Metric

Code to replicate Directional Bias Amplification

To measure Directional Bias Amplification, use the metric code in directional_biasamp.py

Replicate Experiments

Each of the 7 folders contains the code needed to replicate the experiments:

• paint_ex (painting example that creates Figure 2)
• coco_captioning (investigates equalizer model for Figure 3 examples)
• coco_mask (experiment with masking in COCO in Table 1)
• mals (men also like shopping with thresholding in Section 3.3)
• celeba (variance in estimator bias experiments for Section 5.2 and Figure 4)
• fit_bert (section 5.3 "Sentence completion: no ground truth.")
• compas (section 5.3 "Risk prediction: future outcomes unknown.")

Each folder contains a file, run.sh, that contains the necessary scripts to run.

Python Packages

• python = 3.7.3
• numpy = 1.19.1
• torch = 1.4.0
• torchvision = 0.5.0
• responsibly = 0.1.1
• scipy = 1.4.1

Other files

• Data splits are located here