases.
Python 3.9 and PyTorch installation are the most important. All dependencies are marked in requirements.txt, installable in a Python 3.9 environment via:
pip install -r requirements.txt
To automatically download and preprocess all necessary data, run:
cd data && ./download_preprocess_all.sh
Or, you can download each dataset individually with the provided scripts in the data directory.
Note that downloading cifar5m requires the gsutil utility. Also note that cinic10 needs to be preprocessed before use; if it's already downloaded, you can run the preprocess script on your download:
cd data && python preprocess_cinic10.py --data-dir path/to/your/install
In the paper, each experiment was run 3 times. Results are logged to wandb for plotting into figures. All programs are meant to be run with 1 GPU only.
Code for image classification experiments (Section 5.1) is in image_classification, code for visual role-labeing experiments (Section 5.2) is in visual_role_labeling, and code for conditional language generation experiments (Section 5.3) is in language_generation.
Figure 3 top left:
python image_classification/train.py -m 1000 -k 4000 --wandb-group fig3-top --wandb-log
Figure 3 top right:
python image_classification/train.py -m 2500 -k 2500 --wandb-group fig3-top --wandb-log
Figure 3 bottom left:
python image_classification/train.py -m 1000 -k 4000 --subsample-train-set-each-round --wandb-group fig3-bottom --wandb-log
Figure 3 bottom right:
python image_classification/train.py -m 2500 -k 2500 --subsample-train-set-each-round --wandb-group fig3-bottom --wandb-log
Figure 4 left:
python visual_role_labeling/train.py -m 1000 -k 4000 --wandb-group fig4 --wandb-log
Figure 4 right:
python visual_role_labeling/train.py -m 2500 -k 2500 --wandb-group fig4 --wandb-log
Figure 5 left:
python language_generation/train.py -m 1000 -k 4000 --sampling-type nucleus_sampling --wandb-group fig5 --wandb-log
Figure 5 right:
python language_generation/train.py -m 1000 -k 4000 --sampling-type beam_search --wandb-group fig5 --wandb-log
Figure 6 red line (blue & yellow lines are the same as in Figure 5):
python language_generation/train.py -m 1000 -k 4000 --sampling-type beam_search --num-train-epochs 5 --learning-rate 5e-4 --wandb-group fig6 --wandb-log
All the plotting code is in the plotting directory. The code to reproduce the plots for each figure is in the plotting/paper_plots.ipynb notebook. Note that you will first need to run the experiments you want to plot so that data is stored in wandb for the plotting notebook to use.
Theorem 1 upper bound calculations from the paper are implemented in plotting/feedback_bound.py.
Figure 7 is automatically logged with Figures 3 top left and top right.
Figures 8, 15, 16, and 17 are automatically logged with Figure 4.
Figure 9 left:
python image_classification/train.py -m 1000 -k 4000 -a resnet18 --batch-size 128 --wandb-group fig9 --wandb-log
Figure 9 right:
python image_classification/train.py -m 2500 -k 2500 -a resnet18 --batch-size 128 --wandb-group fig9 --wandb-log
Figure 10 left:
python image_classification/train.py -d cinic10 -n 20000 -m 1000 -k 4000 --class-imbalance-factor 2 --wandb-group fig9 --wandb-log
Figure 10 right:
python image_classification/train.py -d cinic10 -n 20000 -m 2500 -k 2500 --class-imbalance-factor 2 --wandb-group fig9 --wandb-log
Figure 11 left:
python image_classification/train.py -m 1000 -k 4000 --class-imbalance-factor 2 --wandb-group fig11 --wandb-log
Figure 11 right:
python image_classification/train.py -m 2500 -k 2500 --class-imbalance-factor 2 --wandb-group fig11 --wandb-log
Figure 12 left:
python image_classification/train.py -m 1000 -k 4000 --class-imbalance-class ship --wandb-group fig12 --wandb-log
Figure 12 right:
python image_classification/train.py -m 2500 -k 2500 --class-imbalance-class ship --wandb-group fig12 --wandb-log
Figure 13 left:
python image_classification/train.py -m 1000 -k 4000 --underfit-model --wandb-group fig13 --wandb-log
Figure 13 right:
python image_classification/train.py -m 2500 -k 2500 --underfit-model --wandb-group fig13 --wandb-log
Figure 14 left:
python image_classification/train.py -n 20000 -m 1000 -k 4000 --wandb-group fig14 --wandb-log
Figure 14 right:
python image_classification/train.py -n 20000 -m 2500 -k 2500 --wandb-group fig14 --wandb-log
Figure 18 left:
python language_generation/train.py -m 2500 -k 2500 --sampling-type nucleus_sampling --wandb-group fig18 --wandb-log
Figure 18 right:
python language_generation/train.py -m 2500 -k 2500 --sampling-type beam_search --wandb-group fig18 --wandb-log
Figure 19 left:
python language_generation/train.py -m 1000 -k 4000 --sampling-type nucleus_sampling -a gpt2-medium --wandb-group fig19 --wandb-log
Figure 19 right:
python language_generation/train.py -m 1000 -k 4000 --sampling-type beam_search -a gpt2-medium --wandb-group fig19 --wandb-log
Figure 20 left:
python language_generation/train.py -m 1000 -k 4000 --sampling-type nucleus_sampling -a gpt2-large --wandb-group fig20 --wandb-log
Figure 20 right:
python language_generation/train.py -m 1000 -k 4000 --sampling-type beam_search -a gpt2-large --wandb-group fig20 --wandb-log
@article{taori2022data,
title={Data Feedback Loops: Model-driven Amplification of Dataset Biases},
author={Rohan Taori and Tatsunori Hashimoto},
url={https://arxiv.org/abs/2209.03942},
year={2022},
}